BRIDGE (protein interfaces)

BRIDGE — Biophysical Representation of Interfaces via Delaunay-based Graph Embeddings — is a coarse-grained graph neural network that learns to describe protein-protein interfaces, released in April 2025 by Prescient Design, the machine learning group inside Genentech. It is a distinct model from the histology and spatial-transcriptomics foundation model of the same acronym; the two share only a name.

The design problem BRIDGE addresses is practical. Predicting therapeutically important properties of an antibody — how tightly it binds its antigen, how viscous it becomes at high concentration — normally requires a structure of the bound complex, and those structures are rarely solved until very late in a program. BRIDGE sidesteps this by learning interface biophysics from a large corpus of experimentally determined complexes, then applying the learned representation to the structures of individual chains in isolation. No docking step is needed: the pretraining task itself acts as a meta-docking operation in graph space.

The pretraining objective is deliberately simple. Each protein chain is reduced to its alpha-carbon atoms, and a Delaunay tessellation over those coordinates defines the graph edges. The network is trained to predict which residue pairs across two chains form Delaunay contacts at the interface. Because Delaunay tessellation prunes spurious long-range connections — for the antibody-antigen complex 6O39, it cuts 7,959 residue pairs within 15 Å down to 2,410, a 69.7% reduction — the graphs stay small and training stays cheap.

#Key Features

• Delaunay-tessellation graphs: Chains are represented as alpha-carbon graphs whose edges come from Delaunay tessellation with a 15 Å cutoff, which removes most spurious contacts while preserving genuine packing relationships.
• Frozen embeddings reused across tasks: The same pretrained encoder produces the representations used for both antibody-antigen affinity regression and antibody viscosity regression, with only a lightweight feed-forward head trained per task.
• No complex structure required: Both downstream predictors take the structures of the individual molecules on their own, so they can be applied before any co-crystal or docked model exists.
• Extremely small parameter budget: The encoder has 847K trainable parameters, orders of magnitude below protein language models such as ESM-3, and trains on a single NVIDIA Quadro P6000 GPU.

#Technical Details

The encoder uses two independent stacks of GATv2 and GCN graph convolution layers followed by fully connected layers, with layer normalization and dropout, producing per-residue embeddings from 27 learnable residue tokens. Interchain contacts are scored by a dot product between residue embeddings from the two chains and a sigmoid, trained with binary cross-entropy against negative sampling. Pretraining used 42,112 chain pairs from DIPS-Plus, a database of experimentally derived structures, reaching a validation F1 of 0.84, precision 0.919, recall 0.775, and AUROC 0.854.

For affinity, embeddings of the two chains are combined by an outer product and flattened into a pair representation fed to a feed-forward regressor. Training used a SAbDab subset with surface plasmon resonance measurements, deduplicated to 1,172 structures after removing 42 PDB entries that overlap the Pierce Lab benchmark. BRIDGE-Affinity reaches Pearson R 0.697 on validation and R² 0.154 on the held-out Pierce Lab set — second only to DG-Affinity (R² 0.367), which used an additional non-public training set, while classical scoring functions such as ZRANK and PyDock post strongly negative R². For viscosity, Fv structures were built with ABodyBuilder2 and the same graph was passed twice to model antibody self-interaction; BRIDGE-Viscosity reaches Spearman R 0.781 on a 75:25 split and 0.978 under leave-one-out cross-validation, above PfAbNet's 0.71.

#Applications

BRIDGE targets large-molecule therapeutic discovery, where developability liabilities need to be triaged early and structural data is scarce. Antibody engineering teams can rank candidates for antigen affinity and for high-concentration viscosity — the property that decides whether a molecule can be formulated for subcutaneous injection — from sequence-derived monomer structures alone. Because the encoder is frozen and inexpensive to query, it also serves as a general-purpose interface descriptor for screening large candidate pools, and the authors point to binder generation and proteome-scale interaction prediction as natural extensions.

#Impact

BRIDGE is a demonstration that a carefully chosen geometric prior can substitute for scale: a sub-million-parameter graph encoder, pretrained on one self-supervised task, transfers to two unrelated therapeutic property predictions and beats the published state of the art on viscosity. The code for training the encoder and both downstream heads is released under Apache-2.0, though no pretrained checkpoint has been published, so reproducing the embeddings requires rerunning pretraining. The viscosity and affinity datasets remain small — 27 approved antibodies plus 38 anti-PDGF variants in the viscosity set — and the work is a preprint that has not been peer reviewed. Within those bounds it makes a useful argument for compact, structure-aware models in antibody engineering.