T-ALPHA

Scoring functions that predict how tightly a small molecule binds a protein are the workhorse of hit identification and lead optimization, but most of them read a single view of the complex — a voxel grid, a molecular graph, or a pair of sequences — and discard the rest. T-ALPHA, from Gregory Kyro, Anthony Smaldone, Yu Shee, Chuzhi Xu and Victor Batista in the Department of Chemistry at Yale University, instead processes a protein-ligand complex through three parallel channels and fuses them in a hierarchical transformer, on the argument that structural connectivity, surface topography, evolutionary context and chemical semantics each carry affinity signal the others miss.

The second half of the paper addresses a practical gap. Affinity models are usually benchmarked across many diverse targets, but a medicinal chemistry campaign cares about one protein. T-ALPHA ships an optional uncertainty-aware self-learning procedure that specializes the released checkpoint to a chosen target using Monte Carlo dropout over unlabeled compounds — no new experimental measurements required. The base checkpoint is fixed and carries all of the headline benchmark results; the self-learning arm is an add-on, not a per-dataset refit.

T-ALPHA sits alongside other cataloged affinity predictors, including the sequence-first HydrAffinity and AQAffinity and the structure-based GatorAffinity, distinguished by how far it pushes multimodal fusion and by its evaluation on predicted structures.

#Key Features

• Three-channel hierarchical fusion: Protein, ligand, and protein-ligand complex are each encoded independently, integrated by cross-attention within each channel, then combined by a meta transformer into a single representation.
• Surface geometry as a first-class input: Following dMaSIF, the protein is converted to an oriented surface point cloud with normals and processed by quasi-geodesic convolutions, capturing pocket curvature and topography that a bond graph cannot express.
• Works on predicted complexes: When the test-set structures are regenerated with Chai-1 instead of using crystallography, T-ALPHA still outperforms every literature model evaluated on crystal structures.
• Target-specific self-alignment without new data: Monte Carlo dropout produces uncertainty-weighted pseudo-labels on unlabeled compounds, which are used for semi-supervised fine-tuning against a single protein of interest.
• Inequality-aware loss: Binding constants reported as ">" or "<" bounds are handled by a custom dynamic-range-aware loss rather than being treated as exact values, as is standard elsewhere.

#Technical Details

The protein channel combines an E(n) equivariant graph neural network over the pocket atom graph, a dMaSIF-style quasi-geodesic surface convolution over 512 sampled surface points, and a 2,560-dimensional ESM-2 sequence embedding. The ligand channel pairs an E(n) EGNN over the molecular graph with 209 RDKit physicochemical descriptors and a 768-dimensional embedding from a 10-block SMILES transformer encoder that the authors pretrained by masked-token prediction on 4,778,512 canonical SMILES assembled from ChEMBL 33, GuacaMol, MOSES, BindingDB and PDBbind. A third EGNN operates on a merged complex graph whose intermolecular edges connect protein and ligand heavy atoms within 4.5 Å.

Training uses PDBbind v2020, whose general set holds 19,443 complexes. On the CASF-2016 core set (290 entries) T-ALPHA reaches RMSE 1.112, MAE 0.875, r² 0.738, Pearson r 0.869 and Spearman ρ 0.860, leading every model reported in the literature; with Chai-1 structures it records RMSE 1.134 and Pearson r 0.857. On the leakage-controlled BDB2020+ set it reaches RMSE 0.969 against 1.01 for IGN and 1.54 for AutoDock Vina, and it leads the LP-PDBbind split. On protein-specific sets it attains Spearman ρ 0.737 for the SARS-CoV-2 main protease and 0.791 for EGFR; self-learning raises Mpro ρ by a further 9.91% and EGFR ρ by 3.41%. Ablations show the complex-level EGNN is the single most important component, its removal degrading CASF-2016 RMSE from 1.112 to 1.247.

#Applications

The intended use is ranking candidate inhibitors during hit identification and lead optimization, especially where no co-crystal structure is available and the complex must come from a structure predictor. Trained parameters and test files are archived on Zenodo, the repository ships scripts/inference.py, and an on-the-fly Colab notebook accepts a protein and a ligand directly, so a medicinal chemist can score compounds without a local GPU pipeline. The self-learning script offers teams working on one target a way to sharpen ranking using only compounds they already have in hand.

#Impact

T-ALPHA is peer-reviewed in the Journal of Chemical Information and Modeling and its code is MIT-licensed, with weights deposited on Zenodo rather than a model hub. The strongest result is not the CASF-2016 leaderboard position — a benchmark the authors themselves note is inflated by protein and ligand overlap with PDBbind — but the pairing of leakage-controlled BDB2020+ and LP-PDBbind gains with the demonstration that predicted structures cost little accuracy. That combination is what makes the model usable outside the curated benchmark setting. The parameter count is not reported, and comparisons draw on literature-reported baselines rather than re-run head-to-head evaluations.