Loading…
Protein language model scoring whether a kinase missense variant alters catalytic activity, fine-tuned from ProtBert on 2,553 curated alterations.
No providers recorded yet. Browse all providers
A missense change in a protein kinase can leave the enzyme untouched, blunt it, switch it on constitutively, or kill catalysis outright — and general-purpose pathogenicity predictors report all four the same way, as "damaging." That distinction is the clinically actionable part: EGFR L858R is a drug target because it activates the kinase, while a kinase-dead substitution at the same gene means something entirely different for a patient. Worse, the residues that control activity are not confined to the textbook elements — the glycine-rich loop, the gatekeeper, the catalytic and activation loops. Many experimentally confirmed activity-altering mutations sit outside them, which is exactly where a position-based rule has nothing to say.
pKAKA attacks that gap by treating kinase activity as a property readable from sequence context. The group of Ze-Xian Liu at Sun Yat-sen University Cancer Center, with collaborators at Zhengzhou University, City University of Hong Kong and South China University of Technology, first curated 2,553 experimentally validated kinase activity-related key alterations (KAKAs) from the literature, then fine-tuned the pretrained protein language model ProtBert — the BERT encoder from the ProtTrans suite, trained on UniRef sequences with no alignment input — on that labelled set. The result, published in the Journal of Genetics and Genomics in 2026, is a single scorer that assigns any human kinase missense substitution a probability that it perturbs catalytic activity.
pKAKA is a fine-tuned derivative rather than a model pretrained from scratch: the ProtBert transformer encoder is the frozen starting point and the KAKA labels supervise the activity-effect head. On held-out testing it reaches an AUC of 0.9593 and outperforms AlphaMissense, the general missense pathogenicity predictor, in the authors' head-to-head comparison — a result worth reading as evidence that kinase activity is a distinct target from pathogenicity, not that one model supersedes the other. Output is a single score scaled between 0 and 1, the categorical probability that the substitution affects activity, rather than a directional call on whether activity rises or falls. pKAKA is used through the KAKA server, which serves the precomputed kinome-wide scores by gene name, UniProt accession or mutation site; the server's dataset download page is not yet open.
The immediate audience is anyone staring at a kinase missense variant of uncertain significance. In tumour sequencing, a panel returns kinase substitutions faster than any lab can assay them, and a score that separates activity-altering from inert changes narrows which ones justify a biochemical follow-up or an inhibitor-sensitivity experiment. The same applies to rare-disease exomes, where kinase genes such as LRRK2 recur, and to mutagenesis design, where picking a substitution expected to abolish catalysis is a routine control.
pKAKA's more durable contribution may be the curated KAKA set itself: a few thousand experimentally grounded, directionally labelled kinase alterations is a scarce commodity, and it gives the field a benchmark for the activity-effect question that general variant benchmarks do not cover. The model demonstrates that a released protein language model plus a small expert-curated corpus is enough to reach usable accuracy on a specialised functional question — the pattern that increasingly replaces bespoke feature engineering in variant interpretation. The practical limits are that the public output collapses the four curated classes into one effect probability, that evaluation is computational rather than newly validated at the bench, and that access runs through a web server rather than code or weights a user can run locally.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.