Kinhibit

Protein kinases are among the most heavily pursued drug targets in oncology and inflammatory disease, and predicting how tightly a candidate inhibitor binds a given kinase decides which molecules reach an assay. Measuring dissociation constants experimentally is slow, and the datasets that do exist are small and heavily skewed toward high-affinity compounds — conditions under which regression models trained from scratch become unstable.

Kinhibit, from Xudong Guo, Zixu Ran and Fuyi Li at Northwest A&F University and the University of Adelaide, addresses this by refusing to learn representations from the affinity data at all. Both halves of the problem are handled by frozen pretrained encoders: ligands by an equivariant graph neural network pretrained with self-supervised contrastive learning over multiple conformers of the same molecule, and kinases by ESM-S, a structure-informed protein language model derived from the ESM-2 family through remote homology detection. Only the projection layers and prediction head are trained. The framework was posted as a preprint in February 2025 and published in Briefings in Bioinformatics in July 2025, benchmarked chiefly against GPT4Kinase, a general-purpose large language model applied to the same MAPK datasets.

#Key Features

• Multiview conformer contrastive pretraining: Each ligand is expanded into ten molecular graphs with different atomic coordinates and conformations; graphs of the same molecule are positive pairs and graphs of other batch molecules are negatives, so the encoder learns conformation-invariant chemical structure without labels.
• Structure-informed protein encoding: Kinases are represented by 1280-dimensional ESM-S embeddings carrying structural signal from remote homology detection; ablations substituting ESM-2 or ProtT5 degrade every metric.
• Explicit knowledge fusion: Projection layers with batch normalization map ligand and kinase embeddings into a shared 128-dimensional space before concatenation, rather than fusing raw features of mismatched dimensionality.
• Classification framing for skewed data: Affinity is predicted as low, medium or high under two published Kd thresholding schemes, sidestepping the instability that direct regression shows on a distribution this imbalanced.
• Bond-level interpretability: Attention over edge features ranks the chemical bonds that most influence a prediction, and the highest-affinity inhibitors concentrate importance on rigid ring systems.

#Technical Details

The molecular encoder is a single-layer E(n) Equivariant Graph Neural Network with input, hidden and output dimensions of 133; nodes carry eight atom features and edges four bond features derived through RDKit. Kinase embeddings enter the projection layer at 1280 dimensions, ligand embeddings at 133, and both are projected to 128; the concatenated 256-dimensional representation feeds a three-class fully connected classifier trained with cross-entropy and Adam at a learning rate of 0.001, with both encoders frozen. Training data is the MAPK set curated for GPT4Kinase from BindingDB, comprising 487 inhibitors across three kinase groups — 161 for RAF, 131 for MEK and 195 for ERK — split 8:2 with no SMILES shared between train and test. Published classification accuracies under the first labelling strategy are 0.912 on RAF, 0.895 on ERK, 0.889 on MEK and 0.808 on the combined MAPK-All set, with a best AUC of 0.956 on RAF. Downstream regression models trained on the learned embeddings and converted back to affinity classes reach 0.963 on MEK, 0.950 on MAPK-All and 0.912 on RAF, ahead of GPT4Kinase, BatchDTA, KIPP, DeepDTA, GraphDTA and AutoDock. The abstract's headline figures of 92.6% and 92.9% correspond to the second labelling strategy's best per-dataset regression results rather than to the classification table. Five-fold cross-validation over k-means clusters of the learned embeddings held accuracy above 0.84 in every fold, and ablations show removing contrastive pretraining drops MAPK-All accuracy from 0.929 to 0.909.

#Applications

Kinhibit is built for virtual screening in kinase-targeted drug discovery, taking SMILES strings and a kinase sequence and returning a ranked affinity assessment before any compound is ordered. The MAPK focus — RAF, MEK and ERK — puts it in the path of oncology programs targeting that cascade, and cross-family experiments extend the evaluation to EGFR, CDK2 and FLT3. A public web server accepts batches of SMILES, with a standalone version for higher-throughput local screening. Because the fusion architecture makes no assumptions specific to kinases, the same recipe applies to other protein-ligand affinity tasks.

#Impact

Kinhibit demonstrates that in low-data protein-ligand settings, the productive move is to pretrain both encoders separately and train only a small fusion head — the ablations quantify what each pretrained component contributes. The t-SNE analysis makes the same point visually: neither molecular graph embeddings nor fingerprints separate affinity levels on their own, and separation emerges only after kinase information is fused in. The limitations are acknowledged by the authors. Evaluation covers few kinase families and 487 inhibitors, cross-family transfer degrades under domain shift, and the thresholds defining the three affinity classes leave residual confusion between adjacent classes. Model weights are archived on Zenodo and the code is public, though the repository ships no license file.