Unified all-atom molecule generation with neural fields

AI summary

Overview

Research area: Machine learning for structure-based drug design, specifically generative modeling of molecules conditioned on protein target structures.

Technical level: Advanced. The paper assumes familiarity with score-based generative models (diffusion, denoising), latent variable models (VAEs), neural fields/implicit representations, and protein structural biology concepts such as complementarity-determining regions (CDRs) and non-canonical amino acids.

Scope: The paper presents FuncBind, a single unified generative framework based on neural fields that can design small molecules, macrocyclic peptides, and antibody CDR loops conditioned on 3D target structures, with both in silico benchmarks and wet-lab validation.

What This Paper Is About

Generative models for drug design are typically built for one molecular modality at a time—small molecules, peptides, or proteins—using modality-specific representations (point clouds, voxels, or residue graphs). This specialization prevents knowledge transfer across modalities and forces researchers to build separate pipelines.

FuncBind addresses this by representing any molecule as a continuous atomic density field (a neural field) and training one score-based generative model over a spatially arranged latent space, enabling a single model to design chemically diverse binders against a target protein.

Key Contributions

1. A modality-agnostic neural field representation with spatial latent structure. Unlike prior work that used a single global embedding to modulate the neural field decoder, FuncBind arranges latents as a spatially ordered feature map grid, enabling the use of CNN/U-Net architectures and scaling to larger molecules such as proteins.

2. A unified conditional generative model trained across three modalities simultaneously. A single 5B-parameter model generates small molecules, macrocyclic peptides (MCPs, including non-canonical amino acids), and antibody CDR loops, handling variable atom and residue counts without modality-specific architecture changes.

3. Competitive or state-of-the-art in silico results plus wet-lab validation. FuncBind matches or exceeds modality-specific baselines on CrossDocked2020 for small molecules, substantially improves CDR H3/loop redesign metrics (1.5–3× on amino acid recovery and RMSD), and produces antibody binders validated by surface plasmon resonance.

4. A new macrocyclic peptide dataset and benchmark. Approximately 190,000 synthetic MCP–protein complexes derived from 641 RCSB PDB structures via a "mutate then relax" protocol, together with new evaluation metrics (Tanimoto similarity, ligand/interface RMSD, TM-score, Vina docking), designed to spur future structure-conditioned MCP generation research.

Main Findings

• Small molecule generation is competitive but not dominant. On CrossDocked2020, FuncBind produced highly diverse molecules but slightly underperformed VoxBind and MolCraft on Vina docking scores and strain energy, and VoxBind on steric clashes. The unified model generated notably higher uniqueness than modality-specialized baselines.

• Antibody CDR redesign is the strongest in silico result. FuncBind achieved roughly 1.5–3× higher amino acid recovery and lower Cα RMSD than DiffAb, AbDiffuser, dyMEAN, AbX, and Rosetta-based RAbD across H1, H2, H3, L1, L2, and L3 loops, despite not applying Rosetta relaxation within its pipeline. Interface energy improvement (IMP) was competitive even without backbone minimization, and improved substantially with post-hoc relaxation.

• Wet-lab validation succeeded. From 10,000 unique generated H3 designs, 190 were synthesized and expressed. Binding rates were 45% for a rigid epitope and 2% for a flexible epitope (rising to 4% with a relaxed threshold), confirmed by SPR measurements.

• Macrocyclic peptide results are strong on geometry and docking. FuncBind achieved lower ligand and interface RMSD than RFPeptide and AfCycDesign, and produced the highest fraction of designs with better Vina docking scores than the seed (41%). Tanimoto similarity was slightly lower, attributed to FuncBind exploring a broader space of non-canonical amino acids and sequence lengths.

• Non-canonical amino acids can be generated de novo. Fewer than 1% of generated amino acids were flagged as chemically "unreasonable" (bad bond geometry), and the model produced novel amino acids not present in the training library that engaged pocket residues the seed and close analogs did not.

• Variable-length design works. FuncBind is not constrained to the seed CDR length; generated H3 length and atom count distributions centered around the seed values, enabling de novo length variation.

Methodology in Plain English

FuncBind repurposes techniques from 3D computer vision. Instead of representing a molecule as a list of atom coordinates, it converts the molecule into a continuous density field on a low-resolution voxel grid—each point in space carries values for how likely each atom type is present there. A 3D convolutional encoder compresses this grid into a spatially organized latent grid (rather than a single vector), which is more compact than voxels and better able to capture local geometry in large molecules.

The decoder reconstructs the atomic density field from any coordinate and the local latent vector at that location, using a multiplicative filter network with Gabor filters—well-suited to sparse atomic densities. This autoencoding step is trained with a reconstruction loss plus a KL regularization term, similar to a VAE.

A second network (a 3D U-Net adapted from image diffusion models) is then trained as a conditional denoiser over the latent space. It receives a noisy latent, a separately encoded target structure, a one-hot modality label, and the noise level, and is trained to output the clean latent—equivalent to learning the score function via the Tweedie–Miyasawa formula. Sampling is done either by reverse-time SDE integration (EDM-style diffusion) or by walk-jump sampling, which uses only one noise level and is faster to train.

To recover chemically valid molecules from generated density fields, FuncBind detects local peaks (via MaxPooling and gradient ascent through the differentiable field) and then uses OpenBabel to infer bonds and residue identities, with a custom procedure for non-canonical amino acids.

Why This Matters

Impact on research. The paper challenges the prevailing assumption that generative drug design requires modality-specific architectures and representations. By demonstrating that one neural field model can handle small molecules, cyclic peptides with non-canonical residues, and antibody loops, it opens the door to transfer learning across chemical spaces and to training on much larger, more heterogeneous datasets. The MCP benchmark fills a real gap—prior structure-conditioned peptide generators could not handle non-canonical amino acids, which are central to real-world peptide therapeutics.

Real-world applications:

• Small molecule hit generation: designing novel ligands inside a target pocket for early-stage drug discovery.
• Antibody engineering: de novo design or optimization of CDR loops against a known epitope, potentially shortening lead discovery cycles.
• Peptide therapeutics: designing macrocyclic peptides that span a chemical space between small molecules and biologics, including non-canonical residues for stability and potency.
• Cross-modality design: enabling generative tasks that mix modalities, such as designing small molecules to disrupt a protein–peptide interface.

Industry relevance. The work comes from Genentech's Prescient Design and Antibody Engineering groups, and antibodies represented 26% of 2024 FDA approvals. A unified, validated generative framework that operates on real drug modalities—and produces experimentally confirmed binders—has direct relevance to pharmaceutical pipelines where separate models currently add engineering overhead.

Future Directions

• Broader transfer learning. The authors observed parity between unified and specialized models on most metrics, with gains mainly in uniqueness—testing whether adding more modalities (e.g., DNA, sugars, cofactors) improves cross-modality transfer remains open.
• Variable-length and de novo design validation. FuncBind can sample CDR loops of varied lengths, but only seed-length designs were experimentally validated. Whether these variable designs bind well is unresolved.
• Improving flexible-epitope performance. The large gap between rigid (45%) and flexible (2%) epitope binding rates suggests that flexible targets are a major unsolved challenge for this and related methods.
• Benchmarking non-canonical amino acid generation. The new MCP benchmark invites comparison with future models that, like FuncBind, drop the constraint of a predefined amino acid library, and may reveal whether unconstrained generation translates into better experimental outcomes.

Target Audience

This paper is most valuable to machine learning researchers working on generative models for structural biology and drug discovery, computational chemists and protein engineers interested in antibody or peptide design, and industrial R&D teams building modality-specific design pipelines who want to understand whether a unified model can replace them. Readers without background in score-based generative models or protein structural biology will find the methods sections demanding, though the experimental results are accessible.

Authors’ abstract

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.