[Paper Review] NeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models

Zhuoran Qiao, Feizhi Ding+11 more|arXiv (Cornell University)|Dec 14, 2024

Cell Image Analysis Techniques8 citations

TL;DR

NeuralPLexer3 (NP3) uses a physics-inspired flow-based generative model to predict biomolecular complex structures with state-of-the-art accuracy and faster inference than prior methods, including AlphaFold3. It achieves strong physical validity and ligand-induced conformational predictions across diverse biomolecular interactions.

ABSTRACT

Structure determination is essential to a mechanistic understanding of diseases and the development of novel therapeutics. Machine-learning-based structure prediction methods have made significant advancements by computationally predicting protein and bioassembly structures from sequences and molecular topology alone. Despite substantial progress in the field, challenges remain to deliver structure prediction models to real-world drug discovery. Here, we present NeuralPLexer3 -- a physics-inspired flow-based generative model that achieves state-of-the-art prediction accuracy on key biomolecular interaction types and improves training and sampling efficiency compared to its predecessors and alternative methodologies. Examined through newly developed benchmarking strategies, NeuralPLexer3 excels in vital areas that are crucial to structure-based drug design, such as physical validity and ligand-induced conformational changes.

Motivation & Objective

• Advance de novo prediction of generalized biomolecular complex structures (proteins, nucleic acids, ligands, ions, PTMs) from sequences and topology.
• Improve physical validity and ligand-induced conformational prediction in structure modeling.
• Enhance training and sampling efficiency over predecessors and competing methods.
• Set up benchmarks (PoseBusters, NPBench, ConfBench) to evaluate interaction types, conformational changes, and physicochemical plausibility.

Proposed method

• Utilize a conditional flow-based generative model with continuous normalizing flows to sample all heavy-atom coordinates from an informative prior.
• Incorporate physics-inspired priors via a globular polymer model with Langevin relaxation to generate plausible initial configurations.
• Apply flow matching with a symmetry-aware permutation module and vector-field reparameterization for stable, efficient training.
• Use an encoder-decoder architecture with anchor-based conditioning, MSA-derived features, and a diffusion-transformer decoder with geometric biases.
• Introduce Flash-TriangularAttention to reduce memory and inference time, enabling larger crop sizes and faster predictions.
• Benchmark NP3 against PoseBusters, NPBench, and CASP15 RNA targets, and report metrics including RMSD, PB-valid, LDDT, and DockQ.

Experimental results

Research questions

• RQ1 Can NP3 surpass AlphaFold3 in protein-ligand binding structure prediction across diverse biomolecular interactions?
• RQ2 How do physics-informed priors and efficient samplers affect physical validity and speed of structure predictions?
• RQ3 What is NP3’s performance on ligand-induced conformational changes and PTMs across multiple biomolecular modalities?
• RQ4 How well do new benchmarks (NPBench, ConfBench) reveal conformational and interaction prediction capabilities of advanced models?

Key findings

• NP3 achieves a 78.4% combined success rate across RMSD and PB-valid criteria on PoseBusters-V2, outperforming AlphaFold3’s 73.1%.
• NP3 attains 80.2% RMSD < 2 Å for ligand-bound predictions, comparable to AlphaFold3’s 80.4%, and 98.8% accuracy in predicting ligand stereochemistry.
• NP3 provides faster inference, delivering results in approximately 30 seconds on a single L40S GPU, versus around six A100-minutes reported for AlphaFold3, and achieves 15× faster predictions relative to AF3 timing statistics.
• On PoseBusters, NP3 achieves 52.7% success for low-homology protein-protein interfaces with DockQ > 0.23 and 87.1% LDDT for protein monomers, closely matching AlphaFold2-Multimer results on similar tasks.
• NP3 shows strong performance across diverse targets in NPBench, CASP15 RNA targets (46.5% vs 47.3% for AF3), and protein-nucleic acid interfaces, indicating broad generalization.
• ConfBench benchmarking indicates NP3 outperforms AF2-M for apo/holo pocket conformations, with 67.4% apo and 69.9% holo correctness overall; kinase-focused results show 77.8% apo and 72.5% holo correctness.