PaperlayerView exact revision on arXiv

Source-linked AI summary

Unified all-atom molecule generation with neural fields

Matthieu Kirchmeyer, Pedro O. Pinheiro, Emma Willett, Karolis Martinkus, Joseph Kleinhenz, Emily K. Makowski, Andrew M. Watkins, Vladimir Gligorijevic, Richard Bonneau, Saeed Saremi

arXiv:2511.15906v1cs.LGq-bio.BM

TL;DR

Structure-based generative models often focus on one molecular modality, despite drug-design applications spanning diverse atomic systems. FuncBind combines neural fields with score-based generation in a unified, target-conditioned approach across small molecules, macrocyclic peptides, and antibody CDR loops. It reports competitive in-silico performance, in-vitro antibody binders, and a new macrocyclic peptide benchmark dataset.

  • Problem

    Structure-based generative models often specialize in one molecule modality, limiting transfer across modalities despite the relevance of multi-modality applications.

    Source: §1 · pp. 1–2

  • Method

    FuncBind represents molecules as neural fields of atomic densities and trains a latent conditional score-based model for all-atom generation across modalities.

    Source: §1 · p. 2

  • Results

    FuncBind matches or outperforms modality-specific baselines in silico across three modalities and produced novel antibody binders in vitro through CDR H3 redesign of two co-crystal structures.

    Source: §1 · p. 2, §5 · p. 10

  • Takeaways & Limitations

    The paper introduces a macrocyclic peptide–protein dataset and benchmark to address the scarcity of established data and benchmarks for this task.

    Source: §4.3 · p. 9

    FuncBind relies on accurate molecular-interface models, which can be costly and restricted, and does not address properties beyond binding such as synthesizability or antibody developability.

    Source: §5 · p. 10

Abstract

from arXiv · show

Generative models for structure-based drug design are often limited to a specific modality, restricting their broader applicability. To address this challenge, we introduce FuncBind, a framework based on computer vision to generate target-conditioned, all-atom molecules across atomic systems. FuncBind uses neural fields to represent molecules as continuous atomic densities and employs score-based generative models with modern architectures adapted from the computer vision literature. This modality-agnostic representation allows a single unified model to be trained on diverse atomic systems, from small to large molecules, and handle variable atom/residue counts, including non-canonical amino acids. FuncBind achieves competitive in silico performance in generating small molecules, macrocyclic peptides, and antibody complementarity-determining region loops, conditioned on target structures. FuncBind also generated in vitro novel antibody binders via de novo redesign of the complementarity-determining region H3 loop of two chosen co-crystal structures. As a final contribution, we introduce a new dataset and benchmark for structure-conditioned macrocyclic peptide generation. The code is available at https://github.com/prescient-design/funcbind.

1 Introduction

Structure-based generative models often specialize in one molecular modality, although drug-design targets span diverse atomic systems. FuncBind uses neural fields and a unified model across three modalities, with competitive in-silico results and antibody binders reported in vitro.

1 Introduction

  • Structure-based drug design seeks molecules that bind a target protein, using its 3D structure to generate candidates with high affinity.The introduction frames generative models as a data-driven alternative to virtual screening and physics-based simulations.Source: §1 · p. 1
  • Most generative models specialize in one modality, while domain-specific representations limit transfer and utility for applications involving multiple modalities.The paper argues modality-agnostic representations can learn physical properties across diverse atomic systems and leverage more training data and challenging metrics.Source: §1 · pp. 1–2
  • FuncBind represents molecules as neural fields mapping 3D coordinates to atomic densities and uses a latent conditional generative model compatible with score-based approaches.The representation handles variable atom and residue counts implicitly and applies across modalities in an all-atom formulation.Source: §1 · p. 2
  • FuncBind matches or outperforms modality-specific baselines in silico across three modalities, produced novel antibody binders in vitro, and adds a ∼190,000-complex MCP dataset.The dataset comprises synthetic macrocyclic peptide–protein complexes.Source: §1 · p. 2

2 Related work

Related work spans point-cloud, voxel, and sequence-based approaches for structure-conditioned molecule design, often tailored to a single modality. The paper positions FuncBind’s neural-field approach as applicable across modalities and highlights remaining gaps in large-molecule scaling, antibody-pose assumptions, and macrocyclic peptide design.

2 Related work

  • Small-molecule generation uses point clouds with graph networks or voxel grids with computer-vision methods, but raw voxel models do not currently scale to larger molecules.Neural fields are continuous analogues to voxels and have matched existing performance while improving memory and computational efficiency.Source: §2 · p. 2
  • Antibody CDR design includes sequence-and-structure co-generation, language-model, and physics-constrained approaches; most assume a provided framework docking pose, unlike.The approach in generates backbone CDR atom positions with RFDiffusion, followed by inverse folding.Source: §2 · pp. 2–3
  • FuncBind is presented as the first neural-field approach applicable across data modalities simultaneously; its feature-map-grid architecture captures local signal and supports expressive denoising architectures.The figure contrasts a global embedding with spatially arranged embeddings.Source: §2 · p. 3, Figure 1 · p. 3
  • Prior macrocyclic peptide work includes target-conditioned generation, but one cited approach cannot handle non-canonical amino acids and another relies on property predictors that generalize poorly out of distribution.Rettie et al. use RFDiffusion; Tang et al. use tokenized SMILES and classifier guidance.Source: §2 · p. 3

3 Method

FuncBind learns spatial latent representations of molecules and trains a conditional denoiser to generate them from target structures. A neural-field decoder and postprocessing pipeline turn sampled latents into molecular structures.

3 Method

  • FuncBind learns molecule latents that modulate a neural-field decoder, then trains a conditional denoiser to sample molecules given a target.The framework uses two training steps, followed by target-conditioned sampling.Source: §3 · p. 3

3.1 Neural field-based latent representation

  • Neural fields represent molecules as coordinate-to-atomic-occupancy functions, while spatial feature maps capture local information and support scaling to larger molecules.The spatial latent representation also accommodates variable atom and residue counts implicitly.Source: §3.1 · p. 3
  • A shared neural-field decoder uses local latent features to produce atomic-density fields, trained across modalities with KL regularization and augmented rotations and translations.The decoder uses multiplicative filter networks with Gabor filters; training upweights coordinates near atom centers.Source: §3.1 · p. 4

3.2 Conditional latent score-based generation

  • The denoiser samples from the encoder’s aggregate posterior, conditioned on the target structure, molecule modality, and noise level.A separate target encoder supplies the target latent representation.Source: §3.2 · p. 4, §3.2.1 · p. 4

3.2.1 Conditioned denoiser

  • A 3D U-Net adapted from image generation predicts clean latents from noisy ones, using rotation and translation augmentation instead of built-in SE(3) equivariance constraints.The model conditions on the target latent, modality class, and noise level.Source: §3.2.1 · p. 4

3.2.2 Sampling strategies

  • Sampling experiments compare continuous-noise diffusion with walk-jump sampling; walk-jump trades easier sampling at larger noise for denoised samples farther from the target distribution.Diffusion samples by integrating a reverse-time SDE from maximum to minimum noise.Source: §3.2.2 · p. 5

3.3 Recovering molecules from generated atomic-density fields

  • Sampled latents become structures through voxel rendering and peak detection, followed by gradient refinement and bond or residue-identity inference.The pipeline uses OpenBabel for bond and applicable amino-acid inference, producing SDF files for molecules and peptides and PDB files for proteins.Source: §3.3 · p. 5

4 Experiments

FuncBind was evaluated on target-conditioned small molecules, antibody CDR loops, and macrocyclic peptides, with a new benchmark introduced for macrocyclic peptide generation. The experiments assess its performance across these modalities, including antibody wet-lab testing and designs with variable lengths and non-canonical amino acids.

4.1 Small molecule generation

  • FuncBind was competitive on CrossDocked2020, slightly underperforming VoxBind and MolCraft on docking-related metrics and strain energy, and VoxBind on clashes.The study sampled 100 ligands per pocket and assessed docking, drug-likeness, synthesizability, diversity, atom count, clashes, and strain energy.Source: §4.1 · p. 6, Table 1 · p. 7

4.2 Antibody CDR redesign

  • FuncBind outperformed other antibody-loop baselines by 1.5 to 3 times on amino acid recovery and Cα RMSD across all loops.Without backbone minimization, its interface energy improvement was competitive with approaches using minimization; applying fast-relax improved IMP but slightly increased RMSD.Source: §4.2 · p. 7, Table 2 · p. 8
  • For one de novo target, generated CDR H3 lengths varied but their length and atom-count distributions centered on the seed’s reference values.The authors leave validation of designs with other lengths for future research.Source: §4.2 · p. 7, Figure 4 · p. 8
  • 45% of tested designs bound the rigid epitope, versus 2% for the flexible epitope, rising to 4% with a relaxed binding threshold.The 190 tested designs were selected from 10,000 unique, seed-length H3 designs using model confidence and sequence-distance clustering.Source: §4.2 · p. 8

4.3 Macrocyclic peptide generation

  • FuncBind achieved superior or similar metrics to MCP baselines, with lower L/I-RMSDs and the highest proportion of designs with better docking scores, but lower Tanimoto similarity.Nearly half of generated samples had better Vina binding affinity than the seed; about 20% had TM scores above 0.5, and many RMSDs were below 1 Å, particularly I-RMSD.Source: Table 3 · p. 9, §4.3 · pp. 9–10
  • The new MCP benchmark contains 186,685 complexes, curated from 641 source complexes by mutation and relaxation; 78% contain non-canonical amino acids.Its non-i.i.d. test set comprises 85 protein pockets.Source: §4.3 · p. 9
  • Fewer than 1% of generated amino acids were labeled unreasonable, and a novel amino acid engaged pocket residues not engaged by the seed or a chemically similar residue.The authors attribute broader pocket exploration to generation without a predefined amino-acid library or initial cyclic-backbone constraints.Source: §4.3 · p. 10, Figure 5 · p. 10

5 Conclusion

FuncBind provides competitive in silico performance against specialized baselines across diverse molecular modalities and generates antibody binders in vitro. Its use depends on accurate interface models, and it does not address all properties needed for real-world drug development.

5 Conclusion

  • FuncBind matches or outperforms specialized baselines in silico across small molecules, macrocyclic peptides, and antibody CDRs.It generates chemically plausible molecules, including non-canonical amino acids.Source: §5 · p. 10
  • In vitro, FuncBind generates binders against de novo targets, including novel and chemically plausible molecules.The conclusion also notes generation of new non-canonical amino acids.Source: §5 · p. 10
  • Scaling behavior remains an open direction, although the denoiser showed no overfitting in tests up to 5B parameters.The conclusion also identifies extending the framework to larger biomolecular systems as future work.Source: §5 · p. 10
  • FuncBind relies on accurate molecular-interface models, which can be costly to obtain and restricted, particularly for large molecules.The work also does not address properties beyond binding, such as small-molecule synthesizability or antibody developability.Source: §5 · p. 10

Appendices

The appendices describe the supplementary material’s organization, broader impacts, representation and model implementation, and sampling procedures. They also explain multimeasurement walk-jump sampling and its finite-noise design.

Appendices

  • The supplementary material covers broader impacts, model and implementation details, antibody redesign results, non-canonical amino-acid inference, and macrocyclic-peptide results, data curation, and splits.Source: Appendices · p. 16

A Broader impacts

  • The broader-impact discussion identifies potential therapeutic-discovery benefits across modalities while noting the gap between in silico prediction and experimental or clinical validation.Source: A Broader impacts · p. 16

B.1 Representation

  • The representation defines a smooth atomic-density field, uses eight shared element types, and merges protein-specific atom labels to support transfer across modalities.The field is defined over a 32Å volume, corresponding to a 128-grid resolution of 0.25Å.Source: §B.1 · p. 16

B.2 Neural Field

  • The neural-field encoder and decoder each have 59M parameters; training applies random rotations and translations before voxelization.The encoder uses a low-resolution grid, while the decoder is a conditional multiplicative filter network with Gabor filters and FiLM-modulated layers.Source: §B.2 · pp. 16–17

B.3 Denoiser

  • The denoiser uses a 5B-parameter UNet, a 3D CNN target encoder, and noise-level preconditioning with log-normal noise sampling.Training uses early stopping and randomly drops conditioning information 10% of the time.Source: §B.3 · p. 17

B.4 Sampling

  • Sampling applies different pocket rotations across MCMC chains to improve uniqueness.Source: §B.4 · p. 17

B.4.1 Denoising diffusion

  • Sampling temperature is set to τ = 0.5 on Crossdocked and τ = 0.33 on MCP-protein complexes.Source: §B.4.1 · p. 17

B.4.2 Walk-Jump sampling

  • WJS-m accumulates noisy measurements so effective noise decreases as σ/√m, while sampling proceeds at a fixed noise scale and the jump denoises using the measurements’ empirical mean.The method requires score functions at only a finite noise level rather than across a continuum; vanilla WJS is WJS-1.Source: §B.4.2 · p. 18

C.1 In silico evaluation

The CDR sampling results examine sequence uniqueness and its relationship to amino-acid recovery. The authors use that relationship as a heuristic for selecting H3 designs for in vitro evaluation.

C.1.1 Uniqueness

  • Higher uniqueness usually corresponds to lower Amino Acid Recovery (AAR) in CDR sampling.Source: §C.1.1 · p. 18
  • Repeated sequences tend to correlate more with the seed, linking lower uniqueness with higher seed similarity.Source: §C.1.1 · p. 18
  • The authors use uniqueness as a heuristic to select H3 designs for in vitro evaluation.Table 4 reports CDR sampling results over unique sequences.Source: §C.1.1 · p. 18

C.1.2 Ablation with Walk Jump Sampling

The in silico comparison reports similar performance for diffusion and multimeasurement WJS, while the unified model has higher uniqueness than the specialized model. The in vitro study tests selected H3 redesigns based on two co-crystal structures.

C.1.2 Ablation with Walk Jump Sampling

C.1.3 Comparison to specialized model

C.2 In vitro validation

  • The in vitro validation redesigns H3 loops for rigid- and flexible-epitope complexes in a de novo setting, excluding similar interfaces from training.Source: §C.2 · p. 20
  • Of 10,000 unique generated H3 designs matching the seed length, 190 were selected for experimental testing.The authors shortlisted 500 designs by repeated-generation counts, then clustered them by weighted sequence edit distance and selected the highest-count design from each group.Source: §C.2 · p. 20

C.2.1 Rigid epitope

FuncBind generated antibody binders in the rigid-epitope redesign experiment, with most submitted designs expressed and purified. The section also reports AAR and RMSD analyses for selected designs.

C.2.1 Rigid epitope

  • Figure 6 presents AAR and RMSD histograms for all generated CDR H3 designs on the rigid epitope.Source: Figure 6 · p. 20, §C.2.1 · p. 20
  • 42% of the 190 submitted designs were binders with pKD values, and 94% were successfully expressed and purified.Across all submitted designs, 54% were binders, including 12% without pKD values; the mean pKD among those with values was 9.56.Source: §C.2.1 · p. 20
  • Table 8 reports average RMSD and AAR for rigid-epitope binders and non-binders.Source: Table 8 · p. 20
  • Figure 7 reports expression, binding affinity, and binding rate for the rigid-epitope designs, with an expression rate of 93.68%.Source: Figure 7 · p. 21

C.3 Flexible epitope

For the flexible epitope, the submitted designs all expressed, but few were confirmed binders. The section also describes how unidentified amino acids are recognized and labeled.

C.3 Flexible epitope

  • Figure 8 presents AAR and RMSD histograms and a sequence logo for designs against the flexible epitope.Source: Figure 8 · p. 21, §C.3 · p. 21
  • 2% of submitted designs were binders, although 100% expressed; measured binder pKDs ranged from 6.95 to 8.47 and averaged 7.74.Source: §C.3 · p. 21
  • Figure 9 reports expression, binding affinity, and binding rate for the flexible-epitope designs, with an expression rate of 100%.Source: Figure 9 · p. 22

D Identifying non canonical amino acids

  • Unidentified amino acids are detected from repeated peptide-backbone patterns around a chiral carbon, then matched by SMILES to a non-canonical amino-acid library or labeled unknown.For output PDB files, a canonical residue name is assigned by matching atom-naming patterns to a known canonical residue.Source: D Identifying non canonical amino acids · p. 22

E.1 Additional results

The additional results describe sampled macrocyclic peptides, amino-acid categories, and novel non-canonical residues, alongside visual summaries of the source and curated datasets.

E.1 Additional results

  • Figure 10 qualitatively compares sampled macrocyclic peptides with their seed, while Figure 10’s samples are shown in three mutant pockets.Source: §E.1 · p. 22, Figure 10 · p. 23
  • Table 10 compares per-residue Tanimoto similarity for molecules sampled with the seed mutant and with the crystal macrocyclic peptide.Source: §E.1 · p. 22
  • Figures 11 and 12 categorize sampled amino acids and show novel non-canonical amino acids absent from the initial test-set library.The novel residues are labeled with formal names and assigned three-letter amino-acid codes.Source: §E.1 · p. 22, Figure 11 · p. 23, Figure 12 · p. 24
  • Figure 5 highlights a newly generated amino acid interacting with pocket residues not engaged by either the seed or a chemically similar amino acid at that position.Source: §E.1 · p. 22
  • Figure 13 summarizes source-dataset macrocyclic-peptide lengths and mutation counts in the curated bound dataset.Source: Figure 13 · p. 24

E.2 Data curation

The curated macrocyclic-peptide dataset expands source complexes through mutation and relaxation, then partitions targets into train, validation, and test sets using sequence-similarity clusters.

E.2 Data curation

  • The curated dataset contains 186,685 MCP-protein structures generated by mutating source MCPs at 1–8 sites and selecting low-interface-score relaxed complexes.Mutations used 213 amino acids; mutations at disulfide- and S-acetyl-cysteine-linked residues were avoided, and about 500 of over 2,000 generated complexes per source structure were selected.Source: §E.2 · p. 24

E.3 Clustering and Splitting

  • Targets were clustered when their longest-common-subsequence similarity exceeded 0.5, yielding 208 clusters for data splitting.Source: §E.3 · pp. 24–25
  • Clusters with more than 100 MCP-protein pairs went to training; 100 remaining clusters were randomly assigned to testing and the rest to validation.Source: §E.3 · p. 25
Loading 2511.15906v1…