A Deep Learning Approach for Predicting Antidepressant Response in Major Depression Using Clinical and Genetic Biomarkers

Abstract

In the wake of recent advances in scientific research, personalized medicine using deep learning techniques represents a new paradigm. In this work, our goal was to establish deep learning models which distinguish responders from non-responders, and also to predict possible antidepressant treatment outcomes in major depressive disorder (MDD). To uncover relationships between the responsiveness of antidepressant treatment and biomarkers, we developed a deep learning prediction approach resulting from the analysis of genetic and clinical factors such as single nucleotide polymorphisms (SNPs), age, sex, baseline Hamilton Rating Scale for Depression score, depressive episodes, marital status, and suicide attempt status of MDD patients. The cohort consisted of 455 patients who were treated with selective serotonin reuptake inhibitors (treatment-response rate = 61.0%; remission rate = 33.0%). By using the SNP dataset that was original to a genome-wide association study, we selected 10 SNPs (including ABCA13 rs4917029, BNIP3 rs9419139, CACNA1E rs704329, EXOC4 rs6978272, GRIN2B rs7954376, LHFPL3 rs4352778, NELL1 rs2139423, NUAK1 rs2956406, PREX1 rs4810894, and SLIT3 rs139863958) which were associated with antidepressant treatment response. Furthermore, we pinpointed 10 SNPs (including ARNTL rs11022778, CAMK1D rs2724812, GABRB3 rs12904459, GRM8 rs35864549, NAALADL2 rs9878985, NCALD rs483986, PLA2G4A rs12046378, PROK2 rs73103153, RBFOX1 rs17134927, and ZNF536 rs77554113) in relation to remission. Then, we employed multilayer feedforward neural networks (MFNNs) containing 1–3 hidden layers and compared MFNN models with logistic regression models. Our analysis results revealed that the MFNN model with 2 hidden layers (area under the receiver operating characteristic curve (AUC) = 0.8228 ± 0.0571; sensitivity = 0.7546 ± 0.0619; specificity = 0.6922 ± 0.0765) performed maximally among predictive models to infer the complex relationship between antidepressant treatment response and biomarkers. In addition, the MFNN model with 3 hidden layers (AUC = 0.8060 ± 0.0722; sensitivity = 0.7732 ± 0.0583; specificity = 0.6623 ± 0.0853) achieved best among predictive models to predict remission. Our study indicates that the deep MFNN framework may provide a suitable method to establish a tool for distinguishing treatment responders from non-responders prior to antidepressant therapy.

Pharmacogenomics‐Driven Prediction of Antidepressant Treatment Outcomes: A Machine‐Learning Approach With Multi‐trial Replication

“…As noted earlier, prior experimental work showed that knockdown of the expression of both TSPAN5 and ERICH3 in neuronally derived cell lines resulted in decreased serotonin release into the culture media . The DEFB1 gene encodes a protein expressed in gastrointestinal mucosa that can inactivate lipopolysaccharides and, in turn, inhibit both inflammation and the biosynthesis of kynurenine, which is enhanced by inflammatory mediators . The facts that the DEFB1 SNPs figured so prominently and that this gene encodes a gut mucosal protein that can inactivate both lipopolysaccharides and gut bacteria highlight the potential importance of the rapidly evolving concept of agut–brain axis …”

“…9 The DEFB1 gene encodes a protein expressed in gastrointestinal mucosa that can inactivate lipopolysaccharides and, in turn, inhibit both inflammation and the biosynthesis of kynurenine, which is enhanced by inflammatory mediators. 10 The facts that the DEFB1 SNPs figured so prominently and that this gene encodes a gut mucosal protein that can inactivate both lipopolysaccharides and gut bacteria highlight the potential importance of the rapidly evolving concept of a gut-brain axis. 25 The identification of these "top hit" SNPs during GWAS was performed for quantitative biological traits (i.e., metabolite concentrations), rather than measures of MDD clinical symptom severity (i.e., HDRS or QIDS-C), as our use of phenotypes represented a conscious attempt to move our analyses toward the biological underpinning of SSRI response.…”

“…A limitation acknowledged by the authors of prior reports is the lack of inclusion of biological factors associated with depression severity or therapeutic response to antidepressants in the prediction models . To address this limitation, recent machine‐learning approaches have used genomics and/or metabolomics data to predict SSRI response, with AUC values of 0.68–0.78 . Although the studies demonstrated the feasibility of integrating biological factors with machine learning to achieve improved prediction performance, they were limited by lack of replication across multiple trials and multiple depression rating scales, in addition to lack of significance from a pharmacogenomics perspective …”

“…[3][4][5]8 To address this limitation, recent machinelearning approaches have used genomics and/or metabolomics data to predict SSRI response, with AUC values of 0.68-0.78. [8][9][10] Although the studies demonstrated the feasibility of integrating biological factors with machine learning to achieve improved prediction performance, they were limited by lack of replication across multiple trials and multiple depression rating scales, in addition to lack of significance from a pharmacogenomics perspective. [8][9][10][11] In the present study, we used a machine-learning workflow (depicted schematically in Figure 1) to study the capabilities of Figure 1 The two-stage analysis workflow.…”

Cell-type-specific methylome-wide association studies implicate neurodegenerative processes and neuroimmune communication in major depressive disorder

“…Prex1 results in autismlike behavior 36 and is associated with antidepressant response in humans 37 . Interestingly, very few associations that were detected in deconvoluted neurons (1 CpG) or glia (14 CpGs) were also detected in bulk, suggesting celltypespecific effects are indeed diluted or obscured in bulk tissue.…”