A GFlowNet-based world model for structural protein–protein interaction prediction on DB5

Structural prediction of protein–protein interactions (PPIs) is essential for understanding biological function and guiding therapeutic development. While deep learning models such as AlphaFold-Multimer have advanced this field, they often struggle with flexible, disordered, or low-homology complexes. In this work, we propose a novel world modelling framework for structural PPI prediction using Generative Flow Networks (GFlowNets). Our method frames molecular complex generation as a reward-driven trajectory sampling problem, guided by receptor protein embeddings and a domain-specific reward function incorporating docking score, drug-likeness (QED), and synthesizability (SAScore). We evaluate our method using the Docking Benchmark 5 (DB5) dataset, spanning rigid, medium, and difficult interaction classes. Compared to existing baselines, our model demonstrates improved structural diversity and competitive plausibility across scenarios. Findings from the study show that the flow method can generate synthesizable molecular structure with a high degree of QED and reduced docking score.