Challenging Federated Learning for Drug–Gene Interaction Prediction: A Comparative Study of FedAvg and FedProx Feasibility
Md. Jisan Mashrafi, Hasan Ahamed Alif, Urmi Haldar, Md Mohibur Rahman, Md Alamgir Miah, Mohammad Hasibul Hasan
Abstract
Drug-gene interaction (DGI) prediction is essential for precision medicine, but centralized model training raises privacy and regulatory concerns. This study uses two approaches to evaluate federated learning (FL) for modeling sparse bipartite drug–gene networks. The first integrates graph convolutional encoders with differential privacy and homomorphic encryption simulations to ensure data confidentiality. The second applies FedAvg and FedProx across simulated institutions, enhanced with gradient sparsification and quantization. Comparative experiments show that FL approaches underperform centralized models by 30–40% in ROC-AUC and incur significant communication costs under high sparsity. Differential privacy further reduces model expressiveness, and FedProx yields only marginal convergence improvement relative to FedAvg. These results indicate that, under extreme data heterogeneity and sparsity conditions, FL provides limited practical benefit for DGI tasks. Our findings highlight the trade-offs between privacy, scalability, and predictive performance, suggesting centralized architectures remain preferable for high-stakes biomedical applications.