Research map / Probabilistic and scientific modeling / Molecules and proteins

Unified all-atom molecule generation with neural fields

NeurIPS 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 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 FuncBind, a modality-agnostic latent neural-field score-based generative model that generates target-conditioned all-atom small molecules, macrocyclic peptides and antibody CDR loops with a single 5B model, and releases a new MCP benchmark.

Problem

Structure-based generative models are usually restricted to one molecular modality (small molecules, proteins, peptides), limiting transfer and applicability; few methods handle non-canonical amino acids for macrocyclic peptides.

Approach

Represent molecules of any modality as continuous atomic-density neural fields encoded into a spatially arranged latent grid, then train a conditional 3D U-Net score-based denoiser (diffusion or walk-jump sampling) on these latents conditioned on target structure and modality, with no SE(3) equivariance but data augmentation.

Key results

• On CrossDocked2020, FuncBind matches SOTA-level results (VinaMin avg -6.34, QED 0.50) but trails VoxBind and MolCraft on docking scores and strain energy.
• On SAbDab CDR inpainting, FuncBind reaches 86.9% AAR on H1 and 47.5% on H3 with RMSD 0.41 and 2.04 Å, outperforming AbDiffuser and DiffAb on AAR and RMSD.
• In vitro, 42% of 190 de novo H3 designs bound the rigid epitope (mean pKD 9.56), while only 2% bound the flexible epitope.

Measured effects reported in the paper

Surprising findings

• Uniqueness was a confound: higher uniqueness of generated CDR sequences correlated with lower AAR, so repeated samples were used as a confidence proxy.
• Unified multi-modality model had higher uniqueness than specialized models with only small differences elsewhere.

Limitations

• Flexible epitope in vitro binding rate was only 2% (4% relaxed), and generated designs match seed lengths in the main evaluation.
• Relies on an accurate target structure and does not handle synthesizability or developability.
• Macrocyclic peptide baselines are limited, and FuncBind underperforms on Tanimoto similarity.

Experimental setup

Datasets: CrossDocked2020 SAbDab RCSB PDB macrocyclic peptide complexes (curated)

Benchmarks: CrossDocked2020 SAbDab CDR inpainting (DiffAb splits) macrocyclic peptide benchmark (new)

Metrics: VinaScore VinaMin VinaDock QED SA diversity strain energy clash AAR RMSD

Baselines: AR Pocket2Mol DiffSBDD TargetDiff DecompDiff MolCraft VoxBind RAbD DiffAb AbDiffuser AbX dyMEAN

Builds on: Kirchmeyer et al. neural field molecule representation VoxBind EDM / Karras diffusion walk-jump sampling multiplicative filter networks

Closest papers by meaning

• Score-based 3D molecule generation with neural fields (NeurIPS 2024 · similarity 0.90)
• UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.88)
• CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.87)
• Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.87)
• Structure-based drug design by denoising voxel grids (ICML 2024 · similarity 0.87)
• Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.87)