RESUMO
Genome interpretation (GI) encompasses the computational attempts to model the relationship between genotype and phenotype with the goal of understanding how the first leads to the second. While traditional approaches have focused on sub-problems such as predicting the effect of single nucleotide variants or finding genetic associations, recent advances in neural networks (NNs) have made it possible to develop end-to-end GI models that take genomic data as input and predict phenotypes as output. However, technical and modeling issues still need to be fixed for these models to be effective, including the widespread underdetermination of genomic datasets, making them unsuitable for training large, overfitting-prone, NNs. Here we propose novel GI models to address this issue, exploring the use of two types of transfer learning approaches and proposing a novel Biologically Meaningful Sparse NN layer specifically designed for end-to-end GI. Our models predict the leaf and seed ionome in A.thaliana, obtaining comparable results to our previous over-parameterized model while reducing the number of parameters by 8.8 folds. We also investigate how the effect of population stratification influences the evaluation of the performances, highlighting how it leads to (1) an instance of the Simpson's Paradox, and (2) model generalization limitations.
Assuntos
Arabidopsis , Genoma de Planta , Folhas de Planta , Sementes , Arabidopsis/genética , Folhas de Planta/genética , Folhas de Planta/metabolismo , Sementes/genética , Sementes/metabolismo , Redes Neurais de Computação , Genômica/métodos , Fenótipo , Modelos Genéticos , GenótipoRESUMO
BACKGROUND: Despite clear evidence of nonlinear interactions in the molecular architecture of polygenic diseases, linear models have so far appeared optimal in genotype-to-phenotype modeling. A key bottleneck for such modeling is that genetic data intrinsically suffers from underdetermination ([Formula: see text]). Millions of variants are present in each individual while the collection of large, homogeneous cohorts is hindered by phenotype incidence, sequencing cost, and batch effects. RESULTS: We demonstrate that when we provide enough training data and control the complexity of nonlinear models, a neural network outperforms additive approaches in whole exome sequencing-based inflammatory bowel disease case-control prediction. To do so, we propose a biologically meaningful sparsified neural network architecture, providing empirical evidence for positive and negative epistatic effects present in the inflammatory bowel disease pathogenesis. CONCLUSIONS: In this paper, we show that underdetermination is likely a major driver for the apparent optimality of additive modeling in clinical genetics today.
Assuntos
Doenças Inflamatórias Intestinais , Dinâmica não Linear , Humanos , Tamanho da Amostra , Doenças Inflamatórias Intestinais/genética , Redes Neurais de Computação , FenótipoRESUMO
High-throughput sequencing allowed the discovery of many disease variants, but nowadays it is becoming clear that the abundance of genomics data mostly just moved the bottleneck in Genetics and Precision Medicine from a data availability issue to a data interpretation issue. To solve this empasse it would be beneficial to apply the latest Deep Learning (DL) methods to the Genome Interpretation (GI) problem, similarly to what AlphaFold did for Structural Biology. Unfortunately DL requires large datasets to be viable, and aggregating genomics datasets poses several legal, ethical and infrastructural complications. Federated Learning (FL) is a Machine Learning (ML) paradigm designed to tackle these issues. It allows ML methods to be collaboratively trained and tested on collections of physically separate datasets, without requiring the actual centralization of sensitive data. FL could thus be key to enable DL applications to GI on sufficiently large genomics data. We propose FedCrohn, a FL GI Neural Network model for the exome-based Crohn's Disease risk prediction, providing a proof-of-concept that FL is a viable paradigm to build novel ML GI approaches. We benchmark it in several realistic scenarios, showing that FL can indeed provide performances similar to conventional ML on centralized data, and that collaborating in FL initiatives is likely beneficial for most of the medical centers participating in them.