A machine learning approach to drug repositioning based on drug expression profiles: Applications to schizophrenia and depression/anxiety disorders
11 A machine learning approach to drug repositioning based on drug expression profiles: Applications to schizophrenia and depression/anxiety disorders
Kai Zhao and Hon-Cheong So *1,2 School of Biomedical Sciences, The Chinese University of Hong Kong, Shatin, Hong Kong KIZ-CUHK Joint Laboratory of Bioresources and Molecular Research of Common Diseases, Kunming Zoology Institute of Zoology and The Chinese University of Hong Kong
Corresponding author: Hon-Cheong So. Email: [email protected]
Abstract
Development of new medications is a very lengthy and costly process. Finding novel indications for existing drugs, or drug repositioning, can serve as a useful strategy to shorten the development cycle. In this study, we present an approach to drug discovery or repositioning by predicting indication for a particular disease based on expression profiles of drugs, with a focus on applications in psychiatry. Drugs that are not originally indicated for the disease but with high predicted probabilities serve as good candidates for repurposing. This framework is widely applicable to any chemicals or drugs with expression profiles measured, even if the drug targets are unknown. It is also highly flexible as virtually any supervised learning algorithms can be used.
We applied this approach to identify repositioning opportunities for schizophrenia as well as depression and anxiety disorders. We applied various state-of-the-art machine learning (ML) approaches for prediction, including deep neural networks, support vector machines (SVM), elastic net, random forest and gradient boosted machines. The performance of the five approaches did not differ substantially, with SVM slightly outperformed the others. However, methods with lower predictive accuracy can still reveal literature-supported candidates that are of different mechanisms of actions. As a further validation, we showed that the repositioning hits are enriched for psychiatric medications considered in clinical trials. Notably, many top repositioning hits are supported by previous preclinical or clinical studies. Finally, we propose that ML approaches may provide a new avenue to explore drug mechanisms via examining the variable importance of gene features.
INTRODUCTION
Development of new medications is a very lengthy and costly process. While investment in research and development has been increasing, there is a lack of proportional rise in the number of drugs approved in the past two decades, especially for drugs with novel mechanisms of actions . There is an urgent need for innovative approaches to improve the productivity of drug development. This is particularly true for some areas like psychiatry, for which there has been lack of therapeutic advances for some time . Finding new indications for existing drugs, an approach known as drug repositioning or repurposing, can serve as a useful strategy to shorten the development cycle . Repurposed drugs can be brought to the market in a much shorter time-frame and at lower costs. With the exponential growth of “omics” and other biomedical data in recent years, computational drug repositioning provides a fast, cost-effective and systematic way to identify promising repositioning opportunities . In this study we describe a general drug repositioning approach by predicting drug indications based on their expression profiles, with a focus on applications in psychiatry. We treat drug repositioning as a supervised learning problem and apply different state-of-the-art machine learning methods for prediction. Drugs that are not originally indicated for the disease but have high predicted probabilities serve as good candidates for repositioning. There are several advantages of this approach. Firstly, the presented approach is a general and broad framework that leverages machine learning (ML) methodologies, a field with very rapid advances in the last decade. This provides great flexibility and opportunities for further improvement in the future as virtually any supervised learning methods can be applied. Newly developed prediction algorithms can also be readily incorporated to improve the detection of useful drug candidates. In addition, the method described here is widely applicable to any chemical or drugs with expression profiles recorded, even if the drug targets or mechanisms of actions are unknown. For example, herbal medicine products may contain a mixture of ingredients with uncertain drug targets; even for many known medications [ e.g. lithium ], their mechanisms of actions and exact targets are not completely known. If transcriptomic profiling has been performed, they can still be analyzed for therapeutic or repositioning potential under the current approach. There has been increasing interest in computational drug repositioning recently, in view of the rising cost of new drug development. Hodos et al. provided a comprehensive and updated review on this topic. Similarity-based methods ( e.g. ref . ) represent one common approach, but as noted by Hodos et al., the dependence on data in the “nearby pharmacological space” might limit the ability to find medications with novel mechanisms of actions. Another related methodology is the network-based approach , which typically requires data on the relationship between drugs, genes and diseases as well as connections within each category (e.g. drug-drug similarities). It can integrate different sources of information but may still be constrained by the focus on a nearby pharmacological space and the choice of tuning parameters in network construction or inference is often ad hoc . The present work is different in that we offer a broad framework for repositioning and we do not focus on one but many different kinds of learning methods. There is comparatively less reliance on known drug mechanisms or the “nearby pharmacological space” as we let the different algorithms “learn” the relationship between drugs, genes and disease in their own ways . We note that kernel-based ML methods such as support vector machine (SVM) are also based on some sort of similarity measures, but here we considered a variety of other approaches and SVM is one of the methods which falls under the broader framework of ML for repositioning. For high-throughput omics data, often only a subset of genes or input features are relevant and a machine learning approach is able to “learn” which features to consider for repositioning. As we shall discuss later, ML approaches also provide a new avenue to explore the mechanisms of different drug classes, by studying the variable importance of input genes. We are particularly interested in drug repositioning for psychiatric disorders in view of the lack of novel treatments in the area. It is worth noting that while psychiatric disorders are leading causes of disability worldwide , there have been limited advances in the development of new pharmacological agents in the last two decades or so . Development of new therapies is also limited by the difficulty of animal models to fully mimic human psychiatric conditions . Investment by drug companies has in general been declining , and new approaches for drug discoveries are very much needed in this field. We will explore repositioning opportunities for schizophrenia along with depression and anxiety disorders. Here depression and anxiety disorders are analyzed together as they are highly clinically comorbid , show significant genetic correlations , and share similar pharmacological treatments . Contributions of this study are summarized below. Firstly, we presented a general approach to drug repositioning based on ML methods, leveraging drug expression profiles as predictors. While previous work has also proposed the use of ML on drug transcriptome profiles for classifying drugs into groups (e.g. anti-cancer drugs, cardiovascular drugs, drugs acting on the central nervous system etc.), we focused on drug repositioning for particular diseases instead of predicting the big therapeutic groups. Secondly, we have performed a comparison of the predictive performances of five state-of-the-art and perhaps most commonly employed ML algorithms, including deep neural networks, support vector machines, elastic net, random forest and gradient boosted trees. Thirdly, we identified new repositioning opportunities for schizophrenia and depression/anxiety disorders, and validated the relevance of the repositioned drugs by showing their enrichment among drugs considered for clinical trials, as well as support by previous literature. We also found that methods with slightly lower predictive accuracy still reveal literature-supported repositioning candidates that are of different mechanisms of actions from the drugs listed by better-performing algorithms. Finally, we explored which genes and pathways contributed the most to our predictions, hence shedding light on the molecular mechanisms underlying the actions of antipsychotics and antidepressants. METHODS
We present a general drug repositioning approach adopting a supervised learning approach. We construct prediction models in which the outcome is defined as whether the drug is a known treatment for the disease, and the predictors are expression profiles of each drug. Drugs that are not originally known to treat the disease but have high predicted probabilities are regarded as good candidates for repositioning.
Drug expression data
The expression data is downloaded from the Connectivity Map (CMap), which captures transcriptomic changes when thee cell lines (HL60, PC3, MCF7) were treated with a drug or chemical . We downloaded raw expression data from Cmap, and performed normalization with the MAS5 algorithm . Expression levels of genes represented on more than one probe sets were averaged. We employed the limma package to perform analyses on differential expression between treated cell lines and controls. Analyses were performed on each combination of drug and cell line, with a total of 3478 instances. Expression measurements were available for 12436 genes. Statistical analyses were performed in R3.2.1 with the help of the R package “longevityTools”. Defining drug indications
Drug indications were extracted from two known resources, namely the Anatomical Therapeutic Chemical (ATC) classification system and the MEDication Indication Resource high precision subset (MEDI-HPS) . We focus on schizophrenia as well as depression and anxiety disorders in this study. From the ATC classification system, two groups of drugs were extracted, including antipsychotics and antidepressants. On the other hand, the MEDI-HPS dataset integrates four public medication resources, including RxNorm, Side Effect Resource 2 (SIDER2) , Wikipedia and MedlinePlus. We used the MEDI high-precision subset (MEDI-HPS) which only include drug indications found in RxNorm or in at least 2 out of 3 other sources . This subset achieves a precision of up to 92% according to Wei et al. . Machine learning methods
We employed different state-of-the-art machine learning approaches including deep neural networks (DNN) , support vector machine (SVM) , random forest (RF) , gradient boosted machine with trees (GBM) and logistic regression with elastic net regularization (EN) to predict indications. Our data is imbalanced as only a minority of the drugs are indicated for schizophrenia or depression/anxiety disorders. We performed both unweighted and weighted analyses in this study; in the weighted analysis, class weights are adjusted such that the minority group (drugs indicated for the disorder) will receive higher weight to achieve a balance between positive and negative instances. In our unweighted model, DNN was implemented in the python package keras. Hyperparameters were chosen by the “fmin” optimization algorithm from “hyperopt”, which employs a sequential model-based optimization approach . The tree-structured Parzen estimator (TPE) was used. The more sophisticated hyper-parameter search strategies provided by sequential model-based methods may produce better results than simpler approaches (e.g. grid search) when the number of hyper-parameters is large, such as in deep learning settings . Fifty evaluations were run for each search of optimal hyper-parameters. Dropout and mixed L1/L2 penalties were employed to reduce over-fitting. The neural networks consisted of two or three layers, with number of nodes selected uniformly from the range [64, 1024]. Dropout percentage was selected uniformly from [0.25, 0.75], and L1/L2 penalty uniformly from [1E-5, 1E-3]. Optimizer was chosen from “adadelta” , “adam” and “rmsprop” , and the activation function chosen from “relu”, “softplus” or “tanh”. One hundred epochs were run for each model and we extracted the model weights corresponding to the best epoch. We also attempted ‘hyperopt’ for the weighted analysis, however the predictive performance was unexpectedly poor due to unclear reasons yet to be revealed. We therefore turned to hyperparameter selection with grid search, with some adjustments in the parameter ranges. A two-layer neural network was constructed with dropout and mixed L1/L2 to avoid over-fitting. The number of neurons in the first hidden layer was selected from {1000, 1500, 2000}, the number in the second layer from {500, 1000, 1500}, dropout rate from {0.4, 0.5, SVM, RF and GBM models were implemented in “scikit-learn” (sklearn) in python. Hyper-parameter selection was performed by the built-in function GridSearchCV in sklearn. For SVM, we chose radial basis function as the kernel. The two hyper-parameters C and gamma were selected from [-5, 2] and [-6, 2] in log10-space respectively. For RF, the number of bagged trees was set to 1000, the maximum number of features used for splitting was selected from {800, 1000, 1500, 2000, 3000, 5000} and min_samples_leaf (the minimum number of samples required at a leaf node) was selected from {1, 3, 5, 10, 30, 50, 80}. For GBM, the number of boosting iterations was selected from a sequence of 100 to 1001 with step size 50, learning rate from {0.005, 0.01, 0.015, 0.02, 0.03, 0.05}, subsampling proportion from {0.8, 1}, maximum depth of each estimator from {2, 3, 5, 10} and maximum number of features from {10, 30, 50, 100, 500, 1000}. Finally, the EN model was implemented by the R package “glmnet” . The elastic-net penalty parameter α was chosen from seq(0, 1, by=0.1), with other settings following the default. Nested Cross-validation
We adopted nested three-fold cross validation (CV) to choose hyper-parameters and evaluate model performances. It has been observed that optimistic bias will result if one uses simple CV to compute an error estimate for a prediction algorithm that itself is tuned using CV . Nested CV avoids this problem and is able to give an almost unbiased estimate of prediction accuracy . The inner loop CVs were used to choose the parameters that optimized predictive performance. In each outer loop CV we made predictions on the corresponding test set using the best model trained from the inner CV loops. To achieve maximum consistency in our comparisons, we compared different methods on the same test set in each loop. Note that the test sets were not involved in model training or parameter tuning. Predictive performance measures
The performances of the machine learning methods were evaluated in the test sets using three metrics, including log loss, area under the receiver operating characteristic curve (ROC-AUC) and area under the precision recall curve (PR-AUC). Log loss compares the predicted probabilities against the true labels. The receiver operating characteristic curve, which plots the sensitivity ( i.e. recall) against (1- specificity), is a very widely used approach to evaluate predictive performances in biomedical applications. The precision-recall curve on the hand plots the precision ( i.e. positive predictive value) against the sensitivity (recall). Since precision depends on the overall proportion of positive labels, the PR-AUC is also dependent on such proportions. Davies et al. suggested that the PR curve may give more informative comparisons when working with imbalanced data. Identifying important genes and pathways
We also performed analyses to reveal the genes which contribute the most to the prediction model. For elastic net, we extracted the genes with non-zero coefficient in at least one cross-validation fold, and the resulting genes were subject to an over-representation analysis (ORA) (using hypergeometric tests) to reveal the pathways involved. For RF and GBM, feature importance was computed using built-in functions in sklearn based on Gini importance (i.e. the average decrease in node impurity). We then performed a gene-set enrichment analysis (GSEA ) based on the genes together with their respective importance scores (the highest score across three folds was taken). For SVM and DNN, there is a lack of widely adopted importance measures, so we focused on the rest of ML methods in this part. Pathway analyses were conducted by the web-based program WebGastalt . Four pathway databases were considered in our analyses, including KEGG, PANTHER, Reactome and Wikipathways. External validation by testing for enrichment of psychiatric drugs considered for clinical trials
We then performed additional analyses to assess if the drugs with high predicted probabilities from our machine learning models are indeed good candidates for repositioning. Briefly, we tested whether the drugs with no known indication for the disease but high predicted probabilities are more likely to be included in clinical trials. In the first step, we filtered off drugs that are known to be indicated for the disease as derived from ATC and MEDI-HPS. This is because we are mainly interested in repositioning other drugs of unknown therapeutic potential, and that the labels of drug indication (from ATC or MEDI-HPS) have already been utilized in the ML prediction steps. Next, we extracted a list of drugs that were included in clinical trials for schizophrenia as well as depression and anxiety disorders. The list was derived from clincialTrial.gov and we downloaded a pre-complied version from https://doi.org/10.15363/thinklab.d212. We then tested for enrichment of those drugs listed in clinicalTrial.gov among the top repositioning results. We performed an enrichment analysis of “drug-sets”, similar to a gene-set analysis approach widely used in bioinformatics . We performed one-tailed t -tests to assess if the predicted probabilities (derived from machine learning models) are significantly higher for psychiatric drugs considered in clinical trials. RESULTS Predictive performance comparison
Unweighted analysis
The average predictive performances (averaged over three folds) of different machine learning methods are listed in Table 1. When considering log loss as the criteria of interest, SVM gave the best result overall, though EN showed the best performance in one of the four datasets. DNN and EN showed quite similar predictive performances. RF and GBM were slightly worse than other methods, but the difference was small. When ROC-AUC was considered as the performance metric, SVM and EN gave similar performances. SVM outperformed EN in the schizophrenia datasets, while EN showed better results in the other two datasets. The performance of DNN was worse than that of SVM and EN, although the differences were not large. The two tree-based methods performed worse especially in the depression/anxiety datasets. We then considered PR-AUC, which is more sensitive to imbalanced data, as the measure of predictive performance. SVM was the best-performing method. EN and DNN followed with very similar performances. Consistent with other performance measures, GBM and RF did not perform as well in the depression/anxiety datasets, but the performance was more comparable for the schizophrenia datasets.
Weighted analysis
Compared with unweighted analysis, we observed improvements in predictive performance for several methods including GBM, RF and deep learning. SVM and EN performed similarly in general. Considering ROC-AUC, deep learning performed the best for depression and anxiety disorders, while RF and EN showed highest ROC-AUC for schizophrenia. SVM achieved the best PR-AUC and log-loss compared to other ML approaches.
Enrichment for psychiatric drugs considered in clinical trials
We further tested whether the top repositioning results are enriched for drugs included in clinical trials for psychiatric disorders. As shown in Table 2, we observed significant enrichment of such drugs for both schizophrenia and depression/anxiety disorders across all methods in the weighted analysis. In addition, most results in the unweighted analysis were also significant. This external validation provides further support to the usefulness of our approach in identifying new repositioning opportunities.
Identifying contributing genes and pathways
Supplementary Tables 1-4 show the top genes as identified by variable importance measures (for RF and GBM) and regression coefficients (for EN). The enriched pathways are shown in Table 3 and Supplementary tables 5-12. Since the number of genes involved is large, we only highlighted a few top enriched pathways here. Interestingly, steroid and cholesterol biosynthesis are among the most significantly enriched pathways for drugs against schizophrenia and depression/anxiety. Notably, abnormalities in the hypothalamic-pituitary-adrenal (HPA) axis have long been suggested as one of the key pathological mechanisms underlying depression . The steroid (cortisol) synthesis inhibitor metyrapone has been shown to be effective for depression in a double-blind randomized controlled trial (RCT) and other studies (reviewed in ref. ), although another trial failed to show any benefits . Antidepressants have also been shown to regulate glucocorticoid receptor functioning in vivo . On the other hand, neuroactive steroids may be implicated in the pathophysiology of schizophrenia . Cholesterol biosynthesis, including regulation by sterol regulatory element-binding protein (SREBP), was frequently top-listed in our pathway analysis. Antipsychotics and some antidepressants are associated with metabolic syndrome and weight gain, and previous in vitro and in vivo studies have shown lipogenic effects of these drugs as controlled by SREBP transcription factors . Interestingly, some studies showed lower cholesterol may be associated with suicidality , depressive symptoms , and poorer cognition in schizophrenia , but these findings are controversial. Whether pathways related to cholesterol synthesis may play a role in the therapeutic effects of psychotropic drugs remain a topic for further investigation. Some other pathways are also worth mentioning. For example, IGF signaling pathway was significantly enriched under antidepressants. IGF-I has been reported to improve depression and anxiety symptoms in clinical samples , and showed antidepressant-like effects in animal models (e.g. ref. ). The 5-HT signaling pathway was also top-listed under antipsychotics. 5-HT has been proposed as a new drug target and improvements in negative and cognitive symptoms have been reported in clinical trials . Top repositioning hits and literature support from previous studies
Table 4 show some of the selected top repositioning candidates with literature support, which will also be discussed below. Tables 5 and 6 show the top 15 repositioning hits by DNN and SVM respectively. Predictions with the weighted ML models were presented here due to better overall predictive performances. More detailed tables showing the top 100 hits for each ML method in both unweighted and weighted analyses are presented in Supplementary Tables 13-16. Note that drugs that are known to be indicated for these disorders according to ATC or MEDI-HPS were excluded from the lists. We noted overlap in top hits derived from different machine learning methods, but some repositioning candidates are unique to one or few ML approaches. This suggests that employing a diverse set of ML methods may be advantageous in “learning” different potential repositioning candidates. We will chiefly focus on candidates from the top 15 hits for each
ML method in the exposition below.
Repositioning candidates for depression/anxiety disorders
Regarding depression and anxiety disorders, many of the top results are antipsychotics, such as trifluoperazine, perphenazine, fluphenazine and thioridazine, among others. Antipsychotics have long been used for the treatment for depression . In earlier studies, phenothiazines (a class of antipsychotic to which many of our top hits belong) was observed to produce similar anti-depressive effects as tricyclic antidepressants . Due to the risk of extra-pyramidal side-effects, typical antipsychotics are less commonly used these days and second-generation (atypical) antipsychotics are more often prescribed. Meta-analyses have shown that atypical antipsychotics are effective as adjunctive or primary treatment for depression . Antipsychotics are also commonly prescribed for severe depressive episodes with psychotic symptoms. A few other drugs on the lists are also worth mentioning. Cyproheptadine (top-listed by SVM, RF, GBM and EN) is a 5-HT receptor antagonist and was shown to improve depression in a small cross-over trial . It was also reported that the drug reduced the neuropsychiatric side-effects of the antiviral therapy efavirenz, including depressive and anxiety symptoms . Chlorcyclizine belongs to the phenylpiperazine class and numerous antidepressants and antipsychotics also belong to this class . Pizotifen, listed by EN, is a 5-HT / antagonist which was shown to possess antidepressant effects in a double-blind RCT . DNN and EN have identified histone deacetylase (HDAC) inhibitors including trichostatin A and vorinostat as top repositioning hits for depression/anxiety and schizophrenia. HDAC have been implicated in the pathogenesis of psychiatric disorders including depression, as reviewed by Fuchikami et al. . HDAC inhibitors have been reported to produce antidepressant-like effects in animal models , although no clinical trials on psychiatric disorders were available. Interestingly, in a recent study which employed gene-set analysis on de novo mutations to uncover repositioning opportunities, HDAC inhibitors were highlighted as candidates for schizophrenia and other neurodevelopmental disorders . Another candidate was tetrandrine, a calcium channel blocker top-listed by DNN, RF, GBM and SVM. Tetrandrine demonstrated antidepressant-like effects in mice in forced swimming and tail suspension tests. The drug also increased the concentration of 5-hydroytrytamine (5-HT) and norepinephrine in mice treated with reserpine or chromic mild stress, and raised the levels of brain-derived neurotrophic factor (BDNF) in the latter case . Repositioning candidates for schizophrenia
With regards to repositioning results for schizophrenia, some of the hits are antidepressants, such as protriptyline, maprotiline and clomipramine, among others. Antidepressants are frequently prescribed in schizophrenia patients due to possibility of comorbid depression or obsessive-compulsive disorder . In meta-analyses antidepressants were also found to improve negative symptoms of schizophrenia . For the antidepressants on the list, maprotiline (listed by RF, GBM, EN, SVM) has been reported to improve negative symptoms in chronic schizophrenia patients as an adjunctive treatment. Other drug clomipramine (listed by DNN, GBM, EN) has been shown to ameliorate not only obsessive-compulsive but also overall schizophrenic symptoms in patients with comorbid disorders . Interestingly, the mood stabilizer valproate was also listed among the top (by DNN and SVM). Valproate may improve clinical response when added to antipsychotics, although the evidence is mainly based on open RCTs . The EN algorithm also “re-discovered” spiperone, an antipsychotic not listed in ATC or MEDI-HPS, as one of the top repositioning hits. Several other drugs less well-known for psychiatric disorders are also worth mentioning. The selective estrogen receptor modulator raloxifene (listed by DNN and EN) was shown to improve schizophrenia symptom scores in double-blind RCTs of post-menopausal women . Another drug nordihydroguaiaretic acid (listed by DNN, GBM, SVM) has antioxidant properties and may be useful in combating oxidative stress in schizophrenia . Pioglitazone, top-ranked by DNN, belongs to the class of thiazolidinediones and has anti-diabetic and anti-inflammatory properties. Although this drug was withdrawn due to unexpected adverse effects on the liver, our finding suggested that other thiazolidinediones may be useful for schizophrenia. Indeed, another drug of the same class known as pioglitazone has been shown to improve negative symptoms in schizophrenia patients in a double-blind RCT . Another RCT also showed improvements in depressive symptoms . Tretinoin (listed by DNN) is a retinoid and retinoid dysfunction has been linked to schizophrenia . Clinical trials with another retinoid (bexarotene) showed some benefits of the drug as an add-on agent in schizophrenia. Again retinoid signaling was implicated for schizophrenia in a recent study on drug repositioning leveraging de novo mutations . Felodipine (listed by GBM) is a calcium channel blocker and GWAS on schizophrenia and bipolar disorder have revealed many genes related to calcium channels ; a recent study also suggested concomitant use of CCB and antipsychotics may be more beneficial than antipsychotics alone . Some hits from the unweighted analysis
The top repositioning candidates from un weighted analysis for each ML method are listed in supplementary tables 13-16. There were a number of overlaps with the candidates from the weighted analysis. Here we highlight a few prioritized drugs (that have not been mentioned above) with literature support. Aspirin (acetylsalicylic acid) is a non-steroidal anti-inflammatory agent (listed by SVM), which has been shown to improve schizophrenia symptoms in a recent meta-analysis of RCTs . Genistein is a phytoestrogen and can bind to estrogen receptors . An animal study showed that genistein may possess anti-dopaminergic actions ; interestingly, clinical studies have shown potential therapeutic benefits of estrogens on schizophrenia . The EN algorithm identified metformin as one of the top repositioning hits for depression/anxiety disorders. A study in Taiwan reported that the risk of depression in diabetic patients was reduced by ~60% for those given metformin with sulfonylurea . Another study reported improved depressive symptoms and cognitive functions for patients with comorbid diabetes and depression . Another drug apigenin, top-listed by GBM, was supported by a number of in vitro and animal studies for possible antidepressant-like and anxiolytic effects . A clinical trial of oral chamomile (which was standardized to contain 1.2% of apigenin) showed benefits for anxiety and depression . Discussions
In this study, we have presented a repositioning approach by predicting drug indications based on expression profiles. We employed and compared five state-of-the-art machine learning methods to perform predictions. We also observed that the top repositioning hits are enriched for psychiatric drugs considered for clinical trials and that many hits are backed up by evidence form animal or clinical studies, supporting the validity of our approach. Concerning the performance of different machine learning classifiers, we have employed five methods in total, and all but one (EN) are non-linear classifiers. SVM is a kernel-based learning approach that is widely used in bioinformatics. On the other hand, deep learning methods (such as DNN) that are based on the principles of representation learning have witnessed rapid advances in the last decade, especially in the field of computer vision. While potentially powerful, current successful applications typically require very large datasets for training, and we suspect that the relatively modest sample size ( N = 3478) of our dataset may have limited DNN to achieve the optimal predictive ability. We have used at most two hidden layers in view of the moderate sample size, and the complexity of the network may be increased with larger samples, although larger samples would lead to greater computational costs. This study shows that deep learning can achieve reasonable performance in drug repurposing, and indeed DNN achieved the best ROC-AUC for depression/anxiety disorders in the weighted analysis. Given the rapid growth in the area, deep learning approaches might be worthy of further investigations. While logistic regression with EN is a linear classifier, it performed well overall though lagging behind SVM. The performances of the two tree-based methods (RF and GBM) were largely comparable with other methods in the weighted analysis, although they were less satisfactory without weighting. Notwithstanding the differences in predictive performances, different algorithms are based on diverse model assumptions and principles, and as shown above, methods with slightly lower predictive accuracy may still reveal useful repositioning candidates that are of different mechanisms of actions. To the best of our knowledge, this is the first study to employ a comprehensive array of machine learning methods on drug expression profiles for drug repositioning of any particular disease; it is also the first application in psychopharmacology. This is also the first work to suggest an ML approach to explore the molecular mechanisms underlying drug actions. In a related work, Aliper et al. made use of the drug transcriptome to predict drug classes e.g. drugs for neurological diseases, drugs for cardiovascular diseases, anti-cancer agents etc. . Here our focus is different and perhaps clinically more relevant in that we directly identify repositioning opportunities for a particular disorder. In addition, we concentrated on the study of psychiatric disorders, which was not explicitly considered in Aliper et al. or other previous works. Interestingly, DNN was reported to be the best performing method in their study. However their study and the present work are not directly comparable as the outcomes studied are different and the evaluation metrics also differ. F1 score was used in Aliper et al. (although the choice of a cut-off probability for classification was not explicitly stated) while we used ROC-AUC, PR-AUC and log loss as performance indicators. Here we aim to provide a proof-of-concept example showing that the application of machine learning methodologies on drug expression profiles may help to identify candidates for repositioning, particularly for psychiatric disorders. The approach is intuitive and also highly flexible. Nevertheless, given the variety of methods and rapid advances in machine learning and computational drug repurposing, there is still room for improvement. Firstly, we only consider drug indications and the drug-induced transcriptomic changes in our prediction model. This makes the method very flexible and widely applicable to any compounds or drugs for which an expression profile is available. The use of drug transcriptome evades the need of specifying targets and knowing the mechanisms of actions, and the approach may even be applicable to a mixture of chemical ingredients as may be the case for herbal medicines. However, it is possible that our methods may be further improved by incorporating other information such as drug targets and chemical structure, if such information is available. As for the prediction algorithm, in our dataset the number of positive labels is small. We tackled this problem by adjusting the weighting of positive and negative instances and indeed found improvements for several ML approaches. Other methodologies to account for imbalanced data are also possible , and this may be a topic for further explorations. We covered five commonly used algorithms here but this coverage is obviously not complete; further studies may benefit from the use of more advanced or recently developed learning methods. We also notice that there is an ongoing effort to expand the coverage of CMap , and that the study with updated full data and documentations have just been released. We are planning to further explore the current framework in the expanded dataset. It is reassuring to observe that many repositioning hits are supported by previous studies and the results enriched for psychiatric drugs considered in clinical trials. However, we stress that further well-designed pre-clinical and clinical studies are necessary before the any results can be brought into clinical practice. We have also made use of ML methods to explore potentially important genes and pathways that may contribute to treatment effects. However, the results require further experimental validations. Computational approaches for repositioning and explorations of drug mechanisms, such as ML-based methods, provide a cost-effective and systematic way to assess and prioritize drug candidates, and might help to reduce the high failure rates in drug development. Given the rising cost in developing a new drug, even a minute reduction in failure rate will represent large savings in absolute terms. Further work might involve combining the current approach with other computational and experimental methods to further improve the accuracy of drug repositioning. It is widely acknowledged that drug development in psychiatry has become stagnant for some years, and that traditional approaches to drug discovery has not been as successful as anticipated. On the other hand, the past few years have seen an extremely rapid development in ML methods and applications; in this regard, we hope that this study will open a new avenue for drug repositioning/discovery, and stimulate further research to bridge the gap between ML and biomedical applications especially drug development. The list of repositioning candidates might also serve as a useful resource for researchers and clinicians working on schizophrenia as well as depression and anxiety disorders, which are illnesses very much in need of new therapies. Supplementary Information is available at https://drive.google.com/open?id=1TyYXm4Vtcc0jB76xHzGeLKonax2ixy4a Acknowledgements
This work is partially supported by the Lo-Kwee Seong Biomedical Research Fund and a Direct Grant from the Chinese University of Hong Kong. We thank Mr. Carlos Kwan-long Chau for help in pathway analysis, and Mr. AO Fu-Kiu Kelvin and Mr. WONG Yui-Hang Harris for assistance in literature search. We also thank Professor Stephen K.W. Tsui and the Hong Kong Bioinformatics Centre for computing support.
Financial Disclosures
The authors declare no competing financial interests. Table 1 Average predictive performance of different machine learning models across four datasets in un weighted (top) and weighted analysis (bottom) Unweighted analysis
Average Log Loss
MEDI-HPS Depression/Anxiety ATC antidepressants MEDI-HPS Schizophrenia ATC antipsychotics
SVM
EN 0.1249
Average ROC-AUC
SVM
EN 0.725
Average PR-AUC
SVM
EN 0.1372 0.1008 0.1586 0.2087 DNN 0.1447 0.0877 0.1577 0.2156 GBM 0.0910 0.0417 0.1426 0.1528 RF 0.1193 0.0639 0.1677 0.1703
Weighted analysis
Average Log Loss
MEDI-HPS Depression/Anxiety ATC antidepressants MEDI-HPS Schizophrenia ATC antipsychotics
SVM
EN 0.5803 0.5344 0.5028 0.5112 DNN 0.1309 0.0990 0.1308 0.1098 GBM 0.1281 0.1032 0.1114 0.0981 RF 0.1234 0.0943 0.1060 0.0939
Average ROC-AUC
SVM 0.7198 0.7718 0.7731 0.7765 EN 0.6610 0.7394 0.7494
DNN Average PR-AUC
SVM
EN 0.0751 0.0896 0.1520 0.2030 DNN 0.1796 0.1107 0.2278 0.2641 GBM 0.1800 0.1168 0.2697 0.2780 RF 0.1771 0.1165 0.2721 0.2707
The learning algorithm with the best performance in each dataset (for each predictive performance measure) is marked in bold. ROC-AUC: area under the curve (AUC) of the receiver operating characteristic (ROC) curve; PR-AUC: Average area under the curve (AUC) of the precision-recall (PR) curve. SVM: support vector machines; EN: logistic regression with elastic net regularization; DNN: deep neural networks; RF: random forest; GBM, gradient boosted machines with trees. MEDI-HPS: MEDication Indication - High Precision Subset; ATC: Anatomical Therapeutic Chemical classification. The drug indications (treated as the outcome variable in prediction) are defined for two kinds of disorders (schizophrenia as well as depression and anxiety disorders) using two different sources (MEDI-HPS and ATC), hence a total of four datasets. Please refer to the main text for details. Table 2 Enrichment for psychiatric drugs included in clinical trials among the repositioning hits
Dataset unweighted analysis weighted analysis P -value q -value P- value q- value SVM
MEDI-HPS Depression/Anxiety
ATC antidepressants
MEDI-HPS Schizophrenia
ATC antipsychotics
DNN
MEDI-HPS Depression/Anxiety
ATC antidepressants
MEDI-HPS Schizophrenia
ATC antipsychotics EN MEDI-HPS Depression/Anxiety
ATC antidepressants
MEDI-HPS Schizophrenia
ATC antipsychotics
GBM
MEDI-HPS Depression/Anxiety
ATC antidepressants
MEDI-HPS Schizophrenia
ATC antipsychotics RF MEDI-HPS Depression/Anxiety
ATC antidepressants
MEDI-HPS Schizophrenia
ATC antipsychotics
P-values < 0.05 and q-values <0.05 are in bold. Table 3 Selected enriched pathways based on variable importance of genes in ML models
Method Name eNet-ORA_Reactome Cholesterol biosynthesis 7 1.78E-06 eNet-ORA_Wikipathway Sterol Regulatory Element-Binding Proteins (SREBP) signalling 8 1.73E-04 eNet-ORA_KEGG Steroid biosynthesis - Homo sapiens (human) 5 2.38E-04 gbm_KEGG Fat digestion and absorption - Homo sapiens (human) 34 8.95E-03 gbm_Panther Insulin/IGF pathway-protein kinase B signaling cascade 34 1.05E-02 rf_Wikipathway Mismatch repair 9 1.38E-01 rf_Wikipathway ID signaling pathway 16 1.58E-01 rf_Wikipathway Statin Pathway 25 1.61E-01 rf_Wikipathway Photodynamic therapy-induced HIF-1 survival signaling 37 1.63E-01 eNet-ORA_Panther TGF-beta signaling pathway 4 1.70E-01 gbm_Reactome STING mediated induction of host immune responses 11 2.33E-01 rf_KEGG Amino sugar and nucleotide sugar metabolism - Homo sapiens (human) 40 2.43E-01 rf_Wikipathway Phase I biotransformations, non P450 7 2.51E-01 rf_Wikipathway Target Of Rapamycin (TOR) Signaling 30 2.56E-01 rf_KEGG Alanine, aspartate and glutamate metabolism - Homo sapiens (human) 28 2.67E-01
ATC antipsychotics and MEDI-HPS schizophrenia (weighted analysis) rf_Wikipathway Sterol Regulatory Element-Binding Proteins (SREBP) signalling 64 0.00E+00 eNet-ORA_Reactome Cholesterol biosynthesis 7 9.33E-05 eNet-ORA_KEGG Steroid biosynthesis - Homo sapiens (human) 4 2.01E-03 rf_Wikipathway Statin Pathway 25 2.51E-02 eNet-ORA_Wikipathway Fatty Acid Beta Oxidation 5 2.98E-02 eNet-ORA_Reactome Asparagine N-linked glycosylation 15 4.36E-02 eNet-ORA_KEGG Metabolic pathways - Homo sapiens (human) 37 4.87E-02 eNet-ORA_Reactome Synthesis of UDP-N-acetyl-glucosamine 3 8.11E-02 gbm_Panther 5HT3 type receptor mediated signaling pathway 14 8.58E-02 eNet-ORA_KEGG Citrate cycle (TCA cycle) - Homo sapiens (human) 3 8.96E-02 gbm_Reactome G1/S-Specific Transcription 18 1.32E-01 eNet-ORA_Reactome Antigen Presentation: Folding, assembly and peptide loading of class I MHC 4 1.32E-01 gbm_Panther Androgen/estrogene/progesterone biosynthesis 9 1.40E-01 rf_Wikipathway Photodynamic therapy-induced unfolded protein response 23 1.42E-01 eNet-ORA_Reactome COPII (Coat Protein 2) Mediated Vesicle Transport 6 1.44E-01 We aggregated pathway analysis results from 4 databases, namely KEGG, Reactome, Panther and Wikipathways. Pathways that were highly similar were filtered. FDR, false discovery rate; eNet, elastic net; gbm, gradient boosted machine; rf, random forest; ORA, over-representation analysis, applicable to elastic net. Only results from weighted analysis are included here. Table 4 Some literature-supported candidates selected from top hits derived from machine learning methods (known antipsychotics and antidepressants are not included in this list)
Drug Method Relationship with disease
Depression/anxiety
Cyproheptadine SVM, RF, GBM, EN 5-HT2 receptor antagonist, improve depression in a small cross-over trial Chlorcyclizine DNN, RF, GBM, SVM phenylpiperazine group to which many other antidepressants and antipsychotics belong Pizotifen EN 5-HT2A/2C antagonist, positive result in an RCT TrichostatinA/Vorinostat DNN, EN HDAC inhibitors may have antidepressant effects as shown in animal models Tetrandrine DNN, RF, GBM CCB; antidepressant-like effects in mice; may increase 5-HT, NE and BDNF concentrations Apigenin GBM Antidepressant and anxiolytic effects in animal models and in an RCT Metformin EN may reduce depression risk among DM subjects
Schizophrenia
Valproate DNN, SVM open RCTs reported symptom improvement when used as adjunctive treatment Raloxifene DNN, EN improve SCZ symptoms in an RCT of post-menopausal women Nordihydroguaiaretic acid DNN, GBM, SVM antioxidant; oxidative stress implicated in SCZ Pioglitazone DNN Another drug in the same class (pioglitazone) improved SCZ symptoms in RCT Tretinoin DNN Retinoid; dysfunction in retinoid signaling may be implicated in SCZ Felodipine GBM CCB; CCB added to antipsychotics may be beneficial Aspirin SVM NSAID; may improve SCZ symptoms as shown in RC Genistein GBM Phytoestrogen; animal model show possible anti-dopaminergic effects Please refer to the main text for detailed discussions and references. As a number of top results were known antipsychotics or antidepressants (please refer to the main text for details), these were not presented in the above table. SCZ, schizophrenia; RCT, randomized controlled trial; HDAC, Histone deacetylases; CCB, calcium channel blocker; 5-HT, serotonin; NE, norepinephrine; BDNF, brain-derived neurotrophic factor; NSAID, non-steroidal anti-inflammatory drugs. Table 5 Top 15 repositioning candidates derived from the deep learning approach
ATC antidepressants
ATC antipsychotics
Drug_cell-line Pred_prob
Drug_cell-line Pred_prob 1 homochlorcyclizine_MCF7 0.938
1 metixene_MCF7 0.932 2 thioridazine_PC3 0.866
2 metixene_PC3 0.926 3 pimethixene_MCF7 0.830
3 mefloquine_PC3 0.916 4 levomepromazine_PC3 0.829
4 protriptyline_MCF7 0.907 5 deptropine_PC3 0.761
5 norcyclobenzaprine_PC3 0.906 6 fluphenazine_PC3 0.756
6 methylbenzethonium.chloride_MCF7 0.895 7 trifluoperazine_PC3 0.756
7 astemizole_MCF7 0.884 8 cyproheptadine_PC3 0.736
8 troglitazone_MCF7 0.867 9 metixene_PC3 0.706
9 clomipramine_MCF7 0.865 10 tetrandrine_MCF7 0.702
10 nortriptyline_MCF7 0.846 11 norcyclobenzaprine_MCF7 0.700
11 suloctidil_MCF7 0.835 12 fluphenazine_HL60 0.695
12 trimipramine_PC3 0.826 13 flupentixol_PC3 0.691
13 loperamide_MCF7 0.823 14 norcyclobenzaprine_PC3 0.667
14 perhexiline_MCF7 0.791 15 loperamide_MCF7 0.667
15 tetrandrine_MCF7 0.788
MEDI-HPS depression/anxiety
MEDI-HPS schizophrenia
Drug_cell-line Pred_prob
Drug_cell-line Pred_prob 1 trichostatin A_HL60 0.999
1 metixene_PC3 0.945 2 fluphenazine_MCF7 0.864
2 valproic.acid_MCF7 0.941 3 promazine_PC3 0.840
3 nordihydroguaiaretic.acid_MCF7 0.939 4 metixene_PC3 0.817
4 ciclosporin_MCF7 0.926 5 orphenadrine_MCF7 0.791
5 suloctidil_MCF7 0.922 6 piperacetazine_PC3 0.757
6 tetrandrine_MCF7 0.908 7 homochlorcyclizine_MCF7 0.741
7 desipramine_PC3 0.892 8 fluphenazine_HL60 0.700
8 loperamide_MCF7 0.887 9 thioridazine_PC3 0.694
9 tretinoin_MCF7 0.874 10 tetrandrine_MCF7 0.682
10 homochlorcyclizine_MCF7 0.864 11 fluphenazine_PC3 0.654
11 trimipramine_MCF7 0.857 12 norcyclobenzaprine_PC3 0.609
12 loperamide_PC3 0.838 13 clomifene_PC3 0.600
13 profenamine_PC3 0.829 14 levomepromazine_PC3 0.595
14 raloxifene_MCF7 0.799 15 haloperidol_MCF7 0.580
15 thiethylperazine_PC3 0.795 Table 6 Top 15 repositioning candidates derived from support vector machine
ATC antidepressants
ATC antipsychotics
Drug_cell-line Pred_prob
Drug_cell-line Pred_ prob 1 homochlorcyclizine_MCF7 0.764
1 norcyclobenzaprine_MCF7 0.753 2 cyproheptadine_MCF7 0.556
2 protriptyline_MCF7 0.667 3 thioridazine_MCF7 0.420
3 methylbenzethonium.chloride_ PC3 0.611 4 trifluoperazine_MCF7 0.418
4 astemizole_PC3 0.539 5 chlorcyclizine_PC3 0.383
5 metixene_MCF7 0.524 6 levomepromazine_PC3 0.339
6 metixene_PC3 0.517 7 prochlorperazine_MCF7 0.323
7 nordihydroguaiaretic.acid_HL60 0.488 8 spiperone_PC3 0.311
8 loperamide_MCF7 0.474 9 perphenazine_MCF7 0.310
9 suloctidil_MCF7 0.451 10 pimethixene_MCF7 0.300
10 norcyclobenzaprine_PC3 0.430 11 trifluoperazine_PC3 0.300
11 maprotiline_PC3 0.416 12 CP-690334-01_PC3 0.264
12 amitriptyline_PC3 0.391 13 tetrandrine_MCF7 0.250
13 quinisocaine_PC3 0.366 14 loperamide_MCF7 0.232
14 mefloquine_PC3 0.364 15 chlorcyclizine_MCF7 0.228
15 nortriptyline_MCF7 0.341
MEDI-HPS depression/anxiety
MEDI-HPS schizophrenia
Drug_cell-line Pred_prob
Drug_cell-line Pred_ prob 1 fluphenazine_MCF7 0.931
1 nordihydroguaiaretic.acid_HL60 0.646 2 thioridazine_PC3 0.780
2 nordihydroguaiaretic.acid_MCF7 0.615 3 fluphenazine_HL60 0.780
3 methylbenzethonium.chloride_ PC3 0.574 4 homochlorcyclizine_MCF7 0.669
4 quinisocaine_PC3 0.550 5 fluphenazine_PC3 0.455
5 nordihydroguaiaretic.acid_PC3 0.500 6 cyproheptadine_MCF7 0.446
6 norcyclobenzaprine_MCF7 0.420 7 perphenazine_MCF7 0.422
7 niclosamide_PC3 0.419 8 nordihydroguaiaretic acid_HL60 0.356
8 metixene_PC3 0.404 9 thioridazine_HL60 0.345
9 valproic.acid_PC3 0.404 10 haloperidol_MCF7 0.331
10 norcyclobenzaprine_PC3 0.401 11 tetrandrine_MCF7 0.310
11 suloctidil_MCF7 0.395 12 nordihydroguaiaretic acid_PC3 0.310
12 nortriptyline_MCF7 0.375 13 metixene_PC3 0.308
13 genistein_PC3 0.372 14 cyproheptadine_PC3 0.291
14 amitriptyline_MCF7 0.362 15 metitepine_MCF7 0.287
15 protriptyline_MCF7 0.362
9 0
References
1. Pammolli, F., Magazzini, L. & Riccaboni, M. The productivity crisis in pharmaceutical R&D.
Nat Rev Drug Discov , 428-38 (2011). 2. Krystal, J.H. & State, M.W. Psychiatric disorders: diagnosis to therapy. Cell , 201-14 (2014). 3. Hyman, S.E. Psychiatric drug development: diagnosing a crisis.
Cerebrum , 5 (2013). 4. Dudley, J.T., Deshpande, T. & Butte, A.J. Exploiting drug-disease relationships for computational drug repositioning.
Brief Bioinform , 303-11 (2011). 5. Malhi, G.S., Tanious, M., Das, P., Coulston, C.M. & Berk, M. Potential mechanisms of action of lithium in bipolar disorder. Current understanding. CNS Drugs , 135-53 (2013). 6. Hodos, R.A., Kidd, B.A., Shameer, K., Readhead, B.P. & Dudley, J.T. In silico methods for drug repurposing and pharmacology. Wiley Interdiscip Rev Syst Biol Med , 186-210 (2016). 7. Gottlieb, A., Stein, G.Y., Ruppin, E. & Sharan, R. PREDICT: a method for inferring novel drug indications with application to personalized medicine. Mol Syst Biol , 496 (2011). 8. Oh, M., Ahn, J. & Yoon, Y. A network-based classification model for deriving novel drug-disease associations and assessing their molecular actions. PLoS One , e111668 (2014). 9. Liu, Z. et al. Similarity-based prediction for Anatomical Therapeutic Chemical classification of drugs by integrating multiple data sources.
Bioinformatics , 1788-95 (2015). 10. Luo, H. et al. Drug repositioning based on comprehensive similarity measures and Bi-Random walk algorithm.
Bioinformatics , 2664-71 (2016). 11. Napolitano, F. et al. Drug repositioning: a machine-learning approach through data integration.
Journal of Cheminformatics (2013). 12. Li, J. & Lu, Z. A new method for computational drug repositioning using drug pairwise similarity. in Bioinformatics and Biomedicine (BIBM), 2012 IEEE International Conference On
Briefings in Bioinformatics . 14. Vigo, D., Thornicroft, G. & Atun, R. Estimating the true global burden of mental illness.
Lancet Psychiatry , 171-8 (2016). 15. Nestler, E.J. & Hyman, S.E. Animal models of neuropsychiatric disorders. Nat Neurosci , 1161-9 (2010). 16. Lamers, F. et al. Comorbidity patterns of anxiety and depressive disorders in a large cohort study: the Netherlands Study of Depression and Anxiety (NESDA).
J Clin Psychiatry , 341-8 (2011). 17. Kessler, R.C. et al. Anxious and non-anxious major depressive disorder in the World Health Organization World Mental Health Surveys.
Epidemiol Psychiatr Sci , 210-26 (2015). 18. Otowa, T. et al. Meta-analysis of genome-wide association studies of anxiety disorders.
Mol Psychiatry , 1485 (2016). 19. Ballenger, J.C. Anxiety and Depression: Optimizing Treatments. Prim Care Companion J Clin Psychiatry , 71-79 (2000). 20. Aliper, A. et al. Deep Learning Applications for Predicting Pharmacological Properties of Drugs and Drug Repurposing Using Transcriptomic Data.
Mol Pharm , 2524-30 (2016). 21. Lamb, J. et al. The Connectivity Map: using gene-expression signatures to connect small molecules, genes, and disease.
Science , 1929-35 (2006). 22. Pepper, S.D., Saunders, E.K., Edwards, L.E., Wilson, C.L. & Miller, C.J. The utility of MAS5 expression summary and detection call algorithms.
BMC Bioinformatics , 273 (2007). 23. Ritchie, M.E. et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res , e47 (2015). 24. Wei, W.Q. et al. Development and evaluation of an ensemble resource linking medications to their indications.
J Am Med Inform Assoc , 954-61 (2013). 25. Kuhn, M., Letunic, I., Jensen, L.J. & Bork, P. The SIDER database of drugs and side effects. Nucleic Acids Res , D1075-9 (2016). 26. Goodfellow, I., Bengio, Y. & Courville, A. Deep learning , xxii, 775 pages. 27. Cortes, C. & Vapnik, V. Support-Vector Networks.
Machine Learning , 273-297 (1995). 28. Breiman, L. Random forests. Machine Learning , 5-32 (2001). 29. Friedman, J.H. Greedy function approximation: A gradient boosting machine. Annals of Statistics , 1189-1232 (2001). 30. Zou, H. & Hastie, T. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B-Statistical Methodology , 301-320 (2005). 31. Bergstra, J., Yamins, D. & Cox, D.D. Hyperopt: A python library for optimizing the hyperparameters of machine learning algorithms. in Proceedings of the 12th Python in Science Conference arXiv preprint arXiv:1212.5701 (2012). 33. Kingma, D. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014). 34. Hinton, G., Srivastava, N. & Swersky, K. Neural Networks for Machine Learning Lecture 6a Overview of mini- ‐ batch gradient descent. (2012). 35. Friedman, J., Hastie, T. & Tibshirani, R. Regularization Paths for Generalized Linear Models via Coordinate Descent. Journal of Statistical Software , 1-22 (2010). 36. Varma, S. & Simon, R. Bias in error estimation when using cross-validation for model selection. Bmc Bioinformatics (2006). 37. Davis, J. & Goadrich, M. The relationship between Precision-Recall and ROC curves. in Proceedings of the 23rd international conference on Machine learning et al.
Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles.
Proc Natl Acad Sci U S A , 15545-50 (2005). 39. Wang, J., Vasaikar, S., Shi, Z., Greer, M. & Zhang, B. WebGestalt 2017: a more comprehensive, powerful, flexible and interactive gene set enrichment analysis toolkit.
Nucleic Acids Res (2017). 40. de Leeuw, C.A., Neale, B.M., Heskes, T. & Posthuma, D. The statistical properties of gene-set analysis.
Nat Rev Genet , 353-64 (2016). 41. Varghese, F.P. & Brown, E.S. The Hypothalamic-Pituitary-Adrenal Axis in Major Depressive Disorder: A Brief Primer for Primary Care Physicians. Prim Care Companion J Clin Psychiatry , 151-155 (2001). 42. Jahn, H. et al. Metyrapone as additive treatment in major depression: a double-blind and placebo-controlled trial.
Arch Gen Psychiatry , 1235-44 (2004). 43. Sigalas, P.D., Garg, H., Watson, S., McAllister-Williams, R.H. & Ferrier, I.N. Metyrapone in treatment-resistant depression. Ther Adv Psychopharmacol , 139-49 (2012). 44. McAllister-Williams, R.H. et al. Antidepressant augmentation with metyrapone for treatment-resistant depression (the ADD study): a double-blind, randomised, placebo-controlled trial.
Lancet Psychiatry , 117-27 (2016). 45. Shulman, Y. & Tibbo, P.G. Neuroactive steroids in schizophrenia. Can J Psychiatry , 695-702 (2005). 46. Raeder, M.B., Ferno, J., Vik-Mo, A.O. & Steen, V.M. SREBP activation by antipsychotic- and antidepressant-drugs in cultured human liver cells: relevance for metabolic side-effects? Mol Cell Biochem , 167-73 (2006). 47. Ferno, J. et al.
Lipogenic effects of psychotropic drugs: focus on the SREBP system.
Front Biosci (Landmark Ed) , 49-60 (2011). 48. Cantarelli Mda, G., Tramontina, A.C., Leite, M.C. & Goncalves, C.A. Potential neurochemical links between cholesterol and suicidal behavior. Psychiatry Res , 745-51 (2014). 49. Persons, J.E., Robinson, J.G., Coryell, W.H., Payne, M.E. & Fiedorowicz, J.G. Longitudinal study of low serum LDL cholesterol and depressive symptom onset in postmenopause.
J Clin Psychiatry , 212-20 (2016).
50. You, H., Lu, W., Zhao, S., Hu, Z. & Zhang, J. The relationship between statins and depression: a review of the literature.
Expert Opin Pharmacother , 1467-76 (2013). 51. Krakowski, M. & Czobor, P. Cholesterol and cognition in schizophrenia: a double-blind study of patients randomized to clozapine, olanzapine and haloperidol. Schizophr Res , 27-33 (2011). 52. Thompson, J.L. et al.
Effects of human growth hormone, insulin-like growth factor I, and diet and exercise on body composition of obese postmenopausal women.
J Clin Endocrinol Metab , 1477-84 (1998). 53. Burgdorf, J. et al. Insulin-Like Growth Factor I Produces an Antidepressant-Like Effect and Elicits N-Methyl-D-Aspartate Receptor Independent Long-Term Potentiation of Synaptic Transmission in Medial Prefrontal Cortex and Hippocampus.
Int J Neuropsychopharmacol (2015). 54. Grunbaum-Novak, N. et al. Relationship between antidepressants and IGF-1 system in the brain: Possible role in cognition.
European Neuropsychopharmacology , 431-438 (2008). 55. Ellenbroek, B.A. & Prinssen, E.P. Can 5-HT3 antagonists contribute toward the treatment of schizophrenia? Behav Pharmacol , 33-44 (2015). 56. Wang, P. & Si, T. Use of antipsychotics in the treatment of depressive disorders. Shanghai Arch Psychiatry , 134-40 (2013). 57. Robertson, M.M. & Trimble, M.R. Major tranquillisers used as antidepressants. A review. J Affect Disord , 173-93 (1982). 58. Spielmans, G.I. et al. Adjunctive atypical antipsychotic treatment for major depressive disorder: a meta-analysis of depression, quality of life, and safety outcomes.
PLoS Med , e1001403 (2013). 59. Komossa, K., Depping, A.M., Gaudchau, A., Kissling, W. & Leucht, S. Second-generation antipsychotics for major depressive disorder and dysthymia. Cochrane Database Syst Rev , CD008121 (2010). 60. Greenway, S.E., Pack, A.T. & Greenway, F.L. Treatment of depression with cyproheptadine.
Pharmacotherapy , 357-60 (1995). 61. Dabaghzadeh, F. et al. Cyproheptadine for prevention of neuropsychiatric adverse effects of efavirenz: a randomized clinical trial.
AIDS patient care and STDs , 146-154 (2013). 62. Maia Rdo, C., Tesch, R. & Fraga, C.A. Phenylpiperazine derivatives: a patent review (2006 - present). Expert Opin Ther Pat , 1169-78 (2012). 63. Standal, J.E. Pizotifen as an antidepressant. Acta Psychiatr Scand , 276-9 (1977). 64. Fuchikami, M. et al. The potential use of histone deacetylase inhibitors in the treatment of depression.
Progress in Neuro-Psychopharmacology & Biological Psychiatry , 320-324 (2016). 65. Hobara, T., Uchida, S., Otsuki, K., Yamagata, H. & Watanabe, Y. Molecular mechanisms of the antidepressant actions by histone deacetylase inhibitors. Neuroscience Research , E316-E316 (2010). 66. Covington, H.E. et al. Antidepressant Actions of Histone Deacetylase Inhibitors.
Journal of Neuroscience , 11451-11460 (2009). 67. So, H.-C. & Wong, Y.-H. Implications of de novo mutations in guiding drug discovery: A study of four neuropsychiatric disorders. bioRxiv , 173641 (2017). 68. Gao, S., Cui, Y.L., Yu, C.Q., Wang, Q.S. & Zhang, Y. Tetrandrine exerts antidepressant-like effects in animal models: Role of brain-derived neurotrophic factor. Behavioural Brain Research , 79-85 (2013). 69. Mao, Y.M. & Zhang, M.D. Augmentation with antidepressants in schizophrenia treatment: benefit or risk.
Neuropsychiatr Dis Treat , 701-13 (2015). 70. Singh, S.P., Singh, V., Kar, N. & Chan, K. Efficacy of antidepressants in treating the negative symptoms of chronic schizophrenia: meta-analysis. Br J Psychiatry , 174-9 (2010). 71. Sepehry, A.A., Potvin, S., Elie, R. & Stip, E. Selective serotonin reuptake inhibitor (SSRI) add-on therapy for the negative symptoms of schizophrenia: a meta-analysis.
J Clin Psychiatry , 604-10 (2007). 72. Yamagami, S. & Soejima, K. Effect of maprotiline combined with conventional neuroleptics against negative symptoms of chronic schizophrenia. Drugs Exp Clin Res , 171-6 (1989). 73. Berman, I. et al. Treatment of obsessive-compulsive symptoms in schizophrenic patients with clomipramine.
J Clin Psychopharmacol , 206-10 (1995). 74. Wang, Y., Xia, J., Helfer, B., Li, C. & Leucht, S. Valproate for schizophrenia. Cochrane Database Syst Rev , CD004028 (2016). 75. Usall, J. et al. Raloxifene as an Adjunctive Treatment for Postmenopausal Women With Schizophrenia: A Double-Blind, Randomized, Placebo-Controlled Trial.
Journal of Clinical Psychiatry , 1552-1557 (2011). 76. Usall, J. et al. Raloxifene as an Adjunctive Treatment for Postmenopausal Women With Schizophrenia: A 24-Week Double-Blind, Randomized, Parallel, Placebo-Controlled Trial.
Schizophrenia Bulletin , 309-317 (2016). 77. Lu, J.M. et al. Molecular mechanisms and clinical applications of nordihydroguaiaretic acid (NDGA) and its derivatives: an update.
Med Sci Monit , RA93-100 (2010). 78. Wu, J.Q., Kosten, T.R. & Zhang, X.Y. Free radicals, antioxidant defense systems, and schizophrenia. Prog Neuropsychopharmacol Biol Psychiatry , 200-6 (2013). 79. Iranpour, N. et al. The effects of pioglitazone adjuvant therapy on negative symptoms of patients with chronic schizophrenia: a double-blind and placebo-controlled trial.
Hum Psychopharmacol , 103-12 (2016). 80. Smith, R.C. et al. Effects of pioglitazone on metabolic abnormalities, psychopathology, and cognitive function in schizophrenic patients treated with antipsychotic medication: a randomized double-blind study.
Schizophr Res , 18-24 (2013). 81. Goodman, A.B. Three independent lines of evidence suggest retinoids as causal to schizophrenia.
Proc Natl Acad Sci U S A , 7240-4 (1998). 82. Lerner, V., McCaffery, P.J. & Ritsner, M.S. Targeting Retinoid Receptors to Treat Schizophrenia: Rationale and Progress to Date. CNS Drugs , 269-80 (2016). 83. Heyes, S. et al. Genetic disruption of voltage-gated calcium channels in psychiatric and neurological disorders.
Prog Neurobiol , 36-54 (2015). 84. Cipriani, A. et al.
A systematic review of calcium channel antagonists in bipolar disorder and some considerations for their future development.
Mol Psychiatry , 1324-32 (2016). 85. Van Dyke, P. & Thomas, K.L. Concomitant calcium channel blocker and antipsychotic therapy in patients with schizophrenia: Efficacy analysis of the CATIE-Sz phase 1 data. Ann Clin Psychiatry , e1-e11 (2017). 86. Sommer, I.E. et al. Efficacy of anti-inflammatory agents to improve symptoms in patients with schizophrenia: an update.
Schizophr Bull , 181-91 (2014). 87. Wang, T.T., Sathyamoorthy, N. & Phang, J.M. Molecular effects of genistein on estrogen receptor mediated pathways. Carcinogenesis , 271-5 (1996). 88. Suresh, P. & Raju, A.B. Antidopaminergic effects of leucine and genistein on shizophrenic rat models. Neurosciences (Riyadh) , 235-41 (2013). 89. Wahlqvist, M.L. et al. Increased risk of affective disorders in type 2 diabetes is minimized by sulfonylurea and metformin combination: a population-based cohort study.
Bmc Medicine (2012). 90. Guo, M. et al. Metformin may produce antidepressant effects through improvement of cognitive function among depressed patients with diabetes mellitus.
Clinical and Experimental Pharmacology and Physiology , 650-656 (2014). 91. Pathak, L., Agrawal, Y. & Dhir, A. Natural polyphenols in the management of major depression. Expert Opin Investig Drugs , 863-80 (2013). 92. Amsterdam, J.D. et al. A randomized, double-blind, placebo-controlled trial of oral Matricaria recutita (chamomile) extract therapy for generalized anxiety disorder.
J Clin Psychopharmacol , 378-82 (2009). 93. Amsterdam, J.D. et al. Chamomile (Matricaria recutita) may provide antidepressant activity in anxious, depressed humans: an exploratory study.
Altern Ther Health Med , 44-9 (2012).
94. Bengio, Y., Courville, A. & Vincent, P. Representation Learning: A Review and New Perspectives.
Ieee Transactions on Pattern Analysis and Machine Intelligence , 1798-1828 (2013). 95. He, H. & Garcia, E.A. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering , 1263-1284 (2009). 96. Aravind Subramanian, R.N., Steven M. Corsello, David D. Peck, Ted E. Natoli, Xiaodong Lu, Joshua Gould, John F. Davis, Andrew A. Tubelli, Jacob K. Asiedu, David L. Lahr, Jodi E. Hirschman, Zihan Liu, Melanie Donahue, Bina Julian, Mariya Khan, David Wadden, Ian Smith, Daniel Lam, Arthur Liberzon, Courtney Toder, Mukta Bagul, Marek Orzechowski, Oana M. Enache, Frederica Piccioni, Alice H. Berger, Alykhan Shamji, Angela N. Brooks, Anita Vrcic, Corey Flynn, Jacqueline Rosains, David Takeda, Desiree Davison, Justin Lamb, Kristin Ardlie, Larson Hogstrom, Nathanael S. Gray, Paul A. Clemons, Serena Silver, Xiaoyun Wu, Wen-Ning Zhao, Willis Read-Button, Xiaohua Wu, Stephen J. Haggarty, Lucienne V. Ronco, Jesse S. Boehm, Stuart L. Schreiber, John G. Doench, Joshua A. Bittker, David E. Root, Bang Wong, Todd R. Golub. A Next Generation Connectivity Map: L1000 Platform And The First 1,000,000 Profiles. bioRxiv (2017).(2017).