Optimization of GFlowNet Parameters for Molecular Generation in Drug Design

Conditioned GFlowNets are characterized by combinatorial parameters needing carefully selected values, with impaired performance when poorly combined. To address this gap, we propose a novel hybrid of metaheuristics for multi-objective optimization of a subset of conditioned GFlowNet. We specifically adapted a biology-based optimizer to jointly maximize three objectives to learn the combined parameter values suitable to fully train GFlowNet conditioned to pockets of protein structure. Thereafter, drug-like molecules are derived from the model. Using the CrossDock202 and docking benchmark version 5 (DB5), we experimentally evaluated the proposed model.