NeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models

TL;DR

NeuralPLexer3 (NP3) tackles the challenge of accurate biomolecular complex structure prediction, including ligand-induced conformational changes, by employing a physics-informed flow-based model with continuous normalizing flows and flow matching. The method integrates a globular polymer prior, symmetry-corrected prior alignment, and an anchored encoder–decoder architecture to enable fast, high-fidelity sampling of multi-molecule assemblies. NP3 delivers state-of-the-art accuracy on protein–ligand complexes and broad coverage across nucleic acids, covalent modifications, and protein–protein interactions, while achieving substantial speed-ups over prior diffusion-based approaches. The framework is validated through new benchmarks (NPBench and ConfBench) and shows practical promise for target validation, structural hypothesis generation, and atom-level interaction analysis in drug discovery, with efficient operation on standard hardware.

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.

Figures (11)

• Figure 1: NeuralPLexer3 (NP3) accurately predicts biomolecular structures with improved physical quality and prediction speed. (A) Schematics of the NP3 system. To perform a prediction, NP3 uses molecular topology extracted from input biopolymer sequences and small molecule graphs as primitive inputs, with additional conditioning signals from sequence language models, multiple sequence alignments (MSAs), and templates. NP3 adopts a flow-matching framework that samples from an informative prior, which includes basic physical restraints on atom configurations, relaxed by Langevin dynamics using a globular polymer model with harmonic connectivity terms (\ref{['supp:si']}, \ref{['algo:a3:Globular']}). During training, the prior samples are permuted to better align with the ground truth structure. (B) Performance of NP3 on the PoseBusters benchmark buttenschoen_posebusters_2024. Left panel: success rate for predicting the ligand-protein structures to within 2 Å RMSD, with and without additionally requiring that the structures are physically reasonable (PB-valid). Center panel: percentage of predicted structures where ligand stereochemistry is correct. Right panel: timing for running a single inference on a 1024-residue protein. Comparisons are made to NeuralPLexer2 (NP2) and AlphaFold 3 (AF3).
• Figure 2: NP3 pushes the frontier of structure prediction accuracy and training efficiency. (A) Comparison of relative encoder/decoder capacity in terms of floating-point operations per second (FLOPs) among different methods. The asterisk indicates an estimate based on our reproduction of AF3. (B) Model scaling behavior across molecule and interface types as reported by the relation between the average validation set local distance difference test (LDDT) mariani2013lddt scores and the total training floating-point operation count (FLOPs). While intra-molecular prediction accuracy tends to saturate upon reaching a critical model size, inter-molecular interaction prediction benefits from jointly scaling the model and data up to the production model size. (C) Key model training and inference improvements included during the full course of model development and their impact on PoseBusters accuracy. $C$: Compute in GFLOPs; $P$: number of decoder replicas.
• Figure 3: Model confidence estimation. (A) Scatter plot of ligand RMSD against ligand pLDDT and on PoseBusters. (B) PoseBusters ligand RMSD success rate statistics grouped into NP3 confidence percentiles. Prediction success rates are consistently higher for higher confidence prediction bins. (C) Scatter plot of DockQ score against the pDockQ score on NPBench protein-protein and protein-nucleic acid interfaces. (D) PPI DockQ success rate statistics grouped into NP3 confidence percentiles. Vertical lines on the scatter plot indicate the median prediction confidence.
• Figure 4: Illustration of Flash-TriangularAttention kernel and experimental comparison on the peak memory usage and the inference time. (A) The workflow for Flash-TriangularAttention. The main goal is to reduce peak memory usage, which is the bottleneck of triangular attention. Our implementation avoids this explicit broadcast, which is required while using other memory-efficient implementations, thus enabling longer crop size training. (B) Peak memory usage comparison between Flash-TriangularAttention and PyTorch built-in SDPA. Our implementation significantly reduces peak memory usage for longer residues, enabling samples with residues exceeding $1024$ to run on one H100 GPU (80GB). (C) the inference time of the Flash-TriangularAttention and Pytorch built-in SDPA. The figure shows that our implementation significantly reduces inference runtime—for example, by $38\%$ at a residue of $1024$.
• Figure 5: ConfBench conformational prediction. (A) Schematic illustration of types of conformational changes of interest. (B) Benchmark target statistics. 57k global and 52k pocket rearrangement entries with meaningful motions and biologically relevant ligands are identified. (C) The ConfBench scoring protocol. (D) Conformational prediction success rate statistics comparison. A correct conformational prediction is defined as a ConfBench score greater than 0. NP3 outperforms AF2-M 2.3 in predicting pocket conformations on apo targets, while maintaining similar success rates on apo targets and holo targets.