[논문 리뷰] A Variational Information Bottleneck Approach to Multi-Omics Data Integration

Changhee Lee, Mihaela van der Schaar|arXiv (Cornell University)|2021. 02. 05.

Bioinformatics and Genomic Networks참고 문헌 11인용 수 30

한 줄 요약

DeepIMV는 variational information bottleneck을 product-of-experts와 함께 활용해 불완전한 다중-오믹스 뷰를 감독 학습 예측에 유연하게 통합하고, 최첨단 벤치마크를 능가합니다.

ABSTRACT

Integration of data from multiple omics techniques is becoming increasingly important in biomedical research. Due to non-uniformity and technical limitations in omics platforms, such integrative analyses on multiple omics, which we refer to as views, involve learning from incomplete observations with various view-missing patterns. This is challenging because i) complex interactions within and across observed views need to be properly addressed for optimal predictive power and ii) observations with various view-missing patterns need to be flexibly integrated. To address such challenges, we propose a deep variational information bottleneck (IB) approach for incomplete multi-view observations. Our method applies the IB framework on marginal and joint representations of the observed views to focus on intra-view and inter-view interactions that are relevant for the target. Most importantly, by modeling the joint representations as a product of marginal representations, we can efficiently learn from observed views with various view-missing patterns. Experiments on real-world datasets show that our method consistently achieves gain from data integration and outperforms state-of-the-art benchmarks.

연구 동기 및 목표

• 이질적인 오믹스 데이터를 감독 설정에서 missing views와 함께 통합할 동기를 부여한다.
• 타깃 예측을 위해 intra-view 및 inter-view 상호 작용을 활용하는 통합 프레임워크를 개발한다.
• 모든 뷰를 임퓨트하지 않고 임의의 뷰-누락 패턴을 유연하게 처리할 수 있도록 한다.
• 정보 병목을 통해 작업 관련 표현을 학습하여 예측력을 향상시킨다.

제안 방법

• Four-network DeepIMV 아키텍처: V view-specific encoders, V view-specific predictors, a product-of-experts (PoE) 모듈, 및 a multi-view predictor.
• Marginal representations Z_v는 per-view에서 q_theta_v(z|x_v)로 학습되며; joint representation Z는 PoE over marginal posteriors로 형성된다: q_theta(z|X_bar) ≈ p(z) ∏_v∈V q_theta_v(z|x_v).
• Task-focused learning은 joint representation에 정보 병목 목표를 적용해 Y-prediction 신호를 최대화하고 입력 정보를 압축한다: L_IB-J = -I(Y;Z) + β I(X_bar; Z).
• Auxiliary marginal IB losses L_IB-M은 각 뷰에 대해 discriminative하고 task-relevant한 특징을 포착하도록 뷰-특정 인코더를 장려한다: -I(Y_v; Z_v) + β_v I(X_v; Z_v).
• Training은 joint 및 marginal IB 손실의 가중합을 최적화한다: L_Total = L_IB-J + α ∑_v L_IB-M.
• Efficient training with Gaussian marginals enabling closed-form PoE joint posterior, and deterministic/stochastic mappings for encoders and predictors.

실험 결과

연구 질문

• RQ1 Can incomplete multi-view omics data be integrated in a single framework without imputing missing views?
• RQ2 How can marginal (per-view) representations be combined to form a sharp, informative joint representation under view-missing patterns?
• RQ3 Does a product-of-experts joint posterior improve predictive performance over mixtures or stage-wise approaches in incomplete multi-view data?
• RQ4 Do view-specific predictors and marginal IB losses aid in learning task-relevant, view-specific information that benefits the joint model?

주요 결과

• DeepIMV consistently improves predictive performance when integrating incomplete multi-view omics data across TCGA and CCLE datasets.
• The Product-of-Experts formulation enables flexible handling of arbitrary missing-view patterns by ignoring missing views in joint latent construction, avoiding extra imputation steps.
• Incorporating view-specific IB losses and predictors helps encoders specialize in view-relevant information, facilitating the training of the PoE joint representation and improving accuracy.
• Across experiments, DeepIMV outperforms multi-view benchmarks that assume complete views or use reconstruction-focused latent spaces (e.g., MVAE, MOFA).
• Ablation studies show PoE plus marginal IB losses yield better AUROC than MoE-based variants, highlighting the value of task-relevant, view-specific representations.