skip to main content

Title: DeepGAMI: deep biologically guided auxiliary learning for multimodal integration and imputation to improve genotype–phenotype prediction
Abstract Background

Genotypes are strongly associated with disease phenotypes, particularly in brain disorders. However, the molecular and cellular mechanisms behind this association remain elusive. With emerging multimodal data for these mechanisms, machine learning methods can be applied for phenotype prediction at different scales, but due to the black-box nature of machine learning, integrating these modalities and interpreting biological mechanisms can be challenging. Additionally, the partial availability of these multimodal data presents a challenge in developing these predictive models.


To address these challenges, we developed DeepGAMI, an interpretable neural network model to improve genotype–phenotype prediction from multimodal data. DeepGAMI leverages functional genomic information, such as eQTLs and gene regulation, to guide neural network connections. Additionally, it includes an auxiliary learning layer for cross-modal imputation allowing the imputation of latent features of missing modalities and thus predicting phenotypes from a single modality. Finally, DeepGAMI uses integrated gradient to prioritize multimodal features for various phenotypes.


We applied DeepGAMI to several multimodal datasets including genotype and bulk and cell-type gene expression data in brain diseases, and gene expression and electrophysiology data of mouse neuronal cells. Using cross-validation and independent validation, DeepGAMI outperformed existing methods for classifying disease types, and cellular and clinical phenotypes, even using single modalities (e.g., AUC score of 0.79 for Schizophrenia and 0.73 for cognitive impairment in Alzheimer’s disease).


We demonstrated that DeepGAMI improves phenotype prediction and prioritizes phenotypic features and networks in multiple multimodal datasets in complex brains and brain diseases. Also, it prioritized disease-associated variants, genes, and regulatory networks linked to different phenotypes, providing novel insights into the interpretation of gene regulatory mechanisms. DeepGAMI is open-source and available for general use.

more » « less
Author(s) / Creator(s):
; ; ; ; ; ; ; ; ; ;
Publisher / Repository:
Springer Science + Business Media
Date Published:
Journal Name:
Genome Medicine
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract Background

    Current methods for analyzing single-cell datasets have relied primarily on static gene expression measurements to characterize the molecular state of individual cells. However, capturing temporal changes in cell state is crucial for the interpretation of dynamic phenotypes such as the cell cycle, development, or disease progression. RNA velocity infers the direction and speed of transcriptional changes in individual cells, yet it is unclear how these temporal gene expression modalities may be leveraged for predictive modeling of cellular dynamics.


    Here, we present the first task-oriented benchmarking study that investigates integration of temporal sequencing modalities for dynamic cell state prediction. We benchmark ten integration approaches on ten datasets spanning different biological contexts, sequencing technologies, and species. We find that integrated data more accurately infers biological trajectories and achieves increased performance on classifying cells according to perturbation and disease states. Furthermore, we show that simple concatenation of spliced and unspliced molecules performs consistently well on classification tasks and can be used over more memory intensive and computationally expensive methods.


    This work illustrates how integrated temporal gene expression modalities may be leveraged for predicting cellular trajectories and sample-associated perturbation and disease phenotypes. Additionally, this study provides users with practical recommendations for task-specific integration of single-cell gene expression modalities.

    more » « less
  2. Abstract Background

    In Alzheimer’s Diseases (AD) research, multimodal imaging analysis can unveil complementary information from multiple imaging modalities and further our understanding of the disease. One application is to discover disease subtypes using unsupervised clustering. However, existing clustering methods are often applied to input features directly, and could suffer from the curse of dimensionality with high-dimensional multimodal data. The purpose of our study is to identify multimodal imaging-driven subtypes in Mild Cognitive Impairment (MCI) participants using a multiview learning framework based on Deep Generalized Canonical Correlation Analysis (DGCCA), to learn shared latent representation with low dimensions from 3 neuroimaging modalities.


    DGCCA applies non-linear transformation to input views using neural networks and is able to learn correlated embeddings with low dimensions that capture more variance than its linear counterpart, generalized CCA (GCCA). We designed experiments to compare DGCCA embeddings with single modality features and GCCA embeddings by generating 2 subtypes from each feature set using unsupervised clustering. In our validation studies, we found that amyloid PET imaging has the most discriminative features compared with structural MRI and FDG PET which DGCCA learns from but not GCCA. DGCCA subtypes show differential measures in 5 cognitive assessments, 6 brain volume measures, and conversion to AD patterns. In addition, DGCCA MCI subtypes confirmed AD genetic markers with strong signals that existing late MCI group did not identify.


    Overall, DGCCA is able to learn effective low dimensional embeddings from multimodal data by learning non-linear projections. MCI subtypes generated from DGCCA embeddings are different from existing early and late MCI groups and show most similarity with those identified by amyloid PET features. In our validation studies, DGCCA subtypes show distinct patterns in cognitive measures, brain volumes, and are able to identify AD genetic markers. These findings indicate the promise of the imaging-driven subtypes and their power in revealing disease structures beyond early and late stage MCI.

    more » « less
  3. Key points

    Induced pluripotent stem cell‐derived cardiomyocytes (iPSC‐CMs) capture patient‐specific genotype–phenotype relationships, as well as cell‐to‐cell variability of cardiac electrical activity

    Computational modelling and simulation provide a high throughput approach to reconcile multiple datasets describing physiological variability, and also identify vulnerable parameter regimes

    We have developed a whole‐cell model of iPSC‐CMs, composed of single exponential voltage‐dependent gating variable rate constants, parameterized to fit experimental iPSC‐CM outputs

    We have utilized experimental data across multiple laboratories to model experimental variability and investigate subcellular phenotypic mechanisms in iPSC‐CMs

    This framework links molecular mechanisms to cellular‐level outputs by revealing unique subsets of model parameters linked to known iPSC‐CM phenotypes


    There is a profound need to develop a strategy for predicting patient‐to‐patient vulnerability in the emergence of cardiac arrhythmia. A promisingin vitromethod to address patient‐specific proclivity to cardiac disease utilizes induced pluripotent stem cell‐derived cardiomyocytes (iPSC‐CMs). A major strength of this approach is that iPSC‐CMs contain donor genetic information and therefore capture patient‐specific genotype–phenotype relationships. A cited detriment of iPSC‐CMs is the cell‐to‐cell variability observed in electrical activity. We postulated, however, that cell‐to‐cell variability may constitute a strength when appropriately utilized in a computational framework to build cell populations that can be employed to identify phenotypic mechanisms and pinpoint key sensitive parameters. Thus, we have exploited variation in experimental data across multiple laboratories to develop a computational framework for investigating subcellular phenotypic mechanisms. We have developed a whole‐cell model of iPSC‐CMs composed of simple model components comprising ion channel models with single exponential voltage‐dependent gating variable rate constants, parameterized to fit experimental iPSC‐CM data for all major ionic currents. By optimizing ionic current model parameters to multiple experimental datasets, we incorporate experimentally‐observed variability in the ionic currents. The resulting population of cellular models predicts robust inter‐subject variability in iPSC‐CMs. This approach links molecular mechanisms to known cellular‐level iPSC‐CM phenotypes, as shown by comparing immature and mature subpopulations of models to analyse the contributing factors underlying each phenotype. In the future, the presented models can be readily expanded to include genetic mutations and pharmacological interventions for studying the mechanisms of rare events, such as arrhythmia triggers.

    more » « less
  4. ABSTRACT Gene regulatory networks (GRNs) are critical for dynamic transcriptional responses to environmental stress. However, the mechanisms by which GRN regulation adjusts physiology to enable stress survival remain unclear. Here we investigate the functions of transcription factors (TFs) within the global GRN of the stress-tolerant archaeal microorganism Halobacterium salinarum . We measured growth phenotypes of a panel of TF deletion mutants in high temporal resolution under heat shock, oxidative stress, and low-salinity conditions. To quantitate the noncanonical functional forms of the growth trajectories observed for these mutants, we developed a novel modeling framework based on Gaussian process regression and functional analysis of variance (FANOVA). We employ unique statistical tests to determine the significance of differential growth relative to the growth of the control strain. This analysis recapitulated known TF functions, revealed novel functions, and identified surprising secondary functions for characterized TFs. Strikingly, we observed that the majority of the TFs studied were required for growth under multiple stress conditions, pinpointing regulatory connections between the conditions tested. Correlations between quantitative phenotype trajectories of mutants are predictive of TF-TF connections within the GRN. These phenotypes are strongly concordant with predictions from statistical GRN models inferred from gene expression data alone. With genome-wide and targeted data sets, we provide detailed functional validation of novel TFs required for extreme oxidative stress and heat shock survival. Together, results presented in this study suggest that many TFs function under multiple conditions, thereby revealing high interconnectivity within the GRN and identifying the specific TFs required for communication between networks responding to disparate stressors. IMPORTANCE To ensure survival in the face of stress, microorganisms employ inducible damage repair pathways regulated by extensive and complex gene networks. Many archaea, microorganisms of the third domain of life, persist under extremes of temperature, salinity, and pH and under other conditions. In order to understand the cause-effect relationships between the dynamic function of the stress network and ultimate physiological consequences, this study characterized the physiological role of nearly one-third of all regulatory proteins known as transcription factors (TFs) in an archaeal organism. Using a unique quantitative phenotyping approach, we discovered functions for many novel TFs and revealed important secondary functions for known TFs. Surprisingly, many TFs are required for resisting multiple stressors, suggesting cross-regulation of stress responses. Through extensive validation experiments, we map the physiological roles of these novel TFs in stress response back to their position in the regulatory network wiring. This study advances understanding of the mechanisms underlying how microorganisms resist extreme stress. Given the generality of the methods employed, we expect that this study will enable future studies on how regulatory networks adjust cellular physiology in a diversity of organisms. 
    more » « less
  5. Abstract Motivation

    Function annotations of gene products, and phenotype annotations of genotypes, provide valuable information about molecular mechanisms that can be utilized by computational methods to identify functional and phenotypic relatedness, improve our understanding of disease and pathobiology, and lead to discovery of drug targets. Identifying functions and phenotypes commonly requires experiments which are time-consuming and expensive to carry out; creating the annotations additionally requires a curator to make an assertion based on reported evidence. Support to validate the mutual consistency of functional and phenotype annotations as well as a computational method to predict phenotypes from function annotations, would greatly improve the utility of function annotations.


    We developed a novel ontology-based method to validate the mutual consistency of function and phenotype annotations. We apply our method to mouse and human annotations, and identify several inconsistencies that can be resolved to improve overall annotation quality. We also apply our method to the rule-based prediction of regulatory phenotypes from functions and demonstrate that we can predict these phenotypes with Fmax of up to 0.647.

    Availability and implementation

    more » « less