Research map / Reinforcement learning / Visual and representation RL

Deep Reinforcement Learning for Modelling Protein Complexes

ICLR 2024 · Reinforcement learning › Visual and representation RL · approach: RL with tree search for discovery · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GAPN, a PPO-trained generative adversarial policy network that assembles multi-chain protein complexes one docking step at a time, using an RMSD-based reward plus a GCN discriminator adversarial reward.

Problem

Protein complex modelling (PCM) via one-by-one assembly of chains from dimer structures is hard because the docking-path search space grows as N^(N-2) and training data for large complexes is scarce, hurting generalization across chain numbers.

Approach

Formulates chain docking as an MDP and trains a policy network with PPO using a domain-specific negative-RMSD reward plus an adversarial reward from a GCN discriminator that distinguishes ground-truth assembly graphs from generated ones, injecting global assembly rules learned across chain numbers.

Key results

• GAPN achieves mean TM-Score 0.95 / RMSD 1.16 for N=3 and 0.36 / 59.13 for N=26-30, best across all chain ranges versus MoLPC and AF-Multimer.
• Under GT dimers for 11<=N<=30, GAPN improves mean and median TM-Score by about 31% and 53% over MoLPC.
• Inference is about 0.85 s per chain, reported as roughly 600x faster than MoLPC (540 s per chain).

Measured effects reported in the paper

Surprising findings

• Removing the adversarial reward significantly degrades performance, with the gain growing as chain number increases.
• Performance on the large-date-split test set is not substantially worse than on the main test set, though AFM dimers reduce accuracy noticeably.

Limitations

• Assumes dimer structures are given and only optimizes the docking path, not the dimers themselves.
• Performance drops sharply for large complexes (RMSD around 50-60 Å for 16-30 chains) and with AFM-predicted dimers.
• Small test set (180 complexes) due to the cost of running baselines.

Experimental setup

Datasets: PDB complexes (first biological assembly, filtered, CD-HIT clustered)

Benchmarks: 180-complex PDB test set (3 to 30 chains) larger date-split test set

Metrics: TM-Score RMSD inference time per chain (T/N)

Baselines: AlphaFold-Multimer ESMFold Multi-LZerD RL-MLZerD MoLPC

Builds on: proximal policy optimization generative adversarial networks ESM MoLPC

Closest papers by meaning

• Protein Multimer Structure Prediction via Prompt Learning (ICLR 2024 · similarity 0.88)
• Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.85)
• Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN (ICML 2025 · similarity 0.84)
• PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.84)
• PSBench: a large-scale benchmark for estimating the accuracy of protein complex structural models (NeurIPS 2025 · similarity 0.83)
• PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.83)