Machine Learning for Integrating Data in Biology and Medicine: Principles, Practice, and Opportunities

Marinka Zitnik; Francis Nguyen; Bo Wang; Jure Leskovec; Anna Goldenberg; Michael M. Hoffman

Source-linked AI summary

Machine Learning for Integrating Data in Biology and Medicine: Principles, Practice, and Opportunities

Marinka Zitnik, Francis Nguyen, Bo Wang, Jure Leskovec, Anna Goldenberg, Michael M. Hoffman

arXiv:1807.00123v2Jun 30, 2018q-bio.QMcs.CEcs.LGstat.ML

TL;DR

Biomedical complexity spans heterogeneous molecular, phenotypic, lifestyle, and environmental data that no single data type can capture. This Review organizes machine learning approaches for integrating these sources, surveys implementations and applications, and examines challenges and future directions. It concludes that integrative methods are promising but must be selected according to the data, models, and biomedical outcomes involved.

Problem
Single data types cannot capture the complexity of biological systems and disease, motivating methods that combine heterogeneous biomedical evidence.
Method
The Review presents principles and a taxonomy of machine learning methods for integrating biomedical data, alongside implementations and applications across biomedical levels.
Results
The Review describes successful data-integration applications, including disease subtyping, biomarker discovery, protein-function prediction, and lifestyle-genomic analyses.
Takeaways & Limitations
Integrative approaches can provide more comprehensive biomedical understanding, but no single method is expected to perform best for every problem.
Takeaways & Limitations
Models trained on existing biomedical datasets may perform poorly when applied to novel patients, new data, or knowledge outside the training domain.

Abstract

from arXiv · show

New technologies have enabled the investigation of biology and human health at an unprecedented scale and in multiple dimensions. These dimensions include a myriad of properties describing genome, epigenome, transcriptome, microbiome, phenotype, and lifestyle. No single data type, however, can capture the complexity of all the factors relevant to understanding a phenomenon such as a disease. Integrative methods that combine data from multiple technologies have thus emerged as critical statistical and computational approaches. The key challenge in developing such approaches is the identification of effective models to provide a comprehensive and relevant systems view. An ideal method can answer a biological or medical question, identifying important features and predicting outcomes, by harnessing heterogeneous data across several dimensions of biological variation. In this Review, we describe the principles of data integration and discuss current methods and available implementations. We provide examples of successful data integration in biology and medicine. Finally, we discuss current challenges in biomedical integrative methods and our perspective on the future development of the field.

1. Introduction

Modern technologies generate heterogeneous biomedical data across molecular, phenotypic, behavioral, and ecological dimensions, but single data types cannot capture disease complexity. This Review presents data integration as a way to combine these sources for more comprehensive analysis and prediction.

High-throughput and sensor technologies now measure biological systems across DNA, epigenomic, transcriptomic, proteomic, phenotypic, and lifestyle dimensions.
Single-data-type studies can miss patterns needed to understand complex phenomena such as disease.
In a diagnostic example, genomic, laboratory, lifestyle, and biomedical-knowledge data are combined before machine learning predicts near-term disease probability.
Integrating multiple evidence sources can reduce false positives and support a more complete model spanning molecular to environmental information.
The Review surveys integration principles, machine learning methods, implementations, applications, challenges, and future opportunities.

2. Challenges in data integration for biology and medicine

Biomedical integration is difficult because datasets are high-dimensional yet sparse, incomplete, biased, heterogeneous, dynamic, and noisy. These limitations complicate meaningful analysis, extrapolation beyond observed knowledge, and reliable clinical decision support.

Biomedical datasets combine high dimensionality with sparsity, including many genetic variants measured in relatively few individuals.
Measurement limits, physical constraints, and investigative biases produce incomplete and unevenly represented biomedical knowledge.
Changing outcomes, including evolving cancer cells, bacteria, and viruses, require integration methods that account for temporal dynamics.
Models trained on existing datasets may generalize poorly to unique patients, new data, or drug responses transferred across species.
A single model trained on a single dataset captures only a small part of interconnected biomedical knowledge, motivating diverse-source integration.

3. Conceptual organization of methods for data integration

Biomedical data integration methods can be organized by when datasets are combined and by the form of their outputs. Early, intermediate, and late strategies differ in how they represent data and models before producing labels, relationships, or complex structures.

Vertical integration combines models across spatial and temporal scales, whereas horizontal integration combines datasets at one biomedical level.
Integration stage: Early integration transforms datasets into a shared feature table or graph before machine learning analysis.
Integration stage: Dimensionality or representation learning can project high-dimensional datasets into low-dimensional spaces before aggregation.
Integration stage: Late integration fits independent first-level models and combines their predictions or weights using a second-level model or meta-predictor.
Integration stage: Intermediate integration learns a joint representation while estimating model parameters, preserving dataset structure without concatenating inputs or fitting separate models.
Prediction outputs: Integrated analyses can output class labels, entity relationships, or complex structures such as inferred networks and partitions.
Available software tools are an aspect of the practical landscape of integration methods.

4. Focus of this Review

This Review surveys machine learning and statistical approaches for integrating heterogeneous biomedical data across molecular, cellular, patient, and population levels. It also covers implementations, applications, strengths, weaknesses, challenges, and future opportunities.

The Review targets computational researchers interested in machine learning applications to biology and medicine amid rapidly expanding heterogeneous data.
It examines available tools, strengths, weaknesses, outstanding challenges, and opportunities for integrative analysis.
Related reviews address narrower topics including network inference, biological networks, visualization, and other specialized data-integration perspectives.
The presentation follows genetic information from DNA and epigenetic markers through single-cell expression, proteins, cellular phenotypes, patients, and populations.

5. Epigenomic variation and gene regulation

Epigenomic states reflect cell identity and gene regulation, but interpreting them requires combining multiple assays and computationally modeling genomic patterns. SAGA methods segment and cluster genomic regions to support functional annotation, while integrated data improve transcription-factor binding prediction.

The epigenome comprises DNA, its packaging, and associated biomolecules whose state varies across cells with distinct gene-expression programs.
Researchers combine multiple epigenomic assays because individual assay signals are difficult to interpret in isolation.
Semi-automated genome annotation: SAGA methods simultaneously cluster regions by epigenomic similarity and segment neighboring genomic positions into heterogeneous regions.
Semi-automated genome annotation: SAGA segmentation and clustering are largely automated, but human experts interpret clusters and assign putative biological roles.
Semi-automated genome annotation: SAGA methods have annotated human, mouse, and fruit-fly genomes, helping assign functions to genomic regions.
Transcription-factor binding: Existing binding assays can require specific antibodies, genome editing, or more patient-sample material than is available.
Transcription-factor binding: Integrated predictors combine sequence motifs with epigenomic signals such as open chromatin or histone modifications to infer transcription-factor binding.

6. Noncoding variant effects

Predicting the effects of noncoding variants is difficult because most of the genome does not directly encode proteins. Methods therefore use sequence models, conservation, evolutionary selection, and integrated genomic or epigenomic features to assess regulatory or deleterious effects.

Noncoding variants are difficult to interpret, despite many being associated with phenotypic traits or genetic diseases.
gkm-SVM identifies enhancer-associated k-mers and uses an SVM to distinguish enriched sequence patterns from random background sequences.
Deep convolutional neural networks use DNA sequence to predict DNase-seq signal probabilities.
Conservation-based methods assume that sequences preserved across species or populations are more likely to be functionally constrained.
CADD integrates 63 conservation and epigenomic features with a linear-kernel SVM to score variants.
INSIGHT models natural selection across species, whereas fitCons clusters DNase-seq, RNA-seq, and histone-modification data.

7. Integrative single-cell analysis

Single-cell technologies reveal cellular heterogeneity beyond bulk averages but produce sparse, high-dimensional data and introduce substantial technical and biological variation. Integrative methods combine representations, clusterings, or omics dimensions, while scalability and computational complexity remain important constraints.

Single-cell measurements expose variation within tissues and cell populations that pooled assays can obscure.
Cell type discovery and exploration: scRNA-seq measures individual-cell gene expression to characterize population heterogeneity and functional diversity.
Cell type discovery and exploration: scRNA-seq analysis must address high dropout rates and the curse of dimensionality.
Cell type discovery and exploration: Unsupervised methods address these challenges through dimensionality reduction, consensus clustering, probabilistic models, or compressed sensing.
Cell type discovery and exploration: Ensemble methods integrate multiple cell representations or clusterings to produce refined similarity structures.
Cross-sample integration: Integrative analysis across patient groups, tissues, and conditions is limited because biological and technical variation can dominate pooled-cell clustering signals.
Single-cell multi-omics analysis: Single-cell multi-omics methods seek shared and modality-specific information across measurements such as methylation, chromatin accessibility, proteome, and metabolome.
Single-cell multi-omics analysis: Correlation-based methods and two-stage clustering approaches are current strategies for integrating single-cell multi-omics data.

8. Cellular phenotype and function

Cellular phenotype and function studies use heterogeneous biological data to predict protein functions, prioritize genes, and characterize interaction networks. Integrative methods combine sequence, expression, interaction, structural, evolutionary, literature, and other data to address incomplete annotations and complex biological relationships.

Protein function prediction: A matrix-based integration strategy establishes statistical correspondence among diverse datasets, including gene-phenotype associations, expression profiles, literature, and paper annotations.Gene-publication and publication-MeSH relationships illustrate how seemingly disparate matrices can be linked.
Protein function prediction: Protein function prediction draws on sequence similarity, gene expression, protein and genetic interactions, metabolic data, evolution, structure, and biomedical text.These heterogeneous evidence sources can be combined rather than analyzed separately.
Gene prioritization: Gene prioritization ranks candidate genes by similarity to seed genes with known roles, narrowing dozens or hundreds of candidates for experimental validation.Methods use different algorithms, datasets, and strategies, with several public implementations and experimental validations.
Protein function prediction: Supervised protein-function prediction faces more than 40,000 GO terms, multi-label outputs, and physiological functions that require considering proteins jointly.Joint latent factor, multi-label, and ensemble learning approaches address these challenges.
Interaction networks: Biological networks support function and phenotype discovery because interacting proteins tend to share functions or phenotypes, while incomplete PPI knowledge motivates computational prediction.Predicted interactions can combine protein sequence, structural, transcriptomic, and essentiality information.

9. Computational pharmacology

Computational pharmacology integrates molecular, interaction, drug, and patient data to predict drug-target interactions, drug combinations, side effects, and new drug uses. These methods address clinically important problems while confronting sparse interaction data, combinatorial growth, and limitations of handcrafted representations.

Computational pharmacology: Computational pharmacology uses physicochemical, pharmacological, and phenotypic measurements to predict how drugs affect the human body.These data describe chemical properties, drug-target binding and activity, cellular changes, and phenotypic outcomes.
Drug-target interaction prediction: Drug-target interaction prediction supports understanding drug side effects, therapeutic mechanisms, and medical indications.Machine-learning formulations use known interactions as positive labels and drug or target features as inputs.
Drug-target interaction prediction: Ligand-based approaches perform poorly when a target has few known binding ligands or when the inferred chemical model is low quality.This limits performance in sparse-ligand settings.
Drug-drug interaction and drug combination prediction: Drug combinations can improve therapeutic efficacy but increase side-effect risk because one drug may alter another’s activity.Drug-drug interactions produce responses beyond those expected under no interaction.
Drug-drug interaction and drug combination prediction: Drug-drug interaction discovery is difficult because drug pairs grow combinatorially, clinical manifestations vary across patients, and exhaustive testing is costly and time-consuming.Computational methods therefore rank pairs that potentially interact.
Drug-drug interaction and drug combination prediction: Heterogeneous and multimodal networks represent drugs, proteins, and their interactions to predict drug-target interactions, drug-drug interactions, and patient side effects.Decagon models side effects as distinct edge types and uses graph convolutional learning.

10. Disease subtyping and biomarker discovery

Disease subtyping and biomarker discovery seek clinically meaningful patient subgroups and integrated biological markers amid substantial patient heterogeneity. Methods range from simple data concatenation and consensus clustering to latent-variable, network-fusion, and probabilistic pathway models.

Disease subtyping: Disease subtyping clusters clinically heterogeneous patients into more homogeneous subgroups using clinical, molecular, and other patient features.More complete patient characterization is expected to produce more faithful and clinically relevant subtypes.
Disease subtyping: Clinical and molecular datasets can yield inconsistent patient subgroupings, while integrating gene expression and methylation resolved inconsistent medulloblastoma subgroups.The example illustrates why combining data types matters for subtype identification.
Disease subtyping: Concatenating all data types before clustering disregards dataset structure and can further dilute weak signals.This is the simplest commonly used integration strategy.
Disease subtyping: Cluster-Of-Cluster-Assignments first clusters each data type and then reclusters binary assignment vectors, but mainly captures common rather than complementary signals.It was used to integrate five TCGA breast-cancer data types.
Disease subtyping: iCluster jointly estimates a sparse latent space from multiple data types under the assumption that it captures the patients’ true subgrouping.Its Gaussian latent-variable model uses Lasso-type sparsity regularization.
Disease subtyping: Similarity Network Fusion combines patient-similarity networks from separate data types through iterative nonlinear diffusion rather than merging measurements directly.This avoids calibrating and comparing heterogeneous measurements in their original spaces.
Biomarker discovery: PARADIGM models gene activity as latent variables in a Bayesian network linking copy-number alteration, expression, protein levels, and protein activity.Comparing reference and disease posterior activity levels yields a quantitative measure of disease-induced alteration.

11. Challenges and future directions

Future progress in biomedical data integration depends on overcoming challenges in combining heterogeneous data, modeling biological complexity across scales, and interpreting machine-learning outputs. Emerging methods and richer phenotyping create opportunities to ask new biomedical questions, but many gold-standard approaches remain under development.

Opportunities: Emerging combinations of heterogeneous data and machine-learning methods enable fundamentally new biomedical questions, while gold-standard methods for several applications are still emerging.Examples include noncoding-variant identification, multi-omics cell profiling, and patient-population stratification.
Data integration challenges: Combining heterogeneous datasets requires careful normalization because technologies can produce structurally and distributively different data.Normalization is described as an essential first step for mixed-technology analyses.
Multi-scale and higher-order approaches: Multi-scale predictive cell models aim to connect molecular measurements with multiple phenotypes, but extending cell models to organisms remains unresolved.The paper also questions whether such models can capture the full extent of biological complexity.
Interpretability and explainability: Black-box machine-learning methods limit biomedical insight because their feature importance and feature–outcome relationships can be difficult to interpret.The review identifies transforming black-box methods into meaningfully interpretable white-box methods as a critical need.
Opportunities: Genomic data generation has become cheaper and faster than phenome-data collection, motivating broader and deeper phenotyping with lower participant burden.Internet-based phenotyping is presented as one fast approach, and studies have begun combining self-reported phenotypes with genomic information.

12. Conclusions

Machine learning approaches can integrate diverse biomedical datasets to better reflect biological complexity. The review concludes that methods should be selected for the domain, data, and outcome, and that integrative approaches may become central to systems biology and medicine.

Conclusions: No single machine-learning method is expected to perform best for every biomedical integration problem.Method selection should reflect domain-specific models, data types, and biomedical outcomes.
Conclusions: As integrative approaches become more available, systems biology and systems medicine are likely to become central computational strategies for generating biomedical knowledge.