A comparator model that rank-orders molecules against a kinase

Methods

Task definition

The model answers a comparison: given one kinase sequence and two ligands, which ligand is more potent against that kinase. Ranking a compound library is a sequence of such comparisons, so this is the primitive operation of virtual screening rather than a proxy for it. Because the orientation of every pair is randomised, a constant output scores exactly 0.500 and a Matthews correlation of 0.000, so the chance line needs no argument. That is not the same as there being no baseline worth beating, and two are reported beside the model below: ranking by molecular size alone, and looking a compound's potency up from another kinase where such data exists.

Data

Measurements are drawn from the Kinase Knowledgebase (KKB), Eidogen-Sertanty, release Q2-2026, restricted to human enzyme assays with a defined potency between pIC50 3 and 11.

From the corpus to the fitted model, every step:

The training half spans 500 kinase genes and 742 distinct sequences, including 13,218 rows carrying a mutated sequence; a mutant is represented by its own sequence rather than folded into the wild type.

Molecular representation

Each ligand is described by a Morgan count fingerprint of radius 2 and 1,024 bits, computed with RDKit from the canonical SMILES, concatenated with fourteen physicochemical and compositional descriptors: molecular weight, heavy-atom count, bond count, rotatable bonds, ring count, and the counts of C, N, O, S, F, Cl, Br, I and P. The descriptors are part of the arm that was scored and are reported for exactness. Their measured contribution is negligible: paired per kinase over the 213 targets with at least 1,000 pairs, adding them is worth +0.0023 ± 0.0030 on count fingerprints. An earlier version of this document explained them by saying a circular fingerprint represents molecular size poorly; that explanation predicted a material gain on BINARY fingerprints, and the measured gain there was +0.0032 ± 0.0031, no different. The explanation is withdrawn and plain Morgan counts would serve as well.

Protein representation

Each kinase sequence is embedded with ESM2 (esm2_t12_35M_UR50D), giving one vector per residue, and reduced to a single fixed 480-dimensional vector by mean pooling over residues. Pooling is fixed rather than learned. Concatenated with the 1,038 ligand features, the vector presented to the model is 1,518 dimensions: 480 protein and 1,038 ligand.

What the model is given, and what it is not

The model is trained on human kinase sequences and ligand structures from the Kinase Knowledgebase, and on nothing else. Its training table carries 673,659 rows over 500 kinases and 742 distinct sequences, of which 13,218 rows are point-mutant sequence records. Every row is an original measurement.

Relation to prior work

The structure-free framing of this project, predicting engagement from protein sequence and ligand chemistry alone, with no three-dimensional pose, using a protein language model together with a chemical language model, follows Fondrie and colleagues, Structure-free, site-resolved contrastive learning extends small-molecule discovery beyond the reach of structure-based modeling (Talus Bioscience, bioRxiv 2026), which introduces Ptarmigan-1.

The debt is specific and worth stating precisely. This project's neural arm was adapted from that architecture: frozen ESM2 residue embeddings and frozen ChemBERTa ligand embeddings projected into a shared space, with the temperature-scaled softmax pooling over residues that Ptarmigan-1 uses to turn residue-level scores into a protein-level call reused here as attention pooling, and its leakage-safe Bemis-Murcko scaffold-split discipline. Trainable projections and head only, both backbones frozen, as a lighter-weight analogue of its LoRA fine-tuning.

The model reported here is not that model. The neural arm was beaten on every measure by the random forest described above, which uses fixed mean pooling and a count fingerprint rather than learned pooling and a chemical language model, and which is what all results in this document describe. The two also answer different questions: Ptarmigan-1 predicts and localises engagement across the proteome, including at cryptic and disordered sites; this model ranks ligands by potency against a kinase for which measured data already exists.

Model and why it is antisymmetric

A random forest regressor of 200 trees (minimum two samples per leaf) is fitted pointwise on (protein vector, ligand features) against measured pIC50. Two ligands are then compared by the difference of the model's two predictions, so that P(A more potent than B) follows the sign of s_A − s_B. This construction is exactly antisymmetric: exchanging the two ligands reverses the prediction identically, and the model cannot return contradictory answers for the same pair presented in a different order. A forest given both ligands as joint input would carry no such guarantee. It also means a library of n compounds is scored in n forward passes rather than n² comparisons.

Evaluation pairs

This model has no training pairs. The forest is fitted pointwise on individual pIC50 values; pairs exist only at evaluation, and every non-tied pair of ligands measured against the same sequence is scored. There is no gap filter on pair construction.

Results are then reported in three bands of true potency difference: under half a log, half to one log, and more than one log. The band most worth quoting is the last, because that is the separation a chemist acts on. For context on what a resolvable difference is, disagreement between two independent publications reporting the same compound against the same kinase has a 90th percentile of 0.715 log units over 204,438 comparisons, rising to 1.218 after deduplication; the source file recommends treating one log as the floor. Pairs closer than that are harder rather than impossible: the model still scores 57.8 percent under half a log, against a 50 percent coin flip.

Held-out evaluation

The advantage that cross-kinase knowledge confers is not assumed away, it is measured. The compound-lookup baseline above is exactly that advantage made explicit, predicting from the same compound’s potency against another kinase, and it scores 73.1 percent. The model scores 76.8 percent on those same pairs, so it adds information beyond knowing the compound elsewhere. What these internal numbers do not measure is performance on chemistry absent from the corpus entirely; the external section speaks to that.

Ligands are withheld by Bemis-Murcko scaffold, keyed on the (sequence, ligand) pair. Every kinase sequence remains in training. No protein and no sequence family is withheld, because a kinase for which no data exists is not the deployment scenario: the kinome is among the most exhaustively characterised regions of the proteome. Evaluation covers 2,965,273 held-out ligand pairs across 413 targets. The kinase is the unit of replication for the kinase-level mean and its standard error, and only for those. Pooled accuracy, the margin deciles and the selectivity figures are computed over pairs that share compounds and are therefore not independent observations; treat them as descriptive rather than as quantities with an interval.

What this model does not use

The headline results are obtained without three bodies of work that are built and available: 4,555 CANDIDATE orthologue sequences across 475 organisms, screened by binding-site identity; per-residue contact maps from 1,417 solved co-complex structures covering 170 targets, with 977 identified ligands and 182 sites excluded because a nucleotide or crystallisation additive rather than an inhibitor defined them; and 821,370 BindingDB measurements retained after removing every compound-target pair appearing in the frozen ChEMBL benchmarks. These are headroom, not caveats.

Data sources

Kinase Knowledgebase (KKB), Eidogen-Sertanty, release Q2-2026, the training and evaluation corpus: eidogen-sertanty.com/kinasekbmarvin.php.

BindingDB, ingested and held for future arms, not used in the results reported here. Gilson et al., Nucleic Acids Res. 44:D1045 (2016), doi:10.1093/nar/gkv1072; bindingdb.org.

ChEMBL, used only as a frozen external benchmark, never trained on; every compound-target pair appearing in it was removed from the BindingDB import. Zdrazil et al., Nucleic Acids Res. 52:D1180 (2024), doi:10.1093/nar/gkad1004; ebi.ac.uk/chembl.

RCSB Protein Data Bank, co-complex structures used for the binding-site analysis reported separately. Berman et al., Nucleic Acids Res. 28:235 (2000), doi:10.1093/nar/28.1.235; rcsb.org.

UniProt, sequence and accession mapping. UniProt Consortium, Nucleic Acids Res. 51:D523 (2023), doi:10.1093/nar/gkac1052; uniprot.org.

Models and software

ESM2 protein language model, checkpoint facebook/esm2_t12_35M_UR50D, 35M parameters, 12 layers, 480-dimensional per-residue representations. Lin et al., Science 379:1123 (2023), doi:10.1126/science.ade2574; code at github.com/facebookresearch/esm.

ChemBERTa chemical language model, checkpoint seyonec/ChemBERTa-zinc-base-v1, 768-dimensional embeddings, used in the comparison arm. Chithrananda et al., arXiv:2010.09885 (2020), arxiv.org/abs/2010.09885.

Morgan / ECFP count fingerprints, radius 2, 1,024 bits, and all physicochemical descriptors computed with RDKit: rdkit.org. Method: Rogers & Hahn, J. Chem. Inf. Model. 50:742 (2010), doi:10.1021/ci100050t.

Bemis-Murcko scaffolds, used to define the held-out ligand split. Bemis & Murcko, J. Med. Chem. 39:2887 (1996), doi:10.1021/jm9602928.

Random forest regressor, scikit-learn: scikit-learn.org. Breiman, Machine Learning 45:5 (2001), doi:10.1023/A:1010933404324.

Pairwise ranking formulation: Burges et al., Learning to Rank using Gradient Descent, ICML 2005, doi:10.1145/1102351.1102363.

Matthews correlation coefficient: Matthews, Biochim. Biophys. Acta 405:442 (1975), doi:10.1016/0005-2795(75)90109-9.