GenNet framework: interpretable neural networks for phenotype prediction

Abstract

14 15 *Jointly supervised this project 16 17 Neural networks have been seldomly leveraged in population genomics due to the 18 computational burden and challenge of interpretability. Here, we propose GenNet, a novel 19 open-source deep learning framework for predicting phenotype from genotype. In this 20 framework, public prior biological knowledge is used to construct interpretable and 21 memory-efficient neural network architectures. These architectures obtain good predictive 22 performance for multiple traits and complex diseases, opening the door for neural 23 networks in population genomics. 24 25phenotypes at more and more independent loci. To illustrate, the latest GWAS for body height 30 based on 700,000 individuals identified more than 3000 near-independent significantly 31 associated single nucleotide polymorphisms (SNPs) 1 . This information, used in combination with 32 annotated biological databases such as: NCBI RefSeq, KEGG, Reactome and GTEx has proven 33 to be highly valuable for understanding the underlying biological mechanisms of complex 34 diseases 2-6 . In this paper, we propose a new framework, GenNet, that integrates these biological 35 data sources for discovery and interpretability in an end-to-end deep learning framework for 36 predicting phenotypes. 37Deep learning is the state of the art in many domains such as medical image analysis and natural 38 language processing because of its flexibility and modeling capabilities 7,8 . In many cases, deep 39 learning yields better performance compared to traditional approaches, since it can model highly 40 non-linear relations and scales very well with data size. However, this often comes at the cost of 41 interpretability, since there is a trade-off between complexity and interpretability 9,10 . 42 Additionally, when it comes to genotype data, the number of learnable parameters increases 43 dramatically because of the large input size, making it infeasible to use classical neural networks 44 in this domain. To overcome previous limitations, we propose a new framework, GenNet, in 45 which different types of biological information are used to define biologically plausible neural 46 network architectures, avoiding this trade-off and creating interpretable neural networks for 47 129 polygenic risk scoring for schizophrenia, which have AUC values on the order of 0.70 (ranging 130 0.49-0.85) 15 . This is noteworthy since in this study the schizophrenia predictions are based on 131 whole exome sequencing data as opposed to GWAS arrays spanning the whole genome.132 133 134 Discussion 135Here, we present a novel framework to train interpretable neural networks for phenotype 136 prediction from genotype. The proposed neural networks have connections defined by prior 137 biological knowledge only, reducing the number of connections and therefore the number of

Explainable artificial intelligence for omics data: a systematic mapping study

“…For example, motif extraction, as presented by Amilpur and Bhukya [ 94 ], could benefit not only GNNs but also other approaches like RNNs or DNNs with only minor adaptations. Other papers propose the development of simplified interpretable NNs, for example, by including prior knowledge into the models so each node represents biological knowledge [ 95 ]. This interpretability approach could also be introduced into DNNs, which might produce even better results due to their increased number of connections.…”

GenNet framework: interpretable deep learning for predicting phenotypes from genetic data

“…GenNet is an open-source framework usable from command line. GenNet and its tutorials, including how to build new layers and networks from prior knowledge, can be found on: https://github.com/arnovanhilten/GenNet/ and Zenodo 44 .…”

A Biologically Interpretable Graph Convolutional Network to Link Genetic Risk Pathways and Neuroimaging Markers of Disease

“…Within the genetics realm, there is a vast literature that associates SNPs and genes to different biological pathways (Mi et al, 2013; Ashburner et al, 2000; Carbon et al, 2021). The works of (van Hilten et al, 2020; Gaudelet et al, 2020) have used this information to design a sparse artificial neural network that aggregates genetic risk according to these pathways in order to predict a phenotypic variable. While an important first step, their ANN contains just a single hidden layer, which does not account for the hierarchical and interconnected nature of the biological processes.…”