TopoBind

TopoBind predicts the binding free energy (ΔG) of antibody-antigen complexes by combining what a protein language model knows about sequence with what topological data analysis can measure about the shape of an interface. It was developed by a Hong Kong consortium spanning the University of Hong Kong, City University of Hong Kong, the Hong Kong University of Science and Technology, the Chinese Academy of Sciences' Centre for Artificial Intelligence and Robotics, and BayVax Biotech, and released as a preprint in August 2025.

Measuring ΔG experimentally, by surface plasmon resonance or isothermal titration calorimetry, is slow and expensive, and physics-based estimation via molecular dynamics or MM/PBSA scales poorly. Sequence-only predictors built on protein language models are fast but have no spatial awareness; structure-based graph and surface models capture local geometry but usually ignore global topological invariants — the connected components, loops, and cavities that persist as a structure is viewed at coarser and coarser scales. TopoBind's contribution is to fuse both signals rather than choose between them, and to show through ablation that each contributes.

The model belongs to the family of thin, task-specific heads mounted on a pretrained upstream foundation model. The heavy lifting of sequence representation is done by ESM-2, whose embeddings are extracted by mean pooling; what the framework adds is the topology encoder, the fusion machinery, and the regressor.

#Key Features

• Persistent homology features: A Vietoris-Rips filtration over atomic point clouds yields birth-death pairs whose lifetimes are summarized as moments and top-k persistences, describing multiscale structure invariant to rigid motion and robust to noise.
• Four structural modalities: Contact-map metrics, interface geometry descriptors, distance-map statistics, and topological invariants are encoded separately and combined by a learnable adaptive fusion layer that weights each family per sample.
• Bidirectional cross-attention: Two stacked cross-attention layers align sequence and structure representations in both directions before a sigmoid gate fuses them, rather than simply concatenating the two modalities.
• Sparse regression head: A Lasso regressor on the fused embedding outperforms an MLP head, indicating that sparsity acts as useful regularization on a dataset of this size.
• Interpretable geometry knobs: The interface contact distance threshold and the number of retained persistence lifetimes are tuned by grid search, peaking at 8.0 Å and six lifetimes with a broad, flat optimum.

#Technical Details

The concatenated antibody and antigen sequence is encoded by the pretrained 3B-parameter ESM-2 model and mean-pooled to a 2,560-dimensional vector, then passed through a two-layer feedforward encoder. Each of the four structural feature families is padded or truncated to 100 dimensions at the 90th percentile and encoded by its own two-layer network with LayerNorm, GELU, and dropout; adaptive gating fuses them into a 256-dimensional topology embedding. Two bidirectional cross-attention layers with eight heads and hidden dimension 256 align the modalities, followed by gated fusion and a Lasso head. Training uses AdamW at a learning rate of 3e-4, weight decay 1e-4, batch size 32, plateau-based scheduling, and early stopping, on a single NVIDIA A100 with mixed precision.

The dataset comprises 1,705 antibody-antigen complexes from 472 unique PDB entries; 303 entries have both sequence embeddings and computed topological features, yielding 1,398 usable instances split 978/209/211 for training, validation, and test with no PDB entry shared across splits. On the test split TopoBind reaches R² 0.3390, Pearson correlation 0.5905, MSE 3.8160, MAE 1.4255, binary classification accuracy 0.7500, and AUC 0.784, ahead of an ESM-2-only MLP (R² 0.2679, AUC 0.737), a Random Forest with PCA and RFECV feature selection (R² 0.2359), and the COMBINE energy-decomposition baseline (R² 0.0058). Ablations put the adaptive fusion module at 0.3201, removal of cross-attention at 0.2716, an MLP head at 0.3251, and ESM-2 features alone with a Lasso head at 0.2896.

#Applications

Fast ΔG estimation is most useful in antibody engineering, where affinity maturation and candidate triage require ranking many designed variants before committing to expression and binding assays. A model that consumes a sequence pair plus a structure and returns a free energy in one forward pass fits into screening pipelines upstream of surface plasmon resonance, and the classification results support its use as a coarse binder-versus-weak-binder filter as well as a regressor.

#Impact

TopoBind's main evidence is methodological: persistent homology and interface geometry carry information that a protein language model does not, and cross-attention extracts more of it than naive concatenation. The scale of the demonstration is modest — 303 structurally characterized complexes, and an R² of 0.339 that leaves most of the variance in binding free energy unexplained — so the result reads as support for topological priors in molecular learning rather than a solved affinity predictor. Code is available under an MIT license, though no trained checkpoint is distributed, and the preprint has not yet been peer reviewed.