KolossuS

KolossuS predicts which kinase phosphorylates which site, from protein sequence alone. It was developed in the Rohit Singh and Scott Soderling labs at Duke University School of Medicine, with collaborators at the International Institute for Integrative Sleep Medicine at the University of Tsukuba, and posted to bioRxiv in April 2025. The problem it addresses is stark: roughly 2% of human genes encode kinases, yet 80% of kinases have fewer than 20 known substrates, and fewer than 5% of catalogued human phosphopeptides have an assigned kinase.

Rather than train a sequence model from scratch, KolossuS is a co-embedding head over a frozen protein language model. Full kinase sequences and 15-residue substrate windows centred on the phosphoacceptor are embedded with ESM-2, average-pooled, and projected by separate multilayer perceptrons into a shared latent space where the cosine similarity between a kinase and a site is the predicted probability of phosphorylation. Positioning both partners in one space makes the model interpretable — kinases that cluster together are predicted to prefer similar substrates — and, crucially, calibrated across kinases rather than only within them, which is what allows an observed phosphoproteome to be matched to its cognate kinases.

The paper pairs the model with CRISPR-based in vivo proximity labelling, using KolossuS to interpret kinase-proximal phosphoproteomic data. Applied to sleep-deprived mice, the combined framework recovered known and previously unreported substrates of the kinase Sik3.

#Key Features

• Sequence-only inference: Predictions need nothing but amino acid sequences, so the model applies directly to phosphoproteomics output, which reports peptide fragments rather than whole proteins or structures.
• Cross-kinase calibration: Family-specific F1-maximizing thresholds fall between 0.15 and 0.37, against near-zero for the closest published comparator, meaning raw scores are comparable across the kinome.
• Species transfer: Trained on human data, the model retains comparable accuracy on mouse and rat phosphorylation events without retraining.
• Interpretable kinase embeddings: The learned kinase space separates the CMGC family from other Ser/Thr kinases — a split absent from raw ESM-2 embeddings — and shows greater within-family dispersion for the AGC and STE families.
• Two backbone sizes: Weights are released for both the ESM-2 15B and ESM-2 650M variants, the smaller one for GPU-constrained settings.

#Technical Details

Training proceeded in two stages over held-out kinase families rather than sequence-identity splits. The model was first trained on atlas-scale position-specific scoring matrices covering 393 human kinases from synthetic peptide array experiments, then fine-tuned on LC/MS in vitro assays of cellular proteins spanning 354 human kinases. Final benchmarking used literature-curated human, mouse, and rat interactions from PhosphoSitePlus, withheld from both training stages; the CK1 and JAK families were reserved for validation and TKL and FAK held out entirely for testing. On human kinases, KolossuS reaches AUROC 75.5% for Ser/Thr and 83.8% for Tyr kinases. The two-stage recipe matters: zero-shot average-pooled ESM-2 embeddings score AUROC 50.3% and AUPR 42.5%, atlas training alone lifts these to 75.4% and 66.7%, and fine-tuning reaches 76.2% and 67.9%.

Against Phosformer-ST, the strongest prior kinase-specificity method, KolossuS is more accurate and substantially more sensitive at a fixed 0.5 threshold, and its advantage widens on Tyr kinases, which Phosformer-ST was not trained on. Ablations found the ESM-2 15B backbone best, with 650M and the newer ESM C 6B both slightly behind. The released model was used to score 466 human kinases against 227,736 substrate sequences and 458 mouse kinases against 100,981 sequences.

#Applications

KolossuS is aimed at researchers who have phosphoproteomic data and no kinase assignment for most of it, including the "dark kinase" problem of enzymes with no annotated substrates. The kolossus package on PyPI provides kolossus-cli and kolossus-extract, which take a FASTA file and a list of kinase-substrate-site triples and return calibrated probabilities, so proteome-wide scoring runs without retraining. A web interface hosted by the Singh lab serves the same predictions. The paper's own use case — interpreting proximity-labelled phosphoproteomes from live mice — shows the intended pairing with spatial experimental data, where sequence specificity and subcellular co-localization are treated as separate, complementary determinants.

#Impact

KolossuS is deliberately narrow: a small projection head on frozen embeddings, doing one task. Its contribution is calibration and coverage rather than architectural novelty — kinome-wide predictions that can be compared across kinases, extending to Tyr kinases and to mouse and rat where prior methods were confined to human Ser/Thr enzymes. Distribution is unusually complete for a preprint-stage tool: checkpoints for both backbones live in the GitHub repository via Git LFS and on the Hugging Face Hub under an MIT license, with a live PyPI package and hosted web predictions. The GitHub repository itself carries no LICENSE file, the preprint has not been peer reviewed, and the accuracy figures are the authors' own and the margins over Phosformer-ST have not been independently reproduced.