PGCA FL: A PATHWAY CONSENSUS-GUIDED FEDERATED LEARNING FRAMEWORK FOR PRIVACY-PRESERVING DISTRIBUTED GENOMIC CANCER CLASSIFICATION
TL;DR
This study introduces a privacy-preserving approach to multi-institution genomic cancer classification, called PCGA-FL (Pathway Consensus Guided Federated Learning), that incorporates biological pathway knowledge into federated optimization.
Abstract
The growing amount of genomic data has accelerated precision medicine by making it possible to analyze genomic data through artificial intelligence (AI) and classify cancers by their genomes. Yet, there are privacy regulations, institutional ownership of data and restricted access to distributed health resources issues that detract from centralized genomic learning. While federated learning (FL) offers a potential solution, the current federated learning paradigms are primarily statistical, and lack the incorporation of biological relationships between genes and pathways. This study introduces a privacy-preserving approach to multi-institution genomic cancer classification, called PCGA-FL (Pathway Consensus Guided Federated Learning), that incorporates biological pathway knowledge into federated optimization. This proposed framework features a Pathway Consensus Index (PCI) to measure pathway-level agreement across clients, a Pathway Aware Local Representation Learning (PLRL) approach to extract biologically meaningful genomic features and a Consensus-Guided Federated Aggregation (CGFA) algorithm for adaptive global model optimization. The framework is tested with the publicly available Cancer Genome Atlas (TCGA) RNA-sequencing (RNAseq) dataset for multiple cancer types, where a multi-institutional federated learning environment is emulated by splitting genomic data across a number of institutional clients. Experimental results show that PCGA-FL gets 98% accuracy, 94% precision, 93% recall, 93.5% F1 score, and 0.96 AUC, and enhances the convergence efficiency and communication overhead reduction of conventional federated learning methods. The proposed framework illustrates how to combine the power of consensus between biological pathways and privacy-preserving federated learning for privacy-preserving and interpretable cancer classification in a simulated multi-institutional federated learning setting.