Deep integrative models for large-scale human genomics

Arnór I. Sigurdsson, David Westergaard, Ole Winther, Ole Lund, Søren Brunak, Bjarni J. Vilhjálmsson, Simon RasmussenbioRxiv 2021 · DOI

TL;DR

This paper introduces "EIR," a deep learning framework for predicting polygenic risk scores (PRSs) that goes beyond traditional linear models by capturing complex, non-linear interactions in genomic data. EIR, powered by a novel "genome-local-net" (GLN) architecture, demonstrates improved predictive performance across a wide range of diseases, particularly autoimmune ones, and effectively integrates diverse clinical data for more robust and calibrated risk assessments.

What Problem Does This Paper Address?

Current methods for predicting polygenic risk scores (PRSs) primarily rely on linear models that capture additive genetic effects. While these have been valuable, they struggle to account for complex interactions between genes (epistasis) and are limited in their ability to integrate diverse data types like clinical measurements, imaging, and other omics data. This limitation hinders the full potential of PRSs for precision medicine, which requires a more nuanced understanding of an individual's disease risk by considering a broader spectrum of biological and clinical factors. The authors aim to bridge this gap by developing a deep learning framework capable of handling large-scale genomic data, identifying non-linear interactions, and integrating multi-modal data for more accurate and comprehensive disease risk prediction.

Methods

The core of the paper is the EIR (Explainable Integrative Risk) framework, built around a novel neural network architecture called genome-local-net (GLN).

  • Genome-Local-Net (GLN): This architecture is specifically designed for large-scale genomic data. It utilizes a custom locally-connected layer (LCL) that is sparsely connected, significantly reducing the number of parameters compared to fully connected layers while still capturing local positional information in the genome. The GLN architecture consists of multiple LCLs followed by residual blocks, and it's fused with tabular data (like age and sex) before final prediction.
  • Multi-Task (MT) Learning: The framework supports training a single model to predict multiple disease outcomes simultaneously. This can improve efficiency and potentially regularization.
  • Data Integration: EIR seamlessly integrates genotype data with other clinical and biochemical measurements (blood, urine, anthropometrics).
  • Explainability: The authors employ techniques like SHAP values to understand which genetic variants and clinical features are driving the model's predictions, making the "black box" of deep learning more interpretable.
  • Handling Missing Data: The models are designed to handle missing genotype data without requiring imputation, simplifying the preprocessing pipeline.

Datasets

The primary dataset used in this study is the UK Biobank (UKBB). The authors utilized individual-level genotype data and various clinical measurements from UKBB participants. They performed analyses on subsets of the data, including:

  • A cohort of 413,736 individuals with British, Irish, or other Western European backgrounds for initial model development and comparison.
  • A larger set of 338 binary disease traits with at least 1,000 cases for broader performance evaluation.
  • A set of 290 diseases for integrative modeling with clinical data.

The paper does not introduce new benchmark datasets but rather applies their framework to existing, large-scale real-world data.

Key Insights and Ideas

  • Beyond Linearity: Deep learning, specifically the GLN architecture, can effectively capture complex non-linear interactions (epistasis) in genomic data, which linear models often miss. This is crucial for understanding the genetic architecture of many diseases.
  • The Power of Integration: Combining genotype data with clinical measurements (blood, urine, anthropometrics) significantly improves predictive performance for a large majority of diseases, highlighting the value of a holistic view of an individual's health.
  • Calibrated Predictions: Integrating genotype data leads to better-calibrated PRS models, meaning the predicted probabilities of disease are more reliable and closer to the true risk, which is essential for clinical decision-making.
  • Interpretable Deep Learning: The GLN model, despite its complexity, can be interpreted, revealing biologically relevant variants and interactions, demystifying deep learning in genomics.
  • Efficiency of MT Learning: While multi-task learning can sometimes lead to a slight performance trade-off, it offers significant gains in training efficiency and parameter reduction, especially for large numbers of tasks.

Key Results

  • GLN Outperforms LASSO: The GLN model consistently outperformed the LASSO (a common linear model) for a wide range of diseases, including autoimmune diseases like Type 1 Diabetes and Rheumatoid Arthritis, achieving statistically significant improvements in ROC-AUC.
  • Improved Performance with Integration: Integrating genotype data with clinical measurements improved PRS performance for 93% of the 290 diseases and disorders considered.
  • Better Calibration: Integrated models showed improved Expected Calibration Error (ECE) and Brier Score (BS) compared to measurement-only models, indicating more reliable probability estimates.
  • Identification of Interactions: The GLN model successfully identified widespread epistasis for Type 1 Diabetes, demonstrating its ability to uncover complex genetic relationships.
  • Efficiency: The framework demonstrated efficient training of large deep neural networks on high-dimensional data, with training times comparable to or better than traditional methods for certain tasks.

Limitations and Future Work

  • Ancestry: The study primarily focused on individuals of Western European ancestry. Developing predictive models that perform robustly across multiple ancestries is a critical future direction.
  • Data Modalities: The study primarily used genotype and tabular data. Future work could explore integrating other data types like high-resolution imaging and multi-omics data.
  • Overfitting in MT Learning: While MT learning offers efficiency, the authors observed that increasing the number of tasks could lead to a decrease in validation performance, suggesting that model capacity might be a limiting factor. Exploring more advanced MT learning architectures could mitigate this.
  • Calibration: Despite improvements, the overall calibration of models still highlights the importance of post-hoc calibration methods and further research into uncertainty estimation and out-of-distribution detection.
  • LD Dilution: The authors note that not filtering for Linkage Disequilibrium (LD) can dilute the activation signal across SNPs, potentially obscuring the true representative SNP. Future work could explore strategies to address this.

Summary updated

Abstract

ABSTRACT Polygenic risk scores (PRSs) are expected to play a critical role in achieving precision medicine. Currently, PRS predictors are generally based on linear models using summary statistics, and more recently individual-level data. However, these predictors mainly capture additive relationships and are limited in data modalities they can use. Here, we developed a deep learning framework (EIR) for PRS prediction which includes a model, genome-local-net (GLN), specifically designed for large scale genomics data. The framework supports multi-task (MT) learning, automatic integration of other clinical and biochemical data, and model explainability. When applied to individual level data in the UK Biobank, we found that GLN outperformed LASSO for a wide range of diseases and in particularly autoimmune diseases. Furthermore, we show that this was likely due to modelling epistasis, and we showcase this by identifying widespread epistasis for Type 1 Diabetes. Furthermore, we trained PRS by integrating genotype, blood, urine and anthropometrics and found that this improved performance for 93% of 290 diseases and disorders considered. Finally, we found that including genotype data provided better calibrated PRS models compared to using measurements alone. EIR is available at https://github.com/arnor-sigurdsson/EIR .
BibTeX
@misc{sigurdsson2021deep,
  author = {Arnór I. Sigurdsson and David Westergaard and Ole Winther and Ole Lund and Søren Brunak and Bjarni J. Vilhjálmsson and Simon Rasmussen},
  title = {{Deep integrative models for large-scale human genomics}},
  year = {2021},
  doi = {10.1101/2021.06.11.447883},
  howpublished = {bioRxiv},
}
Figure