Background

Testosterone therapy is increasingly promoted. No randomized placebo-controlled trial has been implemented to assess the effect of testosterone therapy on cardiovascular events, although very high levels of androgens are thought to promote cardiovascular disease.

Methods
A systematic review and meta-analysis was conducted of placebo-controlled randomized trials of testosterone therapy among men lasting 12+ weeks reporting cardiovascular-related events. We searched PubMed through the end of 2012 using (testosterone or androgen) and trial and (random*) with the selection limited to studies of men in English, supplemented by a bibliograph ic search of the World Health Organization trial registry. Two reviewers independently searched, selected and assessed study quality with differences resolved by consensus. Two statisticians independently abstracted and analyzed data, using random or fixed effects models, as appropriate, with inverse variance weighting.

Results
Of 1,882 studies identified 27 trials were eligible including 2,994, mainly older, men who experienced 180 cardiovascular-related events. Testosterone therapy increased the risk of a cardiovascular-related event (odds ratio (OR) 1.54, 95% confidence interval (CI) 1.09 to 2.18). The effect of testosterone therapy varied with source of funding ( P-value for interaction 0.03), but not with baseline testosterone level (P-value for interaction 0.70). In trials not funded by the pharmaceutical industry the risk of a cardiovascularrelated event on testosterone therapy was greater (OR 2.06, 95% CI 1.34 to 3.17) than in pharmaceutical industry funded trials (OR 0.89, 95% CI 0.50 to 1.60).

Conclusions
The effects of testosterone on cardiovascular-related events varied with source of funding. Nevertheless, overall and particularly in trials not funded by the pharmaceutical industry, exogenous testosterone increased the risk of cardiovascular-related events, with corresponding implications for the use of testosterone therapy.

Keywords:
Testosterone; Cardiovascular; Men; Trial

Background
In observational studies low serum testosterone is associated with cardiovascular disease [1,2]. Testosterone may protect or be a secondary risk marker of other processes [1-3]. On the precautionary principle, expert advice and reviews, largely based on observational evidence, warn that cardiovascular disease may be increased by androgen deprivation therapy [4] and low testosterone [5,6]. Awareness of low testosterone as a treatable condition is being raised [7,8]. Testosterone use is increasing [9-11], possibly as self-medication in response to advertising. In 2004 the Institute of Medicine (IOM) reviewed the evidence on testosterone therapy and concluded, largely based on placebocontrolled trials, that there is not clear evidence of benefit for any of the health outcomes examined [12]. The IOM recommended small-scale trials to establish the efficacy of testosterone therapy where other treatments were not available [12]. To our knowledge, no trial has been designed to assess the effect of testosterone therapy on cardiovascular morbidity or mortality. Previous metaanalyses of randomized placebo-controlled trials found that testosterone therapy resulted in a non-significantly higher risk of cardiovascular events, based on adverse events, but only included trials through March 2005 [13,14]. A more recent meta-analysis included trials through August 2008 but only reported on three specific cardiovascular outcomes, that is, arrhythmia, coronary bypass surgery and myocardial infarction[15]. Given the widespread use of testosterone, the high prevalence of cardiovascular disease in older men and no comprehensive assessment of the effect of testosterone therapy on cardiovascular events, an up-to-date meta-analysis may help inform clinical practice. We carried out a meta-analysis of adverse events from randomized placebocontrolled trials to examine the overall risk of cardiovascular-related events associated with testosterone therapy.

Methods
This meta-analysis follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist [see Additional file 1] and a published protocol (CRD42011001815)[16]. Two reviewers (LX and CMS) independently searched for and selected trials, resolving any differences by consensus. Two statisticians (GF and BJC) extracted information from the selected trials. Additional file 1. PRISMA 2009 Checklist. Format: DOC Size: 63KB Download file This file can be viewed with: Microsoft Word Viewer

Data sources and searches

We (LX and CMS) systematically searched PubMed until 31st December 2012 using (testosterone or androgen) and (random*) and trial with the selection limited to studies in men in English, because a preliminary search only found studies in Engl ish. We (LX and CMS) searched the World Health Organization (WHO) International Clinical Trials Registry Platform for any trial using testosterone as an intervention. From this search, we (LX and CMS) discarded any studies both agreed were irrelevant based on title or abstract and read the remaining. We did a bibliographic search of the selected trials and relevant reviews.

Study selection
We included randomized placebo-controlled trials giving cardiovascular-related events by study arm, because a report may focus on a particular aspect of the trial [17,18] and not report all events that have occurred [17-19]. We excluded trials that only gave treatment-related events in the testosterone arm because these might not include full reporting of events in the placebo arm. Initially, we intended to exclude trials that only reported withdrawals as potentially less comprehensive than reporting of adverse events [18,19], however this turned out to be a very fine distinction, so we included any trial reporting cardiovascular-related events by study arm. We included any randomized controlled trial (RCT) of testosterone, but not other androgens, compared with placebo, including a comparison against a background of other treatments, because men likely to be taking testosterone may also be in treatment for other conditions. We excluded trials of less than twelve weeks duration to assess long-term rather than acute effects of testosterone therapy. We checked for duplication based on overlapping authorship, study description, number of participants and participant characteristics. When duplication occurred we used the study with the most comprehensive description of adverse events.

Outcome
The primary outcome was composite cardiovascular-related events; because we anticipated too few events for robust assessment by cardiovascular event type; and a system-wide composite outcome may be most suitable for adverse events [20]. Cardiovascularrelated events were defined as anything reported as such by the authors, that is, events reported as cardiac disorders, cardiovascular complaints, cardiovascular events, vascular disorders, cardiac or cardiovascular, or where the event description fell within the International Statistical Classification of Disease (ICD) version 10 chapter IX (I00 to I99). Most trials only reported serious adverse events, but a few also reported a wider range of events, so we also examined the effect of testosterone therapy by seriousness. Seriousness was based on the US Department of Agriculture definition of serious adverse events and the type of cardiovascular event generally considered serious [21]. Serious cardiovascular events were defined as cardiovascular-related events which the authors described as serious adverse events or where the outcome was death, life-threatening, hospitalization, involved permanent damage or required medical/surgical intervention, or was one of the following types of cardiovascular event: myocardial infarction, unstable angina, coronary revascularization, coronary artery disease, arrhythmias, transient ischemic attacks, stroke or congestive heart failure but not deep vein thrombosis.

Data extraction and quality assessment

A statistician (GF) extracted the number of participants randomized and cardiovascular-related events by trial arm. Event classification was checked by a physician (LX). A second statistician (BJC) checked the information extracted. Where trials reported cardiovascular-related events without giving the study arm, we contacted the authors twice by email to ask for cardiovascular-related events by study arm. The reviewers (LX and CMS) independently used an established tool to evaluate the quality of each trial [22], focusing on the quality of reporting of cardiovascular-related adverse events. First, we reported whether cardiovascular-related events were individually listed in a table by study arm, because these are easier to identify unambiguously. Second, we reported whether the type and severity of cardiovascular-related events reported was either pre-specified or identified before the allocation was revealed, because issues have been expressed about the reporting of adverse events [23,24]. Cardiovascular-related events vary in severity making the selection criteria and categorization crucial to an outcome assessed from adverse events.

Sensitivity analysis
We initially planned only to assess whether the effects of testosterone on cardiovascular-related events varied with average baseline testosterone, because we did not expect sufficient trials for sub-group analysis by type of testosterone product or by type of cardiovascular-related event. However, the reporting of adverse events may be open to interpretation [23], and may not be comprehensive [25]. Given potential lack of clarity as to the selection of the cardiovascular-related adverse events reported, we also examined whether the effect of testosterone therapy varied with funding source. Finally, we also considered cardiovascular-related death as an outcome.

Data synthesis and analysis

We used the number of participants randomized as the denominator and included all cardiovascular-related events from the start. We used funnel plots and trim and fill to assess publication bias, that is, missing trials. We used I 2 to assess heterogeneity between trials, using fixed effects models where there was low heterogeneity (I 2 <30%), otherwise using random effects models. We obtained the pooled odds ratio, using the metabin function of the meta package in R 2.14.1 (R Development Core Team, Vienna, Austr ia). We used meta-analysis regression, with inverse variance weighting, to assess whether the effects of testosterone therapy varied with baseline testosterone or funding source, using the rma function of the metafor package in R 2.14.1. Initial analysis show ed the pooled odds ratio was similar using a Peto or a Mantel-Haenszel estimate; we used inverse variance weighting for consistency with the meta-regression. This study is an analysis of published data, which does not require ethics committee approval.

Results
The initial search yielded 1,882 papers, of which 169 were selected for full text scrutiny. Of these 169 papers, 31 concerned different placebo-controlled randomized trials among men of testosterone therapy of 12+ weeks reporting cardiovascular-related events by study arm. We found one additional recent trial from the WHO International Clinical Trials Registry Platform [26]. We did not find any additional such trials from a bibliographic search of these 32 papers or from eight reviews [27-34]. Two additional trials [35,36] selected for full text scrutiny had cardiovascular-related events shown in a plot from one previous meta-analysis [13], but not the other meta-analyses [14,15] or in the relevant publications [35-37]. The search did not find one small non-randomized trial [38] included in two previous meta-analyses [13,15]. The search found one additional trial [39] potentially relevant to the earlier metaanalyses [13,14], two additional trials [40,41] potentially relevant to the most recent meta-analysis [15] and 11 subsequent trials. We sought clarification concerning events by study arm for 10 trials as set out in Additional file 2. Six authors never responded [37,40,42-45], three responded but did not provide any relevant information [36,46,47] and one provided information [48]. We included this later trial [48], one of the others that gave cardiovascular deaths, but not other cardiovascularrelated events, by study arm [46] and one that gave vascular events, but not all cardiovascular-related events by study arm [44]. Figure 1 shows the search strategy resulting in 27 placebo-controlled randomized trials. Additional file 2. Trials where authors contacted for additional information and responses.[13,35-37,40,42-48,100]. Format: DOCX Size: 32KB Download file

Figure 1. Selection process for the placebo-controlled randomized trials of the effects of testosterone therapy on cardiovascular-related events. Table 1 shows the 27 trials over 25 years of 2,994 mostly middle-aged or older men (1,733 testosterone and 1,261 placebo) with low testosterone and/or chronic diseases, who experienced 180 cardiovascular-related events. Most of the trials were in Western settings. Thirteen trials were supported by the pharmaceutical industry. Two trials were stopped early [49,50], one because of adverse events among the men allocated to testosterone [50] and one because it would not be feasible to demonstrate - in the foreseeable future - a beneficial effect of testosterone [on mortality] by continuing the study [49]. Table 1. Characteristics of placebo-controlled randomized trials giving the effects of testosterone therapy on cardiovascularrelated events among men

Quality assessment
The quality assessment (see Additional file 3) shows that two trials provided a table with a comprehensive list of cardiovascularrelated events by study arm and eight trials provided a summary table of cardiovascular-related events by study arm. For 17 trials cardiovascular-related events were surmised from withdrawals and/or adverse events as given in Additional file 4, including one where the cardiovascular-related events were surmised from a P-value [65]. The type and severity of adverse events to be reported was pre-specified in one trial [61]. In the two trials terminated early the adverse events motivated termination and were identified

before treatment allocation was known [49,50]. Otherwise it was sometimes unclear whether the definition or classification was made by masked assessors. Additional file 3. Quality assessment of the selected placebo-controlled RCTs of the effects of testosterone therapy on cardiovascular-related events (CRE)[19],[26],[39],[41],[44],[46],[48]-[68]. Format: DOCX Size: 54KB Download file

Additional file 4. Description of cardiovascular-related events in the selected placebo-controlled RCTs [19],[26],[39],[41],[44],[46],[48]-[67],[101]. Format: DOCX Size: 61KB Download file

Data synthesis
The funnel plot (Figure 2) shows several small studies (on the left hand side) where testosterone reduced cardiovascular-related events, however, there were no similar small studies where testosterone increased cardiovascular-related events. The forest plot (Figure 3) shows that the trials were homogeneous (I2 = 7.8%). Testosterone increased the risk of a cardiovascular-related event in a fixed effect model, odds ratio (OR) 1.54, 95% confidence interval (CI) 1.09 to 2.18. Trim and fill revised the OR to 1.69 (95% CI 1.21 to 2.38). When the analysis was restricted to serious events, whose categorization is shown in Additional file 4, the estimate was very similar (OR 1.61 (95% CI 1.01 to 2.56)) and was revised to 2.01 (95% CI 1.30 to 3.14) by trim and fill.

Figure 2. Funnel plot of placebo-controlled randomized trials examining the effects of testosterone therapy on cardiovascular-related events.

Figure 3. Forest plots of placebo-controlled randomized trials examining the pooled effect of testosterone therapy on cardiovascular-related events.

Sensitivity analysis
The cardiovascular-related event rate was lower in trials funded by the pharmaceutical industry (4% (66/1,651)) than in other trials (8% (114/1,343)). In a meta-regression model, risk of cardiovascular-related events on testosterone therapy varied with the source of funding (P-value for interaction 0.03) but not with baseline testosterone ( P-value for interaction 0.70). In trials funded by the pharmaceutical industry testosterone had no effect on cardiovascular-related events, but in the other trials testosterone therapy substantially increased the risk of a cardiovascular-related event (Figure 4). Finally, 33 cardiovascular-related deaths were identified (22 testosterone arm and 11 placebo arm), for which the odds ratio was similar 1.42 (95% CI 0.70 to 2.89) to the estimate for all cardiovascular-related events, and was revised to 1.57 (95% CI 0.78 to 3.13) by trim and fill.

Figure 4. Forest plots of placebo-controlled randomized trials examining the pooled effects of testosterone therapy on cardiovascular-related events by source of funding: upper panel funded by the pharmaceutical industry and lower panel not funded by the pharmaceutical industry.

Discussion

This updated meta-analysis of placebo-controlled randomized trials, with a much larger number of participants than previous metaanalyses [13-15], adds to the previous findings by showing that testosterone therapy increases cardiovascular-related events among men. The risk of testosterone therapy was particularly marked in trials not funded by the pharmaceutical industry. The risks of cardiovascular-related events were similar by baseline testosterone. Several possible explanations exist for our findings. First, not all trials of testosterone therapy reported all cardiovascular-related events by study arm [36,46]. Trials favoring testosterone may be unpublished. However, the funnel plot (Figure 2) and trim and fill suggested trials favoring the placebo may be missing. Second, endogenous and exogenous testosterone may have different effects, with endogenous testosterone being protective, consistent with the observational evidence[1,2] and with testosterone declining with age when cardiovascular disease increases with age. However, a recent Mendelian randomization study, using genetic variants as an instrumental variable for endogenous testosterone, did not corroborate protective effects of endogenous testosterone on cardiovascular disease risk factors [69]. Another possibility is that serum testosterone is not a good indicator of androgen activity [70], as has long been suggested [71] and recently substantiated by the effective use of anti-androgens in prostate cancer at castrate levels of serum testosterone [72,73]. Third, endogenous testosterone may be beneficial, but other metabolites of exogenous testosterone, raised by testosterone therapy, such as estrogens or dihydrotestosterone, could mediate cardiovascular-related events. Exogenous estrogens do not protect men against cardiovascular disease [74]. However, higher free testosterone rather than higher estradiol appeared to mediate the cardiovascular events in a recent trial of testosterone therapy [75]. Few trials have examined the effects of dihydrotestosterone administration and have usually focused on prostate rather than cardiovascular effects [76-78]. The interplay of testosterone and dihydrotestosterone is complex and challenging to disentangle in RCTs [79]. Nevertheless, exogenous testosterone lowers HDL-cholesterol [15] and raises hemoglobin, hematocrit and thromboxane [15,80], all of which might contribute to cardiovascular disease. Thromboxane promotes clotting and blood vessel constriction. Natural experiments suggest that lower lifetime endogenous androgens are associated with a relatively lower risk of death from ischemic heart disease, based on legally castrated men [81] and men with Klinefelters syndrome[ 82]. Similarly, a meta-analysis of RCTs of androgen deprivation therapy found a non-significantly lower risk of cardiovascular mortality among men allocated to androgen deprivation [83], despite bias to the null by the competing risk of death from prostate cancer. Our findings are consistent with the three previous meta-analyses [13-15], which all indicated a non-significantly higher risk of testosterone therapy for a composite cardiovascular outcome of the events considered, despite discrepancies in some studies [13,14]. This meta-analysis based on many more trials (27), many more men (2,994) and correspondingly more events (180) produced a similar, but more precise estimate, with the confidence interval no longer including no effect. The difference between the estimates by funding source is consistent with other observations [84,85] and could be due to different reporting of adverse events in industry funded trials. Differences by funding source could also be due to differences between trials. Industry funded trials reported fewer cardiovascular-related events, which reduces power although it should not affect the direction of effect. Industry funded trials tended to be in younger men. It is possible, although unusual, for the effects of treatment to be crossed by age [86]. From a clinical perspective the issue is ensuring that the benefits of testosterone therapy outweigh the potential risks. Almost a decade after the IOMs report [12] the efficacy of testosterone therapy for health outcomes where treatments are not already available remains uncertain. Testosterone compared to placebo could be beneficial for glucose metabolism [65], depression[87,88], sexual dysfunction [26,89], bone density [90] and HIV wasting syndrome [91,92], although whether testosterone is better than established treatments for these conditions has not been clearly established. Cardiovascular disease is common in typical users of testosterone therapy, that is, older men. The 10-year risk of a cardiovascular event for US men aged 65 to 69 years is about 28% [93]. Assuming the increased risk of cardiovascular-related events seen here with testosterone therapy would give a number needed to harm of at most 90 per year of testosterone therapy. As such, further research might focus on obtaining evidence without interventions, for example by confirming that the observed negative associations of serum testosterone with specific cardiovascular diseases extends to other androgen biomarkers, such as androgen glucuronides [94], and to other study designs less open to biases, such as Mendelian randomization. Moreover, a gene in the steroidogenesis pathway (CYP17A1) is reliably associated with coronary artery disease [95]. Establishing if CYP17A1 acts, if at all, by raising or lowering androgens would also bring clarity. Finally, consideration could be given to whether further trials should be of agents that raise or lower testosterone.

Strengths and limitations

Despite providing a meta-analysis of all known placebo-controlled randomized trials, limitations exist. First, the reporting of adverse events may be open to conflicts of interest [24]. The funnel plot and analysis by funding source are consistent with that possibility. A very large market is at stake [9]. Second, in a trial of a therapy, such as testosterone, which may change how men feel or their sex drive, some accidental unmasking may have occurred. Few trials assessed or reported this possibility. Third, some men in the testosterone arm stopped treatment because of increased prostate specific antigen or polycythemia, which would bias towards the null. Fourth, RCTs are not always tagged as such and could be missed [96]. However, we searched broadly and found several potentially eligible trials that had not been included in previous meta-analyses. Fifth, our study cannot include on-going trials, such as the Testosterone Supplementation and Exercise in Elderly Men trial (NCT00112151) and The Testosterone Trial

(NCT00799617). However, these trials are not designed to assess the effect of testosterone on cardiovascular events and will take time to complete. If new trials show testosterone therapy to be strongly protective against cardiovascular disease, it would be against the general run of evidence to date making interpretation uncertain because of heterogeneity [97]. Sixth, the abstractors were not blinded. Seventh, most trials only reported fairly serious cardiovascular-related events, but the severity varied between trials, although for RCTs the reporting of events should be comparable within trials, and the events reported are more or less severe symptoms of cardiovascular disease on the pathway to cardiovascular mortality. An analysis restricted to events which could be identified as serious gave a similar estimate, but was limited by relying on events the authors had chosen to describe in detail by study arm and was revised upwards by trim and fill. Arguably, the standard definitions of event seriousness, including hospitalization, do not apply to frail older men, because they may be particularly prone to hospitalization. On the other hand, frail older people may also be most affected by any decrement to their already poor health, so hospitalization may represent a particularly significant event. Notably, an estimate based solely on deaths also had a similar point estimate, although the confidence interval included no effect because of low power. Eighth, two larger trials were terminated early [49,50] which reduces power and could affect the estimates. However, the terminations took place towards the end of the planned trials and did not specifically concern cardiovascular-related events. Nevertheless, early terminations may have slightly increased the estimate and widened the confidence intervals. However, the interpretation would, most likely, have been similar. Ninth, although meta-analyses are a mainstay of evidence-based medicine they may be less reliable than large RCTs. Meta-analysis may overstate the benefits of treatment [98] however, they are less prone to overstate the harms [98]. Subsequent large RCTs rarely reverse the direction of effect from metaanalysis [99]. Tenth, given the lack of detailed cardiovascular-event reporting secondary analysis by type of cardiovascular event was not possible. Such sub-group analysis would undoubtedly be etiologically valuable. However, from a public health perspective the issue is identifying side-effects, where a composite outcome relating to a particular system (here the cardiovascular system) has been recommended [20]. Finally, the difference observed by source of funding could just be chance variation; however, testosterone therapy increased the risk of cardiovascular-related events overall.

Conclusions
Appropriately prescribed testosterone is undoubtedly beneficial. However, caution needs to be exercised to ensure that the associated health benefits of testosterone therapy outweigh the potential increased risk of cardiovascular-related events, particularly in older men where cardiovascular disease is common.

Abbreviations
CI: confidence interval; CYP17A1: cytochrome P450 17A1; IOM: Institute of Medicine; ICD: International Statistical Classification of Disease; OR: odds ratio; PRISMA: Preferred Reporting Items for Systematic Reviews and Meta-Analyses; RCT: randomized controlled trial; WHO: World Health Organization. Systematic review reference number CRD42011001815.

Competing interests
All authors declare: no support from any organization for the submitted work; BJC has received research funding from MedImmune Inc., and consults for Crucell MV; no other relationships or activities exist that could appear to have influenced the submitted work.

Authors contributions
LX carried out the systematic search and drafted the manuscript. GF and BJC did the data extraction and analysis; they also reviewed the manuscript critically. CMS originated the idea, carried out the systematic search and helped draft the manuscript. All authors had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. LX is the guarantor. All authors read and approved the final manuscript.

Acknowledgements
The authors thank Steffie Woolhandler, David Himmelstein and Heidi Jones for their support.

References
Background

The mind-body nexus has been a topic of growing interest. Further data are however required to understand the specific relationship between mood and anxiety disorders and individual physical health conditions, and to verify whether these psychiatric disorders are linked to overall medical burden.

Methods
This study examined data collected from 942 men, 20 to 97 years old, participating in the Geelong Osteoporosis Study. A lifetime history of mood and anxiety disorders was identified using the Structured Clinical Interview for DSM-IV-TR Research Version, Non-patient edition (SCID-I/NP). The presence of medical conditions (lifetime) was self-reported and confirmed by medical records, medication use or clinical data. Anthropometric measurements and socioeconomic status (SES) were determined and information on medication use and lifestyle was obtained via questionnaire. Logistic regression models were used to test the associations.

Results
After adjustment for age, socioeconomic status, and health risk factors (body mass index, physical activity and smoking), mood disorders were associated with gastro oesophageal reflux disease (GORD), recurrent headaches, blackouts and/or epilepsy, liver disorders and pulmonary disease in older people, whilst anxiety disorders were significantly associated with thyroid, GORD and other gastrointestinal disorders, and psoriasis. Increased odds of high medical burden were associated with both mood and anxiety disorders.

Conclusions
Our study provides further population-based evidence supporting the link between mental and physical illness in men. Understanding these associations is not only necessary for individual management, but also to inform the delivery of health promotion messages and health care.

Keywords:
Mood disorder; Anxiety disorder; Comorbidity; Medical burden; Population-based study; Physical illness

Background
For centuries, it has been hypothesised and debated that a mind-body interaction exists. Over time, an evidence base has grown to support this notion, yet an incomplete understanding prevails[1,2]. Rates of depression and anxiety are generally higher among the physically ill than in the general population; however, results vary according to the type and severity of the chronic disease and methods of ascertainment [3,4]. Within clinical samples, patients with depression have been found to have a significantly higher number of somatic symptoms [5] and a higher risk of chronic diseases [6-8], such as coronary heart disease [9,10], congestive heart failure [11], stroke and dementia [12], diabetes [13,14], gastro oesophageal reflux disease (GORD) [15], osteoarthritis and rheumatoid arthritis [16], psoriasis [17], cancer [18], neurological disorders [19], pulmonary disease [20], liver disease[21], thyroid disorders [22] and asthma [23,24], as well as non-specific syndromes such as obesity [25], anaemia [26], renal dysfunction [27], chronic fatigue syndrome [28], chronic headaches [24], and chronic pain [29]. The strength of the association as well as the extent of the evidence varies for each medical condition. Similarly, anxiety disxorders, have been shown to be related to several medical conditions, such as cardiovascular disease [30], cancer [18], obesity [31] and other metabolic disorders [32], irritable bowel syndrome [33], gastrointestinal problems [34], GORD [15], thyroid disorders [35], psoriasis [17] and a higher number of medical symptoms [36]. In addition, as recently reviewed by Culpepper and colleagues, migraine and chronic pain, diabetes, peptic ulcer, arthritis and pulmonary disease are also associated [2]. The co-occurrence of mood or anxiety disorders and physical illness worsens the impact of symptom burden [36], impacts disease course, links to the deteriorating patients health status and functioning [7,37], affects medication response and treatment adherence [38], and increases the risk of complications [39] and even death [37]. Moreover, the comorbidity between mental and physical illness is relevant in terms of role impairment and work performance [24], as well as quality of life and health service use and costs [40]. The causal pathway between comorbid mental and physical illness is complex and remains unclear. It has been suggested that the bidirectionality of this relationship may involve several mechanisms [41,42] including biological (for example, increased proinflammatory cytokines, hypothalamic-pituitary axis deregulation, autonomic nervous system dysfunction, serotonin depletion, metabolic, immune and endocrine changes) [43], psychological (for example, behavioural, psychodynamic and cognitive factors) [44] and social (for example, impaired social support, loneliness, and social disengagement) [45]. Furthermore, psychotropic

medication use has been linked to a higher incidence of an array of health outcomes, including diabetes, falls and bleeding, osteoporosis, and sudden cardiac death [46-48]. Enhancing the quality of care for those with mental illness not only can improve mood and anxiety symptoms, but also physical health in patients with comorbid mental and physical illness [49]. Furthermore, treatment for these disorders may influence each other [50]. Thus, estimating the prevalence and understanding the association between mood and anxiety disorders and physical illness is important in order to make more accurate diagnoses and to provide integrated and effective treatment, with regard to syndrome reciprocal influences and medication interactions. Given these data, we aimed to describe the relationship between comorbid mental and physical illness in a large, representative sample of men residing in Australia, utilising structured clinical interviews, medical records, and clinical and self-reported data collected as part of the Geelong Osteoporosis Study (GOS) [51].

Methods
Participants
Data were derived from the GOS male cohort, a study originally designed to investigate the epidemiology of osteoporosis in Australia. Originally, an age-stratified, population-based sample of adult men (n = 1,540; response 67%, 20 to 97 years old) was randomly-selected from the Australian Electoral Commission (AEC) rolls for the Barwon Statistical Division between 2001 and 2006. The Barwon Statistical Division is situated in South-Eastern Australia, comprising both rural and urban communities. Further details of the study have been published elsewhere [51]. Of those invited to participate in the five-year follow-up study (n = 1,540), 141 had died, 41 had left the study region, 16 were unable to provide written informed consent, 139 were not able to be contacted and 225 declined to participate resulting in 978 participants (81% of the eligible baseline sample). For this analyses, those who did not undergo psychiatric assessment (n = 17) or provide medical information (n = 19) were excluded, resulting in a final sample of 942 eligible men. This study was approved by the Barwon Health Human Research Ethics Committee and written informed consent was obtained from all participants.

Assessments Psychiatric measures

The Structured Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, 4th edition, text revision (DSM-IVTR) Research Version, Non-patient edition (SCID-I/NP) [52] was used to identify those who had ever experienced a mood disorder, including major depressive disorder, bipolar disorder (types I and II), dysthymia, minor depression, mood disorder due to general medical condition (GMC) and substance-induced mood disorder; and/or an anxiety disorder including panic disorder, agoraphobia, social phobia, specific phobia, obsessive compulsive disorder, generalised anxiety disorder, anxiety disorder due to a medical condition, substance-induced anxiety disorder or anxiety disorder not otherwise specified. All interviews were conducted by personnel with qualifications in psychology, who were trained using live and videotaped interviews under the supervision of a psychiatrist.

Clinical measures
Height was measured to the nearest 0.1 cm using a wall-mounted stadiometer and body weight was measured to the nearest 0.1 kg using electronic scales. Body mass index (BMI; kg/m2) was calculated from these measurements. Lumbar spine and femoral neck bone mineral density (BMD) was measured by dual energy X-ray absorptiometry (DXA) using GE Lunar Prodigy (Prodigy GE Lunar Madison, WI, USA). Long term stability of the machine was tested three times weekly by scanning an anthropomorphic phantom (Hologic). Low bone mass was classified according to the World Health Organisation definition; one or more standard deviations below the young normal mean at the femoral neck or spine (L2-L4, posterior-anterior projection) [53].

Medical conditions
The presence of medical conditions (lifetime) was self-reported and confirmed by medical records, medication use or clinical data, where possible. Musculoskeletal disease was defined by low BMD or self-reported osteoarthritis, ankylosing spondylitis, psoriatic arthritis, rheumatoid arthritis, muscle disease or weakness, low trauma fracture, or self-reported or medicated gout (agents used in gout or hyperuricaemia). Thyroid disorders included self-reported or medicated (thyroid hormones and anti-thyroid agents) hyperthyroidism, hypothyroidism, Graves disease, Hashimotos disease, thyroiditis and other unspecified thyroid conditions. Metabolic risk factors included a diagnosis of diabetes, defined by self-report or current medication use (oral hypoglycaemic agents and insulin preparations), obesity defined by a BMI >30 kg/m2[54], self-reported or medicated hypercholesterolemia (use of cholesterol-lowering agents) and hypertension. Hypertension was characterised by a combination of self-report and current medication use (antihypertensive, beta-adrenergic blocking agents and diuretics) or a systolic blood pressure of >140 mmHg or diastolic blood pressure of >90 mmHg. Diagnosis of GORD was based on self-report as were gastrointestinal diseases which include peptic ulcer, chronic gastritis and causes of malabsorption, such as pancreatitis, gastric surgery, chronic diarrhoea, irritable bowel syndrome, inflammatory bowel syndrome and coeliac disease. Cardiovascular disease was defined by a diagnosis of hypertension (as

detailed above) or a combination of self-reported or medicated (beta-adrenergic blocking agents, diuretics, anti-arrhythmic, antiangina, cardiac inotropic agents, adrenergic stimulants, other cardiovascular agents) stroke, transient ischaemic accident (TIA), ischaemic cardiomyopathy, arrhythmia and valve and vascular diseases. Syncope and seizures included self-reported epilepsy, blackouts, fainting and dizzy spells. Self-reported migraine and other recurrent headaches were grouped together. Lifetime pulmonary disease included self-reported asthma, emphysema and chronic bronchitis. Psoriasis was defined by self-report. Liver disease diagnosis, including fatty liver, hepatitis, cirrhosis and liver failure, was also based on self-report. Cancer was defined as any malignant tumour including non-melanoma skin cancer. High medical burden was classified as a diagnosis of three or more conditions listed above.

Other measures
Smoking status, medication use and physical activity level were obtained by self-report [51]. Reported medication use was crossreferenced with the participants list of medications or containers and was considered current if the participants reported u se at the time of assessment. Men were classified as active if they participated in light to vigorous activity on a regular basis. Alcohol consumption (average grams consumed daily over a 12 month period) was determined by a validated food frequency questionnaire [55]. Tobacco smoking was documented and grouped as current or not. In order to determine socioeconomic status (SES), residential addresses were matched to the corresponding Australian Bureau of Statistics (ABS) Census Collection District (CCD). ABS software was used to determine the Socio-Economic Index for Areas (SEIFA) from the 2006 ABS Census data. SEIFA values summarize the characteristics of subjects within an area, thereby providing a single measure to rank the level of disadvantage at the area level, not of the individual subject. SEIFA values indicate the level of advantage or disadvantage within each CCD, small geographic areas that incorporate approximately 250 houses [56]. These values inform the Index of Relative Socioeconomic Advantage/Disadvantage (IRSAD), which accounts for various socioeconomic parameters including high and low income, educational attainment ranging from no qualifications to a tertiary degree or higher, the type of occupation from unskilled employment to professional positions, and some measure of wealth (such as owning a car, or number of bedrooms in a dwelling). The relative SES of participants was ascertained by categorizing the sample into quintiles based on cutpoints for the study region [57-59].

Statistical analysis
Differences in characteristics between those with and without mood and anxiety disorders were compared using chi-square analyses for discrete variables and MannWhitney for non-parametric continuous variables. Logistic regression was used to calculate odds ratios (OR) with 95% confidence interval (95% CI) of lifetime medical conditions for those with lifetime mood/anxiety disorders in comparison to those with no mood/anxiety disorder. Models were adjusted for age (model I) and then for age, BMI, SES, physical activity, alcohol consumption and smoking status (model II). Interactions between exposure variables were examined. Each analysis was performed with and without participants with mood disorders due to GMC; no differences in outcome were detected, therefore, those with mood disorder due to GMC were included in all analyses. Values ofP <0.05 were accepted as significant (including interaction terms). Statistical analyses were performed using Minitab (version 16; Minitab, State College, PA, USA).

Results
A total of 160 men (17%) had a lifetime history of mood disorder, of which 139 (86.9%) were identified with major depressive disorder, 10 (6.3%) dysthymia, 9 (5.6%) minor depression, 3 (1.9%) bipolar disorder, 3 (1.9%) mood disorder due to a GMC and 2 (1.3%) substance induced mood disorder. Of the 60 participants (12.2%) with a history of anxiety disorder, 23 (38.3%) met criteria for panic disorder, 15 (25%) post-traumatic stress disorder, 7 (11.7%) specific phobia, 6 (10%) social phobia, 5 (8.3%) generalized anxiety disorder, 5 (8.3%) obsessive compulsive disorder and 3 (5%) anxiety disorder not otherwise specified. Participants with a history of mood or anxiety disorder were younger than those without; otherwise the groups were similar (Table 1). Table 1. Characteristics for the total group and men with or without lifetime history of mood disorder and anxiety disorder

Mood disorder
After adjustments for age, mood disorder was associated with increased odds of GORD (OR 2.41, 95% CI 1.48 to 3.00, P = 0.027), recurrent headaches (OR 2.46, 95% CI 1.29 to 4.69, P = 0.007), syncope and seizures (OR 2.41, 95% CI 1.31 to 3.09, P = 0.002) and liver disorders (OR 2.41, 95% CI 1.40 to 4.98, P = 0.003). Furthermore, a history of mood disorder was associated with increased odds of high medical burden (age-adjusted OR 1.80, 95% CI 1.16 to 2.77, P = 0.008). Mood disorders tended to be associated with musculoskeletal disease (OR 1.43, 95% CI 0.93 to 2.21, P = 0.104) and gastrointestinal disorders (OR 1.47, 95% CI 0.93 to 2.30, P = 0.096). Age was an effect modifier in the association between mood disorders and pulmonary disease. Among older men (60 years), a lifetime history of mood disorder was associated with increased odds of pulmonary disease (age-adjusted OR 2.75, 95% CI 1.45 to 5.21, P = 0.002); no association was detected for those younger than 60 years ( P >0.05). These relationships remained significant after further adjustment for BMI, SES, physical activity, alcohol consumption and smoking status (Table 2). No

association was observed between mood disorders and thyroid disorders, metabolic risk factors, cardiovascular disease, cancer and psoriasis. Table 2. Age-adjusted (model I) and fully-adjusted (model II) odds ratios for medical comorbidities in men with lifetime history of mood and anxiety disorders

Anxiety disorder
After adjustment for age, anxiety disorder was associated with increased likelihood of thyroid disease (OR 5.29, 95% CI 1.63 to 7.14, P = 0.005), GORD (OR 3.84, 95% CI 2.01 to 7.33, P<0.001), gastrointestinal disease (OR 2.22, 95% CI 1.20 to 4.10, P = 0.011) and high medical burden (OR 2.82, 95% CI 1.43 to 5.53, P = 0.003). Anxiety disorder also tended to be associated with metabolic risk factors (OR 1.92, 95% CI 0.96 to 3.83, P = 0.066). Further adjustment for BMI, SES, physical activity, alcohol consumption and smoking status did not affect these outcomes (Table 2). The relationship between lifetime anxiety disorder and metabolic risk factors and lifetime anxiety disorders and psoriasis reached significance following adjustment for age, BMI, SES, physical activity, alcohol consumption and smoking status (OR 2.20, 95% CI 1.07 to 4.53, P= 0.032 and OR 2.77, 95% CI 1.00 to 7.67, P = 0.049, respectively). There were no associations detected between anxiety disorders and musculoskeletal disease, recurrent headaches, syncope and seizures, cardiovascular disease, pulmonary disease, liver disorders or cancer.

Discussion
Our population-based study reports associations between mood and anxiety disorders and physical illness. Mood disorders were associated with increased risk of many of the disease groups (GORD, neurological features, such as recurrent headaches and syncope and seizures, liver disorders and pulmonary diseases in older men). Anxiety disorders presented a different profile; thyroid disorders, GORD, gastrointestinal disease, metabolic risk factors, and psoriasis were more common among individuals with anxiety disorder. Importantly, mood and anxiety disorders were both associated with high medical burden. Our results confirm other population-based data investigating the association between mood disorder and GORD [60,61]. In a sample of over 60,000 participants residing in Norway, those identified with depressive, anxiety and comorbid symptoms, as measured by the Hospital Anxiety and Depression Scale, had a two- to three-fold increased risk of reflux symptoms, self-reported severe symptoms of recurrent heartburn or acid regurgitation [62]. Moreover, mood disorders have also been shown consistently to be more prevalent in patients with GORD in clinical practice[15,63]. Chronic headaches are high prevalence disorders, with as many as 46% of the adult population reporting headaches, 11% migraines and 42% tension headaches [64]. Similar to our findings and other population-based studies [65,66], Scott et al. [24] utilising data pooled from 18 general population surveys reported those identified with either non-comorbid depression or comorbid depression and anxiety, as measured by the Composite International Diagnostic Interview, had an age and gender adjusted odds ratio of 2.5 and 4.0, respectively, for chronic headaches. In regard to syncope and seizures, our results support Morgan et al. [67] who reported fainting, blackouts and dizziness, for which there is no adequate physical explanation, were associated with undiagnosed psychiatric disorders. Depression has also been shown to co-occur with epilepsy. Recently, the association has been viewed according to the diathesis-stress model, with depression resulting from the chronic stress associated with the threat of recurring seizures, as there is little evidence linking specific epilepsy related factors (for example, focus site) to low mood [68]. Akin to our results, Wilhelm and colleagues [69] reported an association between depression and liver pathology in a populationbased survey of 10,641 adults; however, the relationship was not sustained following adjustment for demographic and behavioural confounders including drinking behaviour. Excessive alcohol consumption is known to be highly correlated with liver disease [21]; however, the association persisted following correction for alcohol in our study. In our sample, only older men with a mood disorder were at increased risk of pulmonary disease. Although the association between asthma and mood disorder is evident across the full adult age range [23], chronic bronchitis and emphysema are more common among older people and most likely contribute to the different pattern observed for older and younger men. Furthermore, pulmonary disease can be considered in stages reflecting severity, a factor we could not explore[70]. Musculoskeletal and gastrointestinal disorders tended to be associated with mood disorders, although we speculate that the heterogeneous groupings we employed may have diluted associations in our sample population. Mood disorders and symptomology have previously been shown to be associated with low BMD and subsequent fracture, where both biological and medication related processes are thought to play a role [71]. Similarly, mood disorders have been shown to be associated with gastrointestinal disorders in both clinical and population-based samples [72-74]. In contrast to other population-based studies, we did not detect a relationship between mood disorders and thyroid disorders, metabolic risk factors, cardiovascular disease, cancer and psoriasis in this sample of men. The link between mood disorders and cardiovascular disease, thyroid disorders and metabolic risk factors, in particular, is well documented [22,24,30]. We hypothesize

that these results may be influenced by a healthy participant bias, a common issue where study participants are required to be healthy enough to attend the research centre. Furthermore, disease grouping and data collection limited our ability to identify the degree of illness severity (that is, severe stroke, TIA and medically controlled hypertension were grouped) or time since onset of the physical condition. As with mood disorders, GORD and gastrointestinal disorders were also associated with anxiety disorders. Anxiety has been previously considered to be a non-disease related factor that impacts negatively on quality of life associated with GORD [75] and, as previously discussed, is commonly observed within population-based samples reporting GORD symptoms [62]. Furthermore, data from the present study concur with previous studies indicating that those with anxiety disorder have higher rates of gastrointestinal disorders [34]. Psoriasis has been previously associated with anxiety disorders. Harter and colleagues [34] demonstrated that patients with a lifetime history of generalised anxiety disorder or panic disorder reported significantly higher lifetime prevalence rates for dermatological disorders, including psoriasis. Why this association is seen with anxiety disorders only in our sample population is unclear. Thyroid disorders and metabolic risk factors have also been associated with anxiety disorders. In our study, anxiety was linked to a combination of the most common risk factors for metabolic disorders, such as hypertension, diabetes, hypercholesterolemia and obesity, which Culpepper and colleagues reported to be all likely due to a prolonged stress response in patients with untreated anxiety disorder [2]. Similarly, Simon et al. [35] reported an increased risk of thyroid disorders in patients with generalised anxiety disorder, social phobia or panic disorder and recommended screening for thyroid dysfunction in patients with anxiety disorders. Our data suggest that both mood and anxiety disorders are associated with overall medical burden. These findings are consistent with previous studies showing both depression and anxiety to be associated with an increased number of somatic conditions [5,76]. It is plausible that this relationship may be due to mood and anxiety disorders causing physiological changes, as well as poorer self-care and treatment adherence, which in turn increase medical burden [36]. On the other hand, an increased physical burden, causing functional impairment, may result in the development of an anxiety and/or mood disorder, which have been shown to have a negative impact on clinical outcomes [77]. A major strength of our study is that the sample population spans the full adult age range, in contrast to previous studies that have mainly focused on older patients. An age interaction was only evident when exploring the association between mood disorders and pulmonary disease; thus the relationship between mental illness and each of the physical conditions was consistent across all ages. Further strengths of the study include the measurement of mood and anxiety disorders, diagnoses were made utilising semistructured clinical interviews (SCID-IV), a gold standard tool, and the consideration of several possible confounding factors. However, our observations must be considered with caution. Ascertaining medical conditions by a self-report checklist may be compromised by imperfect recall and response bias error. Although many of these disorders were based on self-report of a physicians diagnosis, which have been demonstrated to generally agree with medical record data [78], where possible this was confirmed by medical records, medication use or clinical data. The cross-sectional nature of the study does not allow verification of a cause-effect relationship between mental and physical illness, longitudinal studies are needed to determine the directionality of this relationship. Finally, we were unable to make a distinction between diseases with an episodic course and those with ongoing symptoms, which may impact differently on psychological status.

Conclusion
Our study provides further population-based evidence supporting the link between mental and physical illness in men. Recognising the link between these illnesses is important in order to manage the development of mood and anxiety disorders in response to a medical condition, or the presence of physical pathologies due to predisposing psychiatric disorders. Integrating treatment approaches and monitoring comorbidity of mental and physical illnesses is likely to improve disease course and outcome as well as enhance patients functional and health status, potentially reducing healthcare utilisation.

Abbreviations
ABS: Australian Bureau of Statistics; BMD: Bone mineral density; BMI: Body mass index; CI: Confidence interval; DSM-IV-TR: Diagnostic and Statistical Manual for Mental Disorders, 4th edition, text revision; DXA: Dual energy X-ray absorptiometry; GMC: General medical condition; GORD: Gastro oesophageal reflux disease; GOS: Geelong Osteoporosis Study; IRSAD: Index of Relative Socioeconomic Advantage/Disadvantage; SCID-I/NP: Structured clinical interview for DSM-IV-TR research version, nonpatient edition; SEIFA: Socioeconomic Index for Areas; SES: Socioeconomic status; TIA: Transient ischaemic accident.

Competing interests
Livia Sanna, Amanda L Stuart, Mark A Kotowicz, and Sharon L Brennan have no conflicts of interest, including specific financial interests and relationships and affiliations relevant to the subject matter or materials discussed in the manuscript. Paolo Girardi has in

the past three years received research support from Lilly and Janssen, participated in advisory boards for Lilly, Organon, Pfizer, and Schering, and received honoraria from Lilly and Organon. Julie A Pasco has received speaker fees from Amgen, Eli Lilly and Sanofi-Aventis and funding from the Geelong Region Medical Research Foundation, Barwon Health, Perpetual Trustees, the Dairy Research and Development Corporation, The University of Melbourne, the Ronald Geoffrey Arnott Foundation, ANZ Charitable Trust, the American Society for Bone and Mineral Research, Amgen (Europe) GmBH and the NHMRC. Michael Berk has received Grant/Research Support from the NIH, Simons Foundation, CRC for Mental Health, Stanley Medical Research Institute, MBF, NHMRC, Beyond Blue, Geelong Medical Research Foundation, Bristol Myers Squibb, Eli Lilly, Glaxo SmithKline, Organon, Novartis, Mayne Pharma, Servier and Astra Zeneca. He has been a paid consultant for Astra Zeneca, Bristol Myers Squibb, Eli Lilly, Glaxo SmithKline, Janssen Cilag, Lundbeck and Pfizer and a paid speaker for Astra Zeneca, Bristol Myers Squibb, Eli Lilly, Glaxo SmithKline, Janssen Cilag, Lundbeck, Organon, Pfizer, Sanofi Synthelabo, Solvay and Wyeth. Lana J Williams has received Grant/Research support from Eli Lilly, Pfizer, The University of Melbourne, Deakin University and the NHMRC.

Authors contributions
LS and ALS took part in the conception and design of the study, acquisition of the data, data cleaning and statistical analysis, interpretation of the analysis and took primary responsibility for writing the manuscript. JAP and MAK took part in the conception and design of the study, interpretation of the analysis and critically revised the manuscript. MB, PG and SLB took part in the interpretation of data and critically revised the manuscript. LJW took part in the conception and design of the study, interpretation of the analysis, critically revised and supervised the writing of the manuscript. All authors read and approved the final manuscript.

Acknowledgments
The study was funded by

Background
Surgeons are usually exposed to high workloads leading to fatigue and stress. This not only increases the likelihood of mistakes during surgery but also puts pressure on surgeons to use drugs to counteract fatigue, distress, concentration deficits, burnout or symptoms of depression. The prevalence of surgeons taking pharmacological cognitive enhancement (CE) or mood enhancement (ME) drugs has not been systematically assessed so far.

Methods
Surgeons who attended five international conferences in 2011 were surveyed with an anonymous self-report questionnaire (AQ) regarding the use of prescription or illicit drugs for CE and ME and factors associated with their use. The Randomized Response Technique (RRT) was used in addition. The RRT guarantees a high degree of anonymity and confidentiality when a person is asked about stigmatizing issues, such as drug abuse.

Results
A total of 3,306 questionnaires were distributed and 1,145 entered statistical analysis (response rate: 36.4%). According to the AQ, 8.9% of all surveyed surgeons confessed to having used a prescription or illicit drug exclusively for CE at least once during lifetime. As one would expect, the prevalence rate assessed by RRT was approximately 2.5-fold higher than that of the AQ (19.9%; 95% confidence interval (CI), 15.9% to 23.9%, N = 1,105). An even larger discrepancy between the RRT and AQ was observed for the use of antidepressants with a 6-fold higher prevalence (15.1%; 95% CI, 11.3% to 19.0%, N = 1,099) as compared to 2.4% with the AQ. Finally, logistic regression analysis revealed that pressure to perform at work (odds ratio (OR): 1.290; 95% CI, 1.000 to 1.666; P = 0.05) or in private life (OR: 1.266; 95% CI, 1.038 to 1.543; P = 0.02), and gross income (OR: 1.337; 95% CI, 1.091 to 1.640; P = 0.005), were positively associated with the use of drugs for CE or ME.

Conclusions
The use of illicit and prescription drugs for CE or ME is an underestimated phenomenon among surgeons which is generally attributable to high workload, perceived workload, and private stress. Such intake of drugs is associated with attempts to counteract fatigue and loss of concentration. However, drug use for CE may lead to addiction and to overestimation of ones own capabili ties, which can put patients at risk. Coping strategies should be taught during medical education.

Keywords:

Cognitive enhancement; Surgeon(s); Prescription drug(s); Stimulant(s)

Background
Surgeons are often exposed to an excessive workload leading to mental and physical exhaustion, for example, fatigue, sleep deprivation (especially in shift work), burnout, and even to higher rates of suicide [1-5]. This increases the likelihood of mistakes during surgery [6-9]. In order to maintain high cognitive performance, surgeons can come under pressure to counteract fatigue, distress, concentration deficits, burnout or symptoms of depression by the use of enhancing substances. Warren and colleagues discussed possible reasons for surgeons to consider the use of substances for cognitive enhancement (CE) or mood enhancement (ME), including such issues as internal (to maximize ones own potential) and external pressure (by the employer or the public), patient safety or surgeons well-being [10]. However, there is no scientific evidence so far that either external or internal pressure leads to the consumption of prescription or illicit drugs to cope with these stressors. Nevertheless, performance enhancing drugs may be especially attractive to surgeons due to the fact that they appear to be a time-saving and, to some extent, an effective and easy alternative to more time-consuming coping strategies (for example, napping, sleep, relaxation techniques, and so on) or the use of coffee which may cause tachycardia and worsening tremor [10,11]. Thus, the use of prescription and illicit drugs could be a coping strategy to manage poor working conditions. However, (psycho-) stimulants (amphetamines (AMPH), methamphetamine, methylphenidate (MPH)), modafinil as well as antidementives and antidepressants have no consistent effects for CE or ME in healthy non sleep-deprived subjects [12-20]. Nevertheless, reduction of cognitive performance due to sleep deprivation is a common problem in shift work. Stimulants and modafinil have been demonstrated to attenuate disruption in cognitive performance and mood during night-shift work and sleep deprivation [20-25]. Previous studies revealed the use of stimulants and modafinil for CE purposes among students having at least an immediate provigilant effect [14,26,27]. Beyond that, antidepressants such as selective serotonin reuptake inhibitors (SSRI) are known to be used for ME, although previous studies showed no immediate or delayed mood enhancing effect [18,28,29]. Furthermore, it is unclear whether ME affects only mood or also other aspects such as self-esteem or self-representation. Previous studies among high school and university students using anonymous questionnaires (AQ) have shown prevalence rates for the use of prescription stimulants (MPH) to be 0.8% and 2.9% for illicit stimulants (AMPH) whereas 10.5% decided to use caffeine tablets for CE [27,30,31]. Furthermore, an online poll conducted by the journal Nature depicts a lifetime prevalence rate of 20% for stimulants, modafinil or beta blockers for CE purposes among participating academics[32]. A recent study using the Randomized Response Technique (RRT) shows prevalence rates of even 20% for the use of prescription and illicit drugs among university students for CE [33]. Studies using RRT guarantee an especially high degree of privacy, anonymity, and confidentiality when a person is prompted to answer sensitive questions about socially undesirable or illicit behavior [34-38]. This present survey reports the first data on prescription and illicit drug use for CE and ME among surgeons, together with usage association factors. We estimated prevalence rates of CE and ME among surgeons using AQ and RRT.

Methods
The data for this study resulted from a survey conducted in 2011 among 3,306 German-speaking surgeons who attended five international conferences of the German Society of Surgery (Deutsche Gesellschaft fr Chirurgie). After the first conference, potential participants were asked if they had been assessed before; those who had been were excluded from a second participation. Based on previous research about prescription and illicit drugs for CE, an AQ about the use of prescription and illicit drugs for enhancing cognitive functions or mood was developed and distributed to the participants during the conferences. The AQ asked about the non-medical use of stimulants with the particular intention of CE and the non-medical use of antidepressants with the particular intention of ME during lifetime, last year, last month and last week (frequency). Furthermore, we asked for the age of first use. Beyond that, the AQ included questions about potential risk factors associated with usage of drugs and questions about biometrical parameters (for example, gender, age, age of first use, and so on.). We were mainly interested in healthy participants using drugs specifically for CE or ME. Therefore, the data of participants with self-reported psychiatric disorders (for example, depression, attention deficit/hyperactivity disorder, (ADHD)) who had physicians prescriptions for any drug were excluded. Participants were asked to drop the questionnaire into black boxes after having filled in the questionnaire anonymously. The study was carried out according to the Principles for Medical Research Involving Human Subjects according to the Declaration of Helsinki. The study was approved by the local Ethics Committee (Landesrztekammer Rheinland-Pfalz) (No. 837.321.08 (6318)). Participants gave informed consent by returning the questionnaire and were informed about this procedure in the introduction section of the questionnaire; this procedure was approved by the above mentioned local Ethics Committee.

The AQ contained questions about the use of drugs for CE and ME as well as questions using RRT. After a brief introduction about the RRT stressing the anonymity of this technique, questions were presented to participants as follows: Please consider a certain birthday (yours, your mothers, etc.). Is this birthday in the first third of a month (1st to 10th day)? If yes, please proceed to Question A; if no, please proceed to Question B. Question A: Is this birthday in the first half of the year (prior to the first of July)? Question B: Did you ever use prescription and/or illicit drugs (e.g. Methylphenidate, Modafinil, illicit Amphetamines, and so on) without a medical need for cognitive enhancement? Note that only you know which of the two questions you will answer o Yes o No For assessing the use of antidepressants for mood enhancement, we modified Question B as follows: Did you ever use antidepressants without medical need for enhancing your mood and/or self-esteem, self-presentation? The interviewers are not able to know which question the respondent has to answer. Therefore, participants can reply honestly without compromising themselves. Of all participants, 67.1% (245.25/365.25) received the sensitive question (B) and 32.9% (120/365.25) the non-sensitive question (A).

s=a(1p)Np
the proportion of yes responses with respect to the sensitive questions can be estimated from proportion a of total yes responses in the sample. p denotes the probability of receiving the sensitive question (Question B; p = 67.1% of all participants received this question). The probability of answering the non-sensitive question (A) with yes is n = 49.6% (181.25/365.25). A 95% confidence interval (CI) for the unknown prevalence can be computed from the sampling variance Var(s)=a(1a)np2 where n denotes the sample size [33,39].

Forexamplefora=328/11050.30,p=245.25/365.250.67,N=181.25/356.250.50,andn=1105,thisy ieldss=328/1105(1245.25/365.25)181.25/365.25245.25/365.250.1993=19.93%.Var(s)=328/1105(1328/1105)1 105(245.25/365.25)20.000419.SE(s)=0.0004190.0205=2.05%.Thusthe95%confidencein tervalis19.931.962.05,thatis,rangesfrom15.91%to23.95%.

Statistical analyses were performed with SPSS for Windows, Version 17.0. Means are given with their corresponding standard deviation (SD) (mean SE) and Clopper-Pearson confidence intervals (95% CI). AQ questions were analyzed using a multiple logistic regression analysis. For the regression, the variable selection procedure was performed by using stepwise forward selection with a selection level of 0.05. The variables which were analyzed as potential multivariable predictors of the use of prescription or illicit drugs for CE before forward selection were (available parameters): pressure to perform at work, pressure in private life, gross income, gender, age, family status, living with children, type of employer, employment status, hours of work, satisfaction with professional success, evaluation of career opportunities, pressure to perform subjectively evaluated as burdensome and pressure to perform subjectively evaluated as harmful to health. Ordinal variables with five or more categories (pressure to perform at work, pressure in private life, gross income, pressure to perform subjectively evaluated as burdensome and pressure to perform subjectively evaluated as harmful to health) were treated as continuous variables. Table 1 shows the variables included for the regression after forward selection; all variables which significantly influence the drug use for CE/ ME are listed in this table. There are no further variables for which we adjust. The results are presented as odds ratio (ORs) with confidence limits and P-values. The regression has been analyzed by referring to cases without missing values (complete case analysis). Table 1. Risk factors for the use of prescription or illicit drugs for CE among surgeons by multivariable analysis

Results
Participants characteristics
A total of 1,204 (36.4%) of 3,306 distributed questionnaires were returned. Of the participants, 61 had to be excluded: 9 for not being a physician, 11 for having a physician`s prescription for drugs because of mental disorders and 39 (pair wise) for giving

incomplete answers. Thus, the data of 1,145 surgeons entered the final statistical analysis. The mean weekly workload was estimated to be 56.8 13.0 hours. Of the respondents, 56.4% were surgeons in training, 21.5% senior surgeons and 22.1% directors or deputy directors. The pressure to perform optimally at the job was estimated to be severe (3.3 1.2 on a 6-point Likert scale, 0 = not at all, 5 = very much), judged to be moderate to severely burdensome (2.6 1.3) and moderate to severely harmful (2.7 1.3). For further details, see Table 2. Table 2. Participants` characteristics Lifetime-prevalence of CE and ME is higher than last-year prevalence, which in turn is higher than last-month and last-week prevalence rates (see Table 3 and 4). Differences between last-year and last-month prevalence rates for CE and ME are small, whereas the difference between lifetime prevalence and last-year prevalence rate is remarkably higher. Age of first use did not differ significantly between prevalence rates. For more details on lifetime-, last-year, last-month and last-week prevalence rates, as well as for age of first use using the AQ, see Tables 3 and 4. Table 3. AQ results for prevalence rates of the use of prescription drugs + illicit drugs for CE

Table 4. AQ results for prevalence rates of the use of antidepressants for ME

Prevalence rates of CE and ME measured by AQ compared to the randomized response technique (RRT)
Prevalence rates measured by the RRT are considerably higher than prevalence rates measured by AQ. Table 5 shows that with AQ, 8.9% of the surgeons confessed to having used a prescription or illicit drug exclusively for CE at least once during their lifetime. In contrast, the corresponding RRT estimate was approximately 2.5-fold higher than the AQ estimate, that is, 19.9% (95% CI, 15.9% to 23.9%, n = 1,105). An even larger discrepancy between the RRT and AQ was observed for the use of antidepressants with a 6-fold higher prevalence rate, that is, 15.1% (95% CI, 11.3% to 19.0%, n= 1,099) as compared to 2.4% with the AQ. Table 5. AQ and RRT results for lifetime prevalence rates of prescription or illicit drug use for cognitive enhancement (CE) or mood enhancement (ME)

Factors associated with the use of prescription and illicit drugs for CE/ ME
Finally, a logistic regression analysis revealed that pressure to perform at work (OR: 1.327; 95% CI: 1.010 to 1.743; P = 0.042) or in private life (OR: 1.252; 95% CI: 1.015 to 1.543; P = 0.036) and gross income (OR: 1.406; 95% CI: 1.133 to 1.744; P = 0.002) were positively associated with the use of drugs for CE or ME (see Table 4). Logistic regression analysis suggests that other factors play no role in the use of prescription and illicit drugs: gender ( P = 0.809), age (P = 0.620), family status (P = 0.698), living with children (P= 0.720), type of employer (P = 0.151), employment status (P = 0.820), hours of work (P= 0.366), satisfaction with professional success (P = 0.829) and evaluation of career opportunities (P= 0.822).

Discussion
The AQ results of this study indicate that 8.9% of all surveyed surgeons used prescription or illicit drugs with the particular intention of CE by AQ. By contrast, the RRT results showed a higher prevalence of 19.9%. Furthermore, using AQ, 2.4% answered that they had already used antidepressants for ME whereas the RRT revealed a prevalence of 15.1%. Furthermore, prescription or illicit drug use for CE or ME was associated with the pressure to perform at work or in private life and with gross income. On the one hand, there are substantial differences regarding the prevalence rate in the present study. On the other hand, there are significant differences compared with previous studies of drug use for performance enhancement. There are an increasing number of studies dealing with academic performance enhancement, cognitive enhancement or pharmacological neuroenhancement regarding cognition (for example, [10,26,27,32,40-42]). Regarding prevalence rates and associated factors, it is useful to consider several factors as follows: With the exception of the present study, there exists a severe paucity of data about drug use for CE among employed adults. DAK, a German health insurance company, online surveyed via e-mail 20,000 employed members (20- to 50-years old) with a response rate of 15%. Participants were asked about their use of various substances for CE and mental well-being without medical need [43]. Without accurately distinguishing prescription and over-the-counter drugs, the non-representative DAK study showed a lifetime prevalence rate of 5%.

Stated reasons for usage were: depressed mood, anxiety, nervousness, uneasiness, memory deficits, fatigue, and p roblems of concentration [43]. These rationales seem to be the same as among surgeons [1-5]. Furthermore, a non-random online poll by the journal Nature which unfortunately did not specify respondents, demonstrated that 20% of participants had already used prescription drugs for non-medical reasons to improve concentration and improve their focus for a specific task. MPH was the most popular substance, followed by modafinil and beta blockers [32]. MPH and modafinil are also the most prevalently used drugs in our survey. This agrees with the results of our study, although, admittedly, the surveyed groups are not directly comparable. Interestingly, these results match the RRT results of our study. Both surveys, online polls as well as the present RRT study, guarantee a relatively high level of anonymity. This may be one of the most important aspects when assessing pharmacological CE or ME, both potentially highly stigmatizing subjects. Outside of these particular studies, only students substance use for academic performance enhancement has been surveyed. A previous study by our research group among 1,500 high school and university students (over 18 years) using AQ, assessed lifetime prevalence rates of 1.29% for prescription stimulants (MPH, AMPH) and 2.6% for illicit stimulants [27]. Regarding lifetime prevalence, we found that prevalence rates for stimulants in the present study are slightly higher than in our earlier studen ts survey. This may be associated with the older age of surgeons. Furthermore, both studies excluded participants with ADHD or other psychiatric disorders where prescribed psychiatric medications were involved. Many other studies did not exclude these patients, revealing higher prevalence rates including those where participants misuse their own prescribed medication [40]. A meta-analysis by Wilens and colleagues examining prevalence rates of stimulant misuse included 21 US studies with 113,000 participants revealing a past-year prevalence rate of 5% to 9% in grade and high schools and 5% to 35% in college-age individuals [40]. For this important meta-analysis which included many significant studies about stimulant misuse among students, CE is only a side aspect of the study. This explains the substantially higher (past-year) prevalence rate compared to the present study. However, to concentrate and improve alertness have been salient participants reas ons for misuse of stimulants. The most recent study about CE among 2,600 students using the RRT shows a one-year prevalence rate of 20% for the non-medical use of prescription and illicit drugs for CE [33]. These results show a comparable prevalence to those of the present study. Beyond that, Partridge and colleagues revealed that a high percentage of the public media portrayed CE as common which accords with our high prevalence rate for CE [42]. However, this finding contrasts with that of our survey study of the same group among university students leading to the assumption of a phantom debate [44,45]. While we were not able to show a significant influence of gender on the use of potential CE- or ME-substances, Dietz and colleagues revealed that significantly more male than female students used prescribed or illicit drugs for CE. Our results do not confirm this finding. The literature is somewhat inconsistent on this subject. For the illicit use of prescription ADHD medications among college students, DeSantis and colleagues found a significantly higher prevalence rate in male than in female students [46], whereas Teter and colleagues found no gender differences regarding prescription stimulant use among college students [47]. However, studies focusing upon this particular association in the context of a different surveyed group from that of the present study, suggest higher risk behaviors in male compared to female subjects [33,48]. Surveyed surgeons answered that their age of first use of prescription or illicit drugs for CE was 24.0 years. However, our previous study among 1,500 students revealed 17 to 18 years to be the age of first use of prescription or illicit drugs for CE [27]. This is almost 5 to 6 years younger than among surgeons, who themselves had been medical students and later trainee surgeons, decades before. However, study participants are 43 years old (mean) which may imply that two decades ago, the use of CE drugs started substantially later in life. Beyond that, first use of antidepressants for ME was 39 years (mean) compared to an average of 24-yearold participants using CE drugs for the first time. Methodologically, all these studies only allow an indirect comparison of different survey methods. The present study allows us for the first time to compare AQ questions with RRT questions in one single integrated survey about drug use. In this respect, a previous meta-analysis of 38 RRT validation studies by Lensvelt-Mulders and colleagues reported that RRT provides more valid data than other survey methods. This strengthens the validity of the RRT prevalence rates of 19.9% for CE and 15.1% for ME [49]. This underlines the relevance of the survey method in general. In particular, it strengthens the validity and reliability of the higher RRT results of 19.9% and 15.1% for CE and ME compared to the lower prevalence rates using direct questions. We were able to show that pressure to perform at work or in private life, together with gross income, are positively associated with the use of prescription or illicit drugs for CE or ME and are the only factors associated with drug use for this purpose. Further hypothesized factors were revealed to play no role in the use of prescription or illicit drugs for CE. We found a positive association of pressure to perform at work or in private life and gross income with the use of drugs for CE. However, we cannot interpret this finding as a general proof of a direct causal relationship between feeling pressure and the use of CE substances. Furthermore, this association is not tenable for professional life in general. Such factors should be addressed in detail in further studies. At least to our present knowledge, there are no data from empirical studies which allow meaningful comparison with our data: studies of stude nts drug use focusing on CE as well as the previously cited Nature poll did not examine these factors.

Surgeons should know about the effects and side-effects of the substances used for CE or ME, at least regarding prescription drugs, such as methylphenidate, amphetamine tablets (for example, Adderall), atomoxetine, modafinil, antidementive drugs and antidepressants. A survey study by Partridge and colleagues showed that university students already seem to have a realistic idea of the effects and side-effects [44]. According to randomized controlled trials (RCTs), reviews and meta-analyses there are almost no pro-cognitive effects regarding normal healthy non-sleep-deprived subjects on simple and higher cognitive domains [12-17]. Nevertheless, stimulants and modafinil have enhancing effects on simple cognitive domains, such as fatigue, vigilance, psychomotor skills and reaction time in sleep-deprived subjects; furthermore, there are slightly pro-cognitive effects on higher cognitive domains and, beyond that, stimulants have subjective pro-effects [12,14,19-25]. One can presume that the effects on higher cognitive skills are indirect effects which are mediated via simple cognitive skills, for example, vigilance. The fact that sleep deprivation leads to clearer results supports this hypothesis [12]. One would expect surgeons to know these limited effects and to avoid the use of these prescription and illicit drugs for CE. However, every fifth surgeon has already used these drugs. We can only speculate about the reasons. On the one hand, surgeons may not know the missing pro-cognitive effects or overestimate the effects of such drugs. On the other hand, knowledge and even overestimation about pro-cognitive effects in sleep-deprived subjects only confirms that sleep deprivation is a common phenomenon among surgeons. Beyond that, antidepressants (such as SSRI) have no mood enhancing effect in normal healthy subjects at all [12,18]. Nevertheless, 15% have already used antidepressants for ME, which may indicate missing knowledge about the effects of antidepressants in normal healthy subjects or overestimation of the putative effects. Another important factor is the side-effect profile and safety risks of amphetamines which have to be considered. Beyond severe side-effects which are described in package-inserts accompanying these drugs and the results of RCTs, reviews and meta-analyses (for example, jitteriness, agitation, cardiologic side effects, such as tachycardia, hypertension, gastro-intestinal side effects, such as stomach ache, diarrhea, and so on), stimulants can cause addiction and further addictive behavior. Also, the misuse of illicit drugs and prescription drugs without prescription is a federal offense. A number of limitations of the present study are worth identifying here. We obtained a response rate of 36.4% which is a low response rate compared to previous studies using RRT (for example,[27,33]). However, questionnaires were distributed during conference breaks in huge buildings, so that we were hardly able to control potential participants behavior concerning the return of the questionnaires to the black boxes provided. Furthermore, substance use or even misuse can be considered a highly stigmatizing subject leading to low response rates. Thus, a response rate of 36.4% may be considered relatively high and comparable to other studies assessing stigmatizing subjects with anonymous questionnaires [36]. However, the response rate of 36.4% together with the non-random sample limits the generalizability of our findings. Another important factor is the likelihood of a participation bias: Since the response rate is only one third, we do not know in particular whether subjects with more positive attitudes or more negative attitudes on the topic participated disproportionately which may have caused a response bias. Since many more male subjects participated in our study, a potential gender bias exists. This may explain why we did not find gender differences in prevalence rates whereas earlier studies including our own have partly shown that male subjects more often used drugs for CE than female subjects. Beyond that, we asked surgeons for the non-medical use of stimulants for CE and antidepressants for ME. However, we did not specifically ask for the context of use, for example, whether surgeons had used it directly prior to surgical interventions.

Conclusions
The use of illicit and prescription drugs for CE or ME is an underestimated phenomenon among surgeons. The present results indicate that about 15% to 20% of surgeons have used drugs for CE or ME at least once during their lifetimes. This may be attributed to high workload and perceived work-related and private stress. Substances such as modafinil seem to counteract fatigue and loss of concentration and thus may provide simple pharmacological help for stressed surgeons. However, pro-cognitive effects on higher cognitive domains are very limited. Furthermore, the side effects and effects of long-term use (for example, misuse, addiction) of such drugs seem to be underestimated by users. Beyond that, stimulant use may put patients at risk given the fact that previous research has shown that stimulants may lead to overestimation of ones own capabilities. Both factors may be harmful for users. The contemporary debate on cognition -enhancing drugs requires a broader data base on consumption rates in populations at risk, together with careful studies of drug (side) effects to substantiate discussions of ethical and legislative aspects. Therefore, I) information about the restricted usefulness and risks of the use should be

provided, II) guidelines on how to deal with drug use among employees who have contact with patients have to be provided, and III) information about, and the development of, relevant coping strategies has to become an integral part of medical education.

Abbreviations
AMPH: Amphetamine(s); AQ: Anonymous questionnaire; ADHD: Attention deficit/ hyperactivity disorder; CE: Pharmacological cognitive enhancement; CI: Confidence interval; ME: Pharmacological mood enhancement; MPH: Methylphenidate; OR: Odds ratio; RCT: Randomized controlled trial; RRT: Randomized Response Technique; SD: Standard deviation; SSRI: Selective serotonin reuptake inhibitor

Competing interests
The authors declare that they have no competing interests.

Authors contributions
AGF, KL and PS participated in the conception and design of the study. AGF, KL and PS monitored data collection. AGF, KL, CB and in particular IH analyzed and checked the AQ data; PD, PS and RU analyzed the data of the Randomized Response Technique (RRT) which was cross-checked by IH. All authors participated in data interpretation, drafting, and revising the manuscript. All authors read and approved the final manuscript.

Authors information
AGF, CB and KL belong to the Department of Psychiatry and Psychotherapy, University Medical Centre Mainz, Germany. KL and AGF are psychiatrists, CB is a sociologist. IH is a mathematician and expert in statistics belonging to the Institute of Medical Biostatistics, Epidemiology and Informatics (IMBEI) of the University Medical Center of the Johannes-Gutenberg University Mainz. PD and PS belong to the Department of Sports Medicine, Rehabilitation and Disease Prevention, Faculty of Social Science, Media and Sports, Johannes Gutenberg-University Mainz, Germany. PS is trained in internal medicine, expert in sports medicine and an active member of the National and World Anti Doping Agency (NADA and WADA). RU belongs to the Department of Psychology, University of Tbingen, Germany; he is trained in mathematical psychology and s tatistics and is a member of WADAs doping prevalence research group.

Acknowledgements
Caroline McFarlane and Teresina Tassone-Steiger, both medical students, contributed to the study by distributing questionnaires during the conferences. The authors want to thank Prof. Dr. William Keenan for proofreading the manuscript. Financial project funding: German ministry of Research and Education (BMBF) No. 01GP0807 (2009 2011). The BMBF had no influence on the content of this manuscript. Chemosensitivity and tumor metastasis are two primary issues in cancer management. Cancer cells often exhibit a wide range of sensitivity to anti-cancer compounds. To gain insight on the genetic mechanism of drug sensitivity, one powerful approach is to employ the panel of 60 human cancer cell lines developed by the National Cancer Institute (NCI). Cancer cells also show a broad range of invasion ability. However, a genome-wide portrait on the contributing molecular factors to invasion heterogeneity is lacking.

Methods
Our lab performed an invasion assay on the NCI-60 panel. We identified invasion-associated (IA) genes by correlating our invasion profiling data with the Affymetrix gene expression data on NCI-60. We then employed the recently released chemosensitivity data of 99 anti-cancer drugs of known mechanism to investigate the gene-drug correlation, focusing on the IA genes. Afterwards, we collected data from four independent drug-testing experiments to validate our findings on compound response prediction. Finally, we obtained published clinical and molecular data from two recent adjuvant chemotherapy cohorts, one on lung cancer and one on breast cancer, to test the performance of our gene signature for patient outcome prediction.

Results

First, we found 633 IA genes from the invasion-gene expression correlation study. Then, for each of the 99 drugs, we obtained a subset of IA genes whose expression levels correlated with drug-sensitivity profiles. We identified a set of eight genes ( EGFR, ITGA3, MYLK, RAI14, AHNAK, GLS, IL32 and NNMT) showing significant gene-drug correlation with paclitaxel, docetaxel, erlotinib, everolimus and dasatinib. This eight-gene signature (derived from NCI-60) for chemosensitivity prediction was validated by a total of 107 independent drug tests on 78 tumor cell lines, most of which were outside of the NCI-60 panel. The eight-gene signature predicted relapse-free survival for the lung and breast cancer patients (log-rank P = 0.0263; 0.00021). Multivariate Cox regression yielded a hazard ratio of our signature of 5.33 (95% CI = 1.76 to 16.1) and 1.81 (95% CI = 1.19 to 2.76) respectively. The eight-gene signature features the cancer hallmark epidermal growth factor receptor (EGFR) and genes involved in cell adhesion, migration, invasion, tumor growth and progression.

Conclusions
Our study sheds light on the intricate three-way interplay among gene expression, invasion and compound-sensitivity. We report the finding of a unique signature that predicts chemotherapy survival for both lung and breast cancer. Augmenting the NCI-60 model with in vitrocharacterization of important phenotype-like invasion potential is a cost-effective approach to power the genomic chemosensitivity analysis.

Keywords:
NCI-60; Invasion; Metastasis; Microarray; Chemotherapy

Background
Gene-expression profiling of tumors from patient cohorts has been used to develop gene signatures for clinical outcome prediction. Recently, a signature combining estrogen receptor (ER) status and predicted chemo/endocrine responsiveness succeeded in identifying patients with high probability of survival following taxane and anthracycline chemotherapy [1]. However, the biological mechanism of the genes involved in such cohort-initiated genomic predictors may not be easy to elucidate. On the other hand, in vitro chemosensitivity experiments on cancer cell lines, such as the NCI-60 cell line panel, are helpful in elucidating the complex relationship between drug responsiveness and gene expression. Despite this, the more challenging problem of how to translate the elucidated relationship for clinical outcome prediction still awaits more studies. In addition to chemosensitivity, metastasis is another major issue in studying treatment efficacy for many cancers, including invasive breast and lung cancer. Like the drug-sensitivity heterogeneity, tumor cells often exhibit a wide range of invasion ability. Such invasion heterogeneity may exist not only between different cancer types, but also among the individual cells from the same malignant neoplasm of a patient. More subtly, it is possible that the clinical outcome of chemotherapy may hinge on the growthinhibition of the more invasive cells rather than the less invasive cells in the neoplasm. However, a characterization of genes associated with both the invasion potential and drug-sensitivity heterogeneity is lacking. There are four aims in this study: (a) Molecular markers of tumor invasion potential: to identify the set of IA genes whose expression levels are likely to be indicative of the invasion potential of a tumor; (b) Drug sensitivity prediction by tumor-invasion markers: to evaluate how the expression levels of IA genes in a tumor are likely to be indicative of the tumors resistance or responsiveness to an anti-cancer drug; (c) Drug discovery with predictable sensitivity: to find anti-cancer drugs whose efficacies correlate with tumor-invasion potential and can be predicted by tumor-invasion markers; (d) Clinical validation: to demonstrate the use of the IA gene signature for predicting clinical outcome. NCI-60 is a diverse panel of 60 cell lines used by the Development Therapeutics Program (DTP) of the National Cancer Institute (NCI) to screen more than 100,000 compounds since 1990 [2-5]. These human cancer cell lines are derived from patients with leukemia, melanoma, lung cancer, colon cancer, central nervous system cancer, ovarian cancer, renal cancer, breast cancer and prostate cancers. The molecular characteristics of these cell lines have been subjected to various DNA microarray studies using both Affymetrix (Santa Clara, CA, USA) and spotted cDNA array technology [2]. Both drug sensitivity and gene expression profiles of the NCI-60 panel are available from the public domain; however, there is no public invasion phenotype data for NCI-60. So we conducted the invasion assay for 53 solid tumor cell lines from the NCI-60 panel in our lab.

We then conducted a series of statistical analysis to combine information from invasion profiling, gene expression and compoundsensitivity profiling. Figure1 outlines how we approached each of our four aims. We identified a set of 633 IA genes as the likely marker candidates of tumor invasion potential. For each of the 99 anticancer compounds of known mechanism, we studied the genedrug correlation to identify IA genes that can be predictive of a cells responsiveness to the compound. A final set of eight IA genes for chemosensitivity prediction on five selected compounds was obtained. We then validated the gene signature with additional cell lines. To show the clinical relevance of our finding, we searched for published chemotherapy clinical cohorts with related regimens to test the performance of our gene signature. We found two recent studies containing anti-microtubule chemotherapy, one on lung cancer and the other on breast cancer. Our signature succeeded in predicting the clinical outcome for both cohorts.

Figure 1. A schematic diagram illustrating the design of this study.

Methods
Matrigel invasion assay
We purchased the panel of NCI human cancer cell lines (NCI-60) to conduct tumor invasion assays. The suspension leukemia cancer cell lines were excluded. All cells were grown in tissue culture flasks at 37C in 5% CO 2 in RPMI 1640 with 2 mM L-glutamine, and 10% fetal bovine serum, all from Invitrogen, Eugene, OR. The invasion capacities for cell lines were examined by using membrane invasion culture system. The HTS FluoroBlok inserts containing 8-m pores (Falcon, Becton Dickinson, Franklin Lakes, NJ) were coated 30 g matrigel (BD Biosciences, San Jose, CA, USA). The cells were suspended in RPMI containing 10% FBS and seeded into the upper wells of the chamber (2.5 104 cells/well). After incubating for 24 hours at 37C, the membrane of the transwell was fixed for 10 minutes at room temperature with methanol and stained for 30 minutes with 50 g/ml propidium iodide (Sigma, St. Louis, MO). The number of cells in each blot was counted under a microscope with Analytical Imaging Station system (Imaging Research Inc., St. Catharines, ON, Canada). For each cell line, we reported the average invasion cell count after three repeats (n = 3).

Cell line gene expression data

Two gene expression datasets for NCI-60 cell lines, produced by Gene Logic (Gaithersburg, MD, USA) using Affymetrix U95, and U133plus2, respectively, were downloaded from the DTP [6]. Probe mapping between U133plus2 and U95 was provided by Affymetrix. The independent gene expression data of 78 cell lines used in the drug response validation were obtained either from the GEO website, GSE6569 (n = 23, breast cancer) [7], GSE9633 (n = 16, prostate cancer) [8], and GSE4127 (n = 29, lung cancer) [9] or by e-mail request (n = 10, lung, Balko et al.) [10]. All but nine cell lines were outside of the NCI-60 panel (Additional file 1: Table S1). Additional file 1. Supplementary information. Procedures for determining the invasion-associated (IA) genes. II. Combination use of anti-microtubule and targeted therapy agents.Figure S1. Histogram of invaded cell counts (ICC) after subtracting the tissue-group means.Figure S2. Cell adhesion: Integrin-mediated cell adhesion and migration pathway. Figure S3. The heatmap for the correlations between 744 IA gene expression and 99 drug respons e (logGI50) in NCI-60 cell lines. Figure S4. Validation of microarray gene expression data with qPCR. Figure S5. Plots of the eight-gene risk scores between drug sensitive and drug resistant groups of cell lines after removing the nine cell lines that came from the NCI60 panel. The dotted line indicated the mean of each group. Figure S6. KaplanMeier survival curves for survival analysis of the eight-gene signature in breast and lung cancer patients who did not receive systemic treatment. Figure S7. KaplanMeier survival curves for survival analysis of the four-gene signature in lung cancer and breast cancer cohorts. Table S1. The expression level of eight signature genes in the nine NCI-60 cell lines. Table S2.TaqMan probes ID for eight gene expression validation. Table S3. Enrichment analysis of 633 invasion-associated genes by functional ontology enrichment tool in MetaCore. Table S4.The correlation matrix for (A) anti-microtubule and for (B) targeted therapy drugs. Table S5.Genes having significant gene-drug correlation with everolimus, dasatinib, erlotinib, paclitaxel and docetaxel profiles. The final list of eight IA-genes is shown in bold face. Table S6.Correlation between the invasion profile and each of the 99 drug sensitivity profiles of NCI60 cell lines. The significant correlation (P <0.05) is shown in bold face.

Format: DOC Size: 1.6MB Download file This file can be viewed with: Microsoft Word Viewer

Chemosensitivity data
We downloaded the NCI-60 chemosensitivity data for anti-cancer drugs with known molecular mechanisms, including those used in targeted therapy [5]. Drug sensitivity was measured by the negative of log10 GI50 (logGI50). Seven compounds, inactive in all cell lines, were excluded. A total of 99 drugs were analyzed. The drug sensitivity in the 78 independent cell lines was measured either by -logGI50 value [9] or by the negative of log10 half maximum inhibitory concentration (IC50) value [7,8,10].

The real-time reverse transcription quantitative PCR analysis

The expressions of eight signature genes and a control gene TBP in nine lung cancer cell lines of NCI-60 were measured by reverse transcription qPCR(RT-qPCR) with specific Taqman probes and primer sets (Additional file 1: Table S2). The transcripts were amplified with Taqman One-Step RT-PCR Master Mix Reagent (Applied Biosystems) and a detection system (ABI Prism 7900HT, Applied Biosystems). Gene expression was quantified in relation to the expression of the control gene with the use of sequence detector software and the relative quantification method (Applied Biosystems).

Statistical analysis
The comparison of the tumor invasion ability between tissue types was performed by ANOVA. We applied the normal score transformation to preprocess the gene expression data before computing the Pearson correlations. The invasion-associated genes were obtained by computing the Pearson's correlation coefficients between gene expression profiles and invasion ability profile in NCI-60 cell lines. Clustering was used to order the IA probes. Students t-test was used in validating the drug sensitivity prediction of gene signatures. All statistical analyses were performed in the R language environment [11].

Cancer chemotherapy cohorts

Two chemotherapy cohorts were used in this study. The lung cancer cohort came from JBR.10, a randomized controlled trial of adjuvant vinorelbine/cisplatin versus observation alone [12]. We used the adjuvant cisplatin/vinorelbine therapy arm where after surgery (n = 71). The breast cancer cohort of 508 patients came from a prospective multicenter study, conducted at the M. D. Anderson Cancer Center, of which the patients were those with newly diagnosed Human Epidermal Growth Factor Receptor 2 (HER2)negative breast cancer under chemotherapy containing sequential taxane and anthracyclinebased regimens (followed by endocrine therapy if ER positive) [1].

Control cohorts
Four cohorts where the patients were systemically untreated after surgery were used. Three of them were the breast cancer cohorts (GSE2034 n = 286 [13], GSE7390 n = 198 [14] and GSE11121 n = 200 [15]). The fourth, a lung cancer cohort, was the OBS arm where patients were under observation after surgery (n = 62).

Survival analysis
We calculated the patients risk scores from the eight IA genes and classified them into the high -risk or the low-risk groups with the mean of risk score as the threshold value. We calculated the eight-gene risk score for each patient via simple averaging:

Risk score=(GEGFR+GITGA3+GNNMT+GMYLK+GIL32+GGLS+GAHNAK+GRAI14)/8,
Kaplan-Meier survival curves were obtained and compared by log-rank tests. Multivariate Cox proportional hazard regression analysis was used to evaluate independent prognostic factors, such as age, gender, tumor stage nodal status and histological grade.

Results
Invasion heterogeneity
The matrigel invasion assay of the 53 NCI-60 solid tumor cell lines shows a great variation between different cell lines, with the invaded cell counts (ICC) ranging from 129 to 5,514 (Figure2). According to the tissue origins, the 53 cancer cell lines are classified into eight groups: melanoma (ME), lung cancer (LC), colon cancer (CO), central nervous system cancer (CNS), ovarian cancer (OV), renal cancer (RE), breast cancer (BR) and prostate cancer (PR) [16-18]. Even within the same tissue group, there are substantial differences between the ICCs of different cell lines. Additional file 1: Figure S1 gives the overall distribution for the deviation of an individual cell lines ICC from its group mean. The wide range of this distribution sh ows that the within-group variation means are greater than the between-group variation means. This suggests that the tissue of origin may not be an essential factor in characterizing invasion heterogeneity between cancer cell lines. We tested the equality of the ICC group means by ANOVA and found that the significance barely passes the 5% mark ( P = 0.048).

Figure 2. Invasion profiling of NCI60 cancer cell lines. Cell lines are divided into groups by their tissue origin. Each dot in each tissue group gives the invaded cell counts by the matrigel invasion assay for one cell line (n = 3). Dotted lines indicate the mean of invasion cell counts for all cell lines in each tissue group.

Identification of 633 IA genes

We identified the IA genes from Affymetrix U133plus2 NCI-60, and U95 NCI-60 microarray gene expression datasets by a twostage procedure which required that (a) the gene expression profile must be significantly correlated with the invasion profile in both datasets, and (b) the sign of correlation must be consistent in both datasets (see the Additional file 1, Supplementary information). A total of 744 probes, which represented 633 distinct genes, were obtained. We estimated the false discovery rate (FDR) at the confirmation stage to be 0.08 (FDR = (2,417*0.025)/744; where 2,417 = total number of probes considered at the confirmation stage and 0.025 = the P-value cutoff with the sign consistency criterion). We show the expression levels of the 744 probes (Affy U133plus2) in the 53 cell lines with a heat map (Figure 3). The ordering of cell lines in this figure is based on their invasion abilities, the higher ones being on the left. The ordering of the genes is the output of the hierarchical clustering algorithm applied to the 744 expression profiles. The top panel of genes contains 341 probes that correlate negatively with the invasive ability, while the bottom panel of 403 probes correlates positively with the invasion ability.

Figure 3. Heat map for the expression of IA genes. Cell lines are ordered according to the invasion ability (measured by ICC) with the highest ICC placed leftmost. The genes in the top panel have negative correlations with invasion while the genes in the bottom panel have positive correlations.

Functional enrichment analysis of IA genes

We use the MetaCore (Thomson Reuters, New York, NY), a web-based computational platform designed for system biology and drug discovery, to conduct functional enrichment analysis of IA genes. We input the set of 633 IA genes (744 probes) and used the Functional Ontology Enrichment tool with the default settings. The results show that our IA genes were enriched in cell adhesion and cytoskeleton remodeling pathways (Additional file 1: Table S3). Additional file 1: Table S3 list the significantly enriched pathways and networks ( P <0.0001). The results indicate the involvement of cell adhesion and cytoskeleton remolding actin cytoskeleton network ( ACTN1, ACTN4, ATCB, ACTG1, ZYX, VCL and CFL), integrin signaling (ITGA3, ITGB1,CAV1 and CAV2), matrix metalloproteinase signaling (CDH2, CD44, TGFB2, TGFBR2, JAK1, SMAD1and SMAD3), microtubule cytoskeleton (TUBB and TUBB6), and myosin signaling (MYLK1, LIMK1,MYL6 and MYH9). We further examined the Integrin-mediated cell adhesion and migration pathway, (P = 5.49E-8, input/total nodes = 11/48), which has a much higher ratio of root to total nodes and significant P-value. The Integrin-mediated cell adhesion and migration pathway shows that ITGA3 promotes focal adhesion kinase ( FAK) autophosphorylation and creates a binding site for c-Src. EGFR signaling also activates the FAK/Src pathway [19,20]. FAK activation regulates theERK-mediated phosphorylation and activation of Myosin light chain kinase (MYLK) contributes to cell-matrix adhesion dynamics [21]. Integrin recruits FAK and a cytoskeletal protein vinculin and alpha-actin to focal contacts. c-Src and ERK2-mediated phosphorylation of FAK1 promotes its release from focal contacts and ERK2-mediated phosphorylation of paxillin promotes the association of nonphosphorylated FAK1 with paxillin at new or growing focal contact sites [22]. Activation of MYLK together with inactivation of PAK1 contributes to cell-matrix adhesion dynamics.PAK1 phosphorylation leads to the activation of LIM-kinase 1 (LIMK1) [23], inhibition of MYLK, activation of myosin regulatory light chains (MRLC) [24]. FAK is a tyrosine kinase which interacts with the important oncogene c-Src. FAK signaling is important for integrin regulated cell adhesion and migration. Notably, this pathway is the target of dasatinib [25], a drug we address further in this paper. A simplified version of the Integrin-mediated cell adhesion and migration pathway is shown Additional file 1: Figure S2, which features three IA genes, EGFR, ITGA3, MYLK.

Association of drug sensitivity with IA gene expression

To evaluate how the abundance of the IA gene transcripts in a cancer cell correlates with the cells response to each of the 99 anticancer drugs for which the drug-sensitivity profiles on NCI-60 were available, we computed Pearson's correlation between the gene expression profile of each IA gene and the chemosensitivity profile of each compound. A positive correlation means that cell lines

with higher gene expression are more sensitive to the drug, while a negative correlation indicates the opposite. The gene-drug correlation results are shown in a heat map (Additional file 1: Figure S3A). We identified all the significant correlations ( P <0.05), and displayed the findings in Additional file 1: Figure S3B. An IA probe showing a significant correlation with a drug would be called a drug-sensitivity-correlated probe for that drug. We counted the number of drug-sensitivity-correlated IA probes for each compound. A higher count indicates the availability of more IA genes for the sensitivity prediction of that drug. Such a drug would be more likely to have differential anti-cancer effects among the tumors showing differential invasion potential. Drugs with higher counts are this studys primary interest. Individually speaking, the compound with the highest count is zoledronic acid, a bisphosphonate drug used to prevent skeletal fractures in cancer patients and to treat osteoporosis. Interestingly, recent studies indicate zoledronic acid can prevent skeletal metastases through inhibition of invasion and angiogenesis of cancer cells [26,27]. However, we were unable to find the drug sensitivity validation and gene expression data outside the NCI panel for zoledronic acid. To investigate Additional file 1: Figure S3B further, we present each drugs count of the drug-sensitivity-correlated IA probes by grouping according to the drugs action mechanism (Figure4A).

Figure 4. The distribution of drug-sensitivity-correlated IA probes. (A) The number of IA probes with significant (P <0.05) gene-drug correlation with each anti-cancer compound is plotted according to the grouping of the drug mechanism. The dotted line gives the mean of the probe counts in each group. (B) The numbers of significant IA probes in tubulinbinding agents and targeted therapy agents. We find that on average, the group of tubulin-binding agents ranks the highest, followed by the targeted therapy. The count for each compound in these two groups is given in Figure4B. We continued our study on the 17 drugs from these two groups.

Gene-drug heat map for 17 drugs

We represent the 744 by 17 gene-drug correlations as a heat map after clustering the genes (Figure5A). We kept the ordering of compounds to be the same as in Figure5B. Interestingly enough, we found that the top three compounds in the group of target therapy agents, everolimus, dasatinib and erlotinib, showed a color pattern almost completely opposite to the tubulin-binding agents (Figure5). This means that cells with higher expressions of genes, such as MYB and TOB1, tend to be more sensitive to tubulinbinding agents but they also tend to be more resistant to the three-target therapy drugs. Likewise, cells with lower expression of genes, such as EGFR andITGA3, tend to be more resistant to tubulin-binding agents but they also tend to be more sensitive to the three-target therapy drugs. In other words, the efficacy of the two groups of compounds tends to be in the opposite direction. The clustering pattern in Figure5A can also be easily detected from the column-to-column correlations (Additional file 1: Table S4). The correlations among everolimus, dasatinib and erlotinib are much higher than other correlations in the target agents. Similarly, the correlation between paclitaxel and docetaxel is much higher than other correlations in the tubulin-binding agents.

Figure 5. Heatmap of gene-drug correlation. (A) Heatmap showing the gene-drug correlations for tubulinbinding and targeted therapy agents. (B) The specific pattern for the eight-gene signature enlarged from A. Blue, negative correlation; red, positive correlation. Dasatinib is a clinically studied SRC inhibitor for cancer therapy [28]. Our functional enrichment analysis of IA genes shows that the Src signaling pathway played an important role in integrin-mediated cell adhesion and migration pathway (Additional file 1: Figure

S2). Erlotinib inhibits the intracellular phosphorylation of tyrosine kinase associated with the EGFR [29]. Everolimus is an mTOR inhibitor [30]. In the anti-tubulin group, docetaxel is a semisynthetic side chain analog of paclitaxel.

Selection of drug-sensitivity-correlated IA genes for five compounds

In addition to the mechanism sharing and the clustering pattern observed, we also noticed that the numbers of sensitivity-correlated IA genes that these five compounds had are higher than other drugs in the respective mechanism groups (Figur e4A). This prompted us to ask if these five compounds might share any drug-sensitivity-correlated IA genes that could be used for drugsensitivity prediction. By inspecting the drug-sensitivity-correlated IA gene list for these five compounds, 26 common probes are identified (Additional file1: Table S5). Among them, 19 probes are in the Affy HG U-133A chip and 7 probes are in the HG U-133B chip. Because most microarray data are available in the public domain used only the U-133A chip, we exclude the seven U133-B probes. We use the standard deviation (SD) of the expression across all 53 cell lines to further exclude probes showing low expression variation. Only those probes with SD ranking within the top 10% are retained. Finally, we check the sign of correlation between gene expression and drug sensitivity to ensure the consistency in predicting sensitivity or resistance. A final set of eight IA genes (EGFR, AHNAK, GLS, IL32, ITGA3, MYLK, NNMT, and RAI14) are obtained. We confirm the quality of the microarray gene expression data by performing the qPCR assay for these eight genes on nine lung cancer cell lines in NCI-60 (Additional file 1: Figure S4). A total of 72 data points are plotted, each dot representing the expression level of a gene measured by qPCR (horizontal axis) and the microarray (vertical axis). A significant positive trend (Pearson correlation = 0.69, 2.129e-11) is observed, confirming the consistency of the two assays.

Biological functions of eight invasion-associated genes

The eight IA genes we identified are associated with cellular cytoskeleton, cell invasion and oncogenetic signaling, including the well-known cancer driver EGFR, and FAK-Src signaling (ITGA3, MYLK); ITGA3 (integrin, alpha 3) mediates cell survival and invasion through FAK-Src signaling[31-33]. MYLK is a key target of FAKSrc signaling [34]. Glutaminase (GLS) was shown to be up-regulated in MYC induction cancer cells and inhibition of GLS decreased the cancer cell growth[35,36]. Retinoic acid induced 14 (RAI14) is an actin cytoskeleton protein regulated by retinoic acid[37]. High expression of interleukin 32 (IL32) was shown to cause a worse clinical outcome in lung cancer patients [38]. AHNAK nucleoprotein was reported as pseudopod-specific proteins in different metastatic human tumor cell lines [39]. The activity of nicotinamide N-methyltransferase (NNMT) in catalyzing the Nmethylation of nicotinamide is important for biotransformation of drug and xenobiotic compounds [40]. NNMT was identified as a tumor biomarker [41-43], promoting cell migration [44] and cell invasion [45].

Validating gene signature with independent cell lines

We used the simple averaging to combine the expressions of eight IA genes into an eight-gene score. We searched the public domain extensively for drug response experiments involving any of the five featured compounds. Four experiments were found. In the first two experiments [9], two standard anti-microtubule agents, paclitaxel and docetaxel, were applied separately to 29 lung cell lines, which were then divided into drug-resistant and drug-sensitive groups according to the IC50 value of the experiment outcome. From the gene expression data, we calculated the eight-gene score for each cell line and found a significant difference between the drugresistant group and the drug-sensitive group (t-test P = 0.009 in paclitaxel-treated group, and P = 0.013 in docetaxel-treated group, Figure6). In the third experiment [8], dasatinib was applied to 16 prostate and 23 breast cancer cell lines [7], which were then divided into dasatinib-sensitive and dasatinib-resistant groups according to the experiment outcome. The eight-gene scores show a significant difference between the dasatinib-sensitive group and dasatinib-resistant group (t-test P = 0.0044 for breast cancer and P = 0.0044 for prostate cancer, Figure6). The fourth experiment applied erlotinib to 10 lung cancer cell lines. Again, the difference between erlotinib-sensitive and erlotinib-resistant was significant (t- test P = 0.025, Figure6). It should be noted, however, that these plots also showed an overlapping pattern in the distributions between the drug-resistant group and drug-sensitive group. This indicates that the eight gene biomarkers still cannot accurately classify cell lines into a resistant group versus a sensitive group.

Figure 6. Plots of the eight-gene risk scores between drug-sensitive and drug-resistant groups of cell lines. The dotted line indicated the mean of each group.

In these validation experiments, we found nine cell lines from the NCI-60 panels. We removed these cell lines from the validation data and conducted the comparison again (Additional file 1: Figure S5). All significant differences in the first three experiments were still valid. For the fourth experiment, the P-value increased to 0.0788, mainly because the sample size (n = 8) was too small.

Clinical outcome prediction in adjuvant chemotherapy lung and breast cancer patients
To test if the eight genes have the predictive power in clinical outcome evaluation, we searched the public domain for chemotherapy cohort studies involving any of the two anti-microtubule compounds we used in obtaining the eight genes. Two recently published studies on the vinorelbine containing regimen for lung cancer [12] and the taxane-containing regimen for treating invasive breast cancer [1] were obtained. We use the simple average expression of the eight-genes as the risk score to dichotomize patients into two groups. We found that the low-risk group has a significantly longer relapse-free survival time than the high-risk group in the lung cancer cohorts (P = 0.0263, Figure7A) and similarly, the low-risk group shows significantly longer distant relapse-free survival in the breast cancer cohorts (P = 0.00021, Figure7B).

Figure 7. KaplanMeier survival curves for survival analysis of the eight-predicted genes in (A) lung cancer and (B) breast cancer cohorts. We further conducted multivariate Cox proportional hazard regression analysis with our gene signature and other prognostic factors (including age and tumor stage) as the predictors. The result shows that the effect of our gene signature is still significant. The adjusted hazard ratio (HR) is 5.33 (P = 0.003) for the lung cancer cohort, and 1.81 (P = 0.006) for the breast cancer patients (Table1). This shows that our eight-gene signature is an independent predictor for patient outcome. To demonstrate that the predicative capacity of the eight-gene signature is specific to chemotherapy, we applied it to four control cohorts (three for breast cancer and one for lung cancer) of which the patients were systemically untreated after surgery. The eight gene signature failed to predict clinical outcome (Additional file 1: Figure S6). Table 1. Multivariate Cox regression analysis of the eight-gene signature for predicting relapse-free survival in cancer patients

Discussion
The invasive or metastatic potential of a malignant neoplasm and the growth-inhibition of tumor cells by a therapeutic agent are two common denominators of patient survival in cancer systemic therapy. While cell line models have been used to predict treatment response or patient survival by genes associated with drug sensitivity in the cancer cell lines [46-49], these studies have not considered the varying metastasis potential of a tumor. To weigh in the interplay of both factors directly, our pre-clinical gene signature discovery method features the co-integration of invasion phenotypes and compound-sensitity profiles with gene expression at the full genome scale. Our approach is based on the observation that invasion ability and drug sensitivity are both phenotypes of the cell lines available for study. Each phenotype is naturally associated with its own set of molecular determinants. We hypothesize the potential overlap between the set for invasion ability phenotype and the set for drug responsiveness phenotype. By identifying these common determinants, then we may use these shared determinants to estimate the overall invasion potential of the cancer cells in a tumor and also use it to predict the drug response at the same time. However, because the tumor microenvironment in a patient is different from the growth environment of cancer cell lines monitored in a lab, the robustness of an invasion molecular marker becomes an important factor for increasing the chance of success in clinical applications. An alternative strategy of analyzing the three-way interaction of invasion, gene expression and drug response would be to correlate invasion with drug response first. Once the most correlated drugs are identified, then genes correlated with response to these drugs can be used to predict the drug sensitivity. However, we did not pursue this line of analysis further because the phenotye-phenotype

correlation is often weaker than phenotype-genomic determinant correlation. As a matter of fact, our data show that most druginvasion correlations appear weak; only four drugs pass the statistical significance and the best two correlations are only 0.39 and 0.35 (Additional file 1: Table S6). Our approach overcomes the limitations of weak phenotype-phenotype correlation by looking for statistical evidence of correlations directly from the genomic determinant. This helps improve the robustness of the genetic marker thus obtained. Previously, without considering drug sensitivity, our team performed invasion profiling for the nine lung cancer cell lines of NCI-60 to obtain a four gene signature for clinical outcome prediction [50]. We find that among the four genes ANKRD49 and LPHN1 are in the IA gene list and only ANKRD49has a significant correlation with paclitaxel and docetaxel. The four-gene signature failed to predict the survival outcome for the two validation cohorts receiving adjuvant chemotherapy (Additional file 1: Figure S7). To gain robustness of our gene signature, instead of using different panels of tissue origins in NCI-60 to obtain different sets of IA genes for different types of cancer, we used the invasion data from all 53 NCI60 solid tumor cell lines and obtained 633 IA genes. Then a series of statistical analyses were designed to increase the robustness of the final eight-gene signature in predicting drug sensitivity for the selected compounds. The eight-gene score differed between drug-resistant and drug-sensitive cell lines (Figure6). We succeeded in applying our gene signature to one lung cancer cohort and one breast cancer cohort, of which the patients received a regimen containing an anti-microtubule agent. The success in using the same signature to predict patient outcome for different types of cancer showed the robustness of this gene signature. The eight-gene signature showed a positive correlation with the sensitivity of targeted therapy compounds and a negative correlation with the sensitivity of anti-microtubule compounds (Figure5B). Because high values of the eight-gene signature correlate with high invasion potential in cancer cells, this suggests that the direct correlation between the invasion profile and the sensitivity profiles of anti-MT drugs may be negative. This is indeed the case, but the correlation is weak (Additional file 1: Table S6, correlation = 0.16, -0.28 for paclitaxel and docetaxel, respectively). On the other hand, the correlation between the eight-gene score and the drug sensitivity is stronger (0.41, -0.54, respectively). Similarly, the correlation between the eight-gene score and the sensitivity for erlotinib, dasatinib and everolimus is 0.46, 0.52, 0.44, respectively, which is again stronger than the correlation between invasion and drug sensitivity (0.26, 0.24, 0.09, respectively). Therefore, despite the weak correlation between the invasion phenotype and the drug sensitivity phenotype, the eight-gene signature is an effective genomic marker for invasion potential and it can be used to predict the drugs differential growth-inhibition efficacy that varies between cancer cells of higher invasion potential and those of lower invasion potential. When applying to the patients tumor specimen, the eight-gene signature provides an averaged profile of the gene expression by individual cells with varying invasion potential. A low eight-gene score indicates that the overall invasion potential of the tumor is low and the chance of the patients favorable response to regimens containing anti -microtubule compounds increases. On the other hand, a high eight-gene score predicts the abundance of the cells of higher invasion potential, which are harder to eradicate by antimicrotubule compounds, but may be more likely to succumb to the said targeted therapy. This suggests the combined use of targeted therapy like dasatinib or erlotinib with anti-microtubule agents to increase the regimen efficacy of chemotherapy alone. There have been several studies on augmenting the anticancer effect of chemotherapy with targeted therapy. Erlotinib was shown to be more sensitive in the doxorubicin-resistant human breast cancer cell lines and paclitaxel-resistant human ovarian cancer cell lines[51] and the sensitivity was positively correlated with EGFR expression. More references were provided in Additional file 1, Supplementary information Text II. There is room to improve our eight-gene signature for drug-sensitivity prediction. The overlap in distribution between the drugsensitivity group and the drug-resistant group (Figure6) suggests that drug response in cell lines is a very complex phenotype which is not fully characterized by our gene signature. Other genomic components, such as DNA copy number, single-nucleotide polymorphisms, methylation and microRNA, have not been considered in our study. In addition, differences in lab environment may also contribute to the variations observed in the data.

Conclusions
We have shown that augmenting the NCI-60 model with in vitro characterization of important phenotypes like invasion potential is a cost-effective approach to power the genomic chemosensitivity analysis. Our analysis delineates the complex three-way interplay of gene expression, cancer cells invasion potential and cancer cells responsiveness to an anti -cancer compound. We report the identification of a unique eight-gene signature for both lung and breast cancer, which predicted the relapse-free survival of adjuvant chemotherapy patients. The signature features the cancer hallmark EGFR and genes involved in cell adhesion cell migration, cell invasion, tumor growth and tumor progression. The discovery of prognostic biomarkers for chemotherapy patients remains critical toward improving the efficacy of cancer treatment. The eight-gene signature obtained here may be useful for the development of individualized cancer therapy. Our method of gene discovery may be applicable in studying other cancers.

Abbreviations

AHNAK: AHNAK nucleoprotein; Br: Breast cancer; CNS: Central nervous system; CO: Colon cancer; DTP: Development therapeutics program; EGFR: Epidermal growth factor receptor; ER: Estrogen receptor; FAK: Focal adhesion kinase; FDR: False discovery rate; GLS: Glutaminase; HR: Hazard ratio; IA: Invasion-associated; ICC: Invaded cell counts; IL32: Interleukin 32; ITGA3: Integrin, alpha 3; LC: Lung cancer; LIMK1: LIM-kinase 1; ME: Melanoma; MRLC: Myosin regulatory light chains; MYLK: Myosin light chain kinase; NCI: National cancer institute; NNMT: N-methyltransferase; OV: Ovarian cancer; PR: Prostate cancer; RE: Renal cancer.

Competing interests
The authors declare that they have no competing interests.

Authors contributions
YCH carried out the invasion experiments, conceived the drug sensitivity and clinical prediction study design, and wrote the paper. YCH, HYC, SY, CHL, GW and KCL collected public gene expression and statistical data analysis. SLY participated in setting up the experiment platform for the cell line invasion and qPCR validation. SLY, PCY and KCL provided the study materials and reagents. KCL and YCH set up the study aims and rationalized the approach. PCY and KCL helped chart the study strategy and were involved in drafting and finalizing the manuscript. KCL and YCH proposed the conceptual framework to link gene expression variation with heterogeneity of invasion potential and drug sensitivity in cancer cells. All authors read and approved the final manuscript for publication.

Acknowledgments
This research is supported by NSC 98-2314-B-001-001-M