Research map / Probabilistic and scientific modeling / Molecules and proteins
NeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models
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.
Presents NeuralPLexer3, a physics-inspired flow-matching model for biomolecular complex structure prediction, plus NPBench and ConfBench benchmarks, reporting AF3-level accuracy with ~30 s inference on one L40S GPU.
Problem
Biomolecular complex structure prediction for drug discovery struggles with physically implausible outputs, high computational cost, and limited evaluation of ligand-induced conformational changes.
Approach
A conditional flow-matching model with an encoder-decoder architecture, a globular polymer Langevin-relaxed prior, optimal-transport symmetry correction (graph-preserving atom permutation), and rigid-alignment vector field reparameterization, plus confidence training without diffusion rollouts.
Key results
- NP3 achieves 78.4% combined PB-valid success on PoseBusters versus AF3's 73.1%, with 80.2% RMSD<2Å success (AF3 80.4%).
- NP3 generates predictions in about 30 seconds on a single L40S GPU, reported as 15x faster than AF3's roughly six A100-minutes.
- On ConfBench, NP3 reaches 51.9% correctness on 54 apo/holo linkages with conformational change versus 29.6% for AF2-M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeuralPLexer3 instead of AlphaFold 3 | protein-ligand complex prediction · PoseBusters-V2 | combined success rate: 73.1% -> 78.4% | Helped |
| Use NeuralPLexer3 instead of AlphaFold 3 | protein-ligand complex prediction · PoseBusters-V2 | ligand RMSD <2 Å success rate: 80.4% versus 80.2% | No clear effect |
| Use NeuralPLexer3 instead of AlphaFold2-Multimer v2.3 | protein-protein interface prediction · NPBench protein-protein interfaces | DockQ >0.23 success rate: 52.3% -> 52.7% | No clear effect |
| Use NeuralPLexer3 instead of AlphaFold2-Multimer | protein-peptide interface prediction · NPBench protein-peptide interfaces | success rate: 76.6% -> 85.1% | Helped |
| Use NeuralPLexer3 instead of AlphaFold 3 | RNA structure prediction · CASP15 RNA targets | LDDT: 46.5% versus 47.3% | No clear effect |
| Use NeuralPLexer3 instead of AlphaFold2-Multimer | ligand-induced conformation prediction · ConfBench apo/holo pairs with changes >1.5 Å | conformation correctness rate: 29.6% -> 51.9% | Helped |
| Add weighted interface-based training data sampling and cropping (ablation) | NeuralPLexer3 · PoseBusters | accuracy | Helped |
| Add AdaLN-Zero initialization (ablation) | NeuralPLexer3 flow blocks · PoseBusters | accuracy | Helped |
| Add QK-Norm (ablation) | NeuralPLexer3 attention layers · PoseBusters | accuracy | Helped |
| Use the improved rigid structure alignment weighting scheme (ablation) | NeuralPLexer3 training · PoseBusters | accuracy | Helped |
| Add clash and chirality penalties to conformer ranking (ablation) | NeuralPLexer3 sample ranking · PoseBusters | accuracy | Helped |
| Use Flash-TriangularAttention instead of a naive implementation | NeuralPLexer3 attention · large-structure training | peak memory usage: 5× reduction | Helped |
Surprising findings
- NP3 matches AF3 on CASP15 RNA (46.5% vs 47.3%) while conditioning on RNA language models rather than RNA MSAs.
Limitations
- ConfBench scores for NP predictions remain imperfect
- Apo structures used for scoring do not always imply fully ligand-free structures
- NPBench evaluation set is small (1,143 targets) and some ligand interfaces remain challenging (60.7% noncovalent success on the harder subset)
Experimental setup
Datasets: PDB structures before 2020-09-01 OpenProteinSet OrbNet Denali conformers PoseBusters-V2 CASP15 RNA targets PLINDER
Benchmarks: PoseBusters-V2 NPBench CASP15 RNA ConfBench
Metrics: ligand RMSD < 2 Å success rate PB-valid success rate LDDT DockQ > 0.23 success rate ConfBench score
Baselines: AlphaFold 3 NeuralPLexer2 AlphaFold2-Multimer v2.3 Vina GOLD Uni-Mol EquiBind TankBind DiffDock Alchemy_RNA2
Builds on: AlphaFold3 NeuralPLexer2 flow matching diffusion transformer Chinchilla scaling laws
Closest papers by meaning
- ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design (ICLR 2025 · similarity 0.87)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.86)
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.86)
- Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.86)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.