CASTER-DTA

CASTER-DTA (Cross-Attention with Structural Target Equivariant Representations for Drug-Target Affinity) predicts the binding affinity between a protein and a small molecule directly from the protein's three-dimensional structure. Most drug-target affinity (DTA) methods represent the protein as a one-dimensional amino acid sequence or as a contact map derived from that sequence, discarding the geometry that actually determines binding. CASTER-DTA instead encodes the protein as a geometric graph and processes it with SE(3)-equivariant layers, so the learned representation respects the rotational and translational symmetries of real molecular structure.

The model was developed by Rachit Kumar, Joseph D. Romano, and Marylyn D. Ritchie at the University of Pennsylvania, first posted as a preprint in November 2024 and published in Briefings in Bioinformatics in 2025. Its design premise is that structure is now cheap: with AlphaFold 2 and its associated structure database, a predicted backbone is available for essentially any protein of interest, and affinity models should exploit that rather than fall back on sequence. The authors pair the structural protein encoder with a conventional molecular graph network and join the two through a cross-attention layer between protein residues and drug atoms.

That cross-attention layer does double duty. It conditions each modality's representation on the other before pooling, and it exposes per-residue attention weights that can be inspected after the fact, turning an otherwise opaque regression into something a medicinal chemist can interrogate.

#Key Features

• Equivariant structural protein encoder: Geometric vector perceptron (GVP) graph convolutions operate on scalar and vector features of the apo backbone, preserving SE(3) symmetry. Ablations show equivariant layers outperform non-equivariant convolutions on the same graphs.
• Residue-atom cross-attention: Attention weights between protein residues and drug atoms are extracted at inference and correspond to ligand-binding residues annotated in BioLIP, providing structural interpretability alongside the scalar affinity prediction.
• No external pretrained embeddings: Protein features are one-hot amino acid identity plus geometric descriptors rather than protein language model embeddings, keeping the trained model to roughly 764,000 parameters.
• Proteome-scale prediction resource: The authors released 37,776,465 predicted affinities covering 1,615 FDA-approved drugs against 23,391 human protein structures spanning 20,504 unique proteins.
• Hosted inference from a fixed checkpoint: A web server applies the pretrained CASTER-DTA(2,2) model to arbitrary protein structure and SMILES pairs, returning both an affinity estimate and attention visualizations.

#Technical Details

Protein graphs are built from ligand-free structures, sourced first from experimentally determined PDB entries at 100% sequence identity, then from the AlphaFold structure database, and finally by folding remaining sequences locally with ColabFold. Residues are connected to every other residue within 4 Å, including self-loops — a departure from the 30-nearest-neighbor convention used in earlier GVP-GNN work. Molecules are converted from SMILES to 2D graphs with RDKit and processed by GINEConv layers. After cross-attention, both graphs are mean-pooled, concatenated, and passed through fully connected layers for regression. The headline configuration, CASTER-DTA(2,2), uses two protein and two molecule convolution blocks for 764,396 parameters.

Training and evaluation span four DTA benchmarks: Davis (30,056 pairs), KIBA (118,254 pairs), Metz (36,136 pairs), and a filtered BindingDB Kd set (47,569 pairs). Against DeepDTA, GraphDTA, DGraphDTA, DeepGLSTM, and AttentionDTA, CASTER-DTA reaches MSE 0.211 on Davis, 0.282 on Metz, and 0.717 on BindingDB, improving on the best baseline in each case, with Pearson correlations of 0.859, 0.832, and 0.856 respectively. On KIBA it is essentially tied with DGraphDTA (MSE 0.143 versus 0.141).

#Applications

The immediate use is structure-informed virtual screening and drug repurposing: the released proteome-wide matrix lets researchers query which approved drugs are predicted to bind a target of interest without running any inference themselves, and the web server covers one-off protein-drug pairs. Beyond ranking, the attention output supports binding-site hypothesis generation, and the authors demonstrate downstream analyses including assessment of how PharmGKB-catalogued missense variants shift predicted affinity — relevant to pharmacogenomics, where a variant's effect on drug binding is often the mechanism of interest.

#Impact

CASTER-DTA offers a controlled demonstration that SE(3)-equivariance improves affinity prediction over otherwise comparable non-equivariant graph networks, and it does so at a parameter count orders of magnitude below sequence models that lean on large pretrained encoders. Its limitations are stated plainly by the authors: predictions inherit AlphaFold 2's weaknesses on disordered regions and alternative conformations, apo structures cannot capture induced fit, large proteins are fragmented for tractability, and affinity alone says nothing about bioavailability or off-target liability. The code and weights are distributed under a University of Pennsylvania academic license permitting non-commercial research use only, with commercial use requiring permission from the Penn Center for Innovation.