Stochastic semi-supervised learning to prioritise genes from high-throughput genomic screens

Abstract

Access to large-scale genomics datasets has increased the utility of hypothesis-free genome-wide analyses that result in candidate lists of genes. Often these analyses highlight several gene signals that might contribute to pathogenesis but are insufficiently powered to reach experiment-wide significance. This often triggers a process of laborious evaluation of highly-ranked genes through manual inspection of various public knowledge resources to triage those considered sufficiently interesting for deeper investigation. Here, we introduce a novel multi-dimensional, multi-step machine learning framework to objectively and more holistically assess biological relevance of genes to disease studies, by relying on a plethora of gene-associated annotations. We developed mantis-ml to serve as an automated machine learning (AutoML) framework, following a stochastic semisupervised learning approach to rank known and novel disease-associated genes through iterative training and prediction sessions of random balanced datasets across the protein-coding exome (n=18,626 genes). We applied this framework on a range of disease-specific areas and as a generic disease likelihood estimator, achieving an average Area Under Curve (AUC) prediction performance of 0.85. Critically, to demonstrate applied utility on exome-wide association studies, we overlapped mantis-ml disease-specific predictions with data from published cohort-level association studies. We retrieved statistically significant enrichment of high mantis-ml predictions among the top-ranked genes from hypothesis-free cohort-level statistics (p<0.05), suggesting the capture of true prioritisation signals. We believe that mantis-ml is a novel easy-to-use tool to support objectively triaging gene discovery and overall enhancing our understanding of complex genotype-phenotype associations.

Reaching the End-Game for GWAS: Machine Learning Approaches for the Prioritization of Complex Disease Loci

“…However, the AUCs between all algorithms were deemed too similar to conclude one model out-performed all others across datasets. These results were also supported by comparison with a combined framework using all models in prioritization, the stacking classifier, ensuring the highest reliability in the chosen classifier for each disease (Vitsios and Petrovski, 2019). Kafaie et al (2019) aimed to prioritize genes associated with colorectal cancer comparing various models (SVM, random forest, logistic regression with stochastic gradient descent, and K−nearest neighbors).…”

“…However, within benchmarking studies, SVM has not shown itself to be a top-performing model. For example, Vitsios and Petrovski (2019) found it had the lowest AUC (0.83, only slightly lower than the top-performing random forest at 0.85) of their seven models, while Kafaie et al (2019) found SVM performed better than random forest yet worse than logistic regression. The varying performance of SVM also highlights the importance of input data, as Kafaie et al (2019) were one of the only studies to focus on comparing feature selection methods as well as models.…”

“…They found that logistic regression was the highest performing ML model -emphasizing that a classification problem may require simpler solutions. Besides ensemble learning and logistic regression, SVM is also consistently used within studies performing benchmark comparisons (Roshan et al, 2011;Isakov et al, 2017;Maciukiewicz et al, 2018;Vitsios and Petrovski, 2019). SVM aims to plot a decision boundary between groups by measuring hyperplanes -based on the distances between the most extreme samples of each classification group (Smola and Scholkopf, 2004; Figure 2).…”

“…When tailored for understanding GWAS data, ML predictions can provide an improved statistical foundation of evidence to support or improve GWAS results. For instance, ML in GWAS has been applied to identify loci, increase the statistical power of GWAS (Mieth et al, 2016), detect epistatic interactions (Leem et al, 2014), improve polygenic risk scoring produced from GWAS (Pare et al, 2017), and prioritize genes and variants on post-GWAS analysis (Vitsios and Petrovski, 2019). Here we will focus on the ML applications developed for post-GWAS prioritization.…”

“…Since then several computational methods for prioritizing GWAS associated loci have been developed with growing attention on ML applications (Fridley et al, 2011;Gagliano et al, 2015;Raj and Sreeja, 2018;Wu et al, 2018). ML for prioritizing GWAS results has used common models ( Figure 2) such as logistic regression, decision tree classifiers such as -e.g., gradient boosting machines (GBM) and random forests (Wang et al, 2013;Oh et al, 2017), -and support vector machines (SVM; Vitsios and Petrovski, 2019), with more recent advances including deep learning models (Khan et al, 2018;Zhou et al, 2018).…”