Layer-Selected ESM-2 Embeddings for Kinase Classification

Protein kinase sequences, domain-level ESM-2 embeddings with mid-layer averaging (layers 20-33), homology-aware train/test splits at 40%, 50%, and 70% sequence identity, trained logistic regression models with Platt calibration, and complete experimental results for kinase functional family classification. This dataset supports the finding that mid-layer averaging (layers 20-33) in ESM-2 outperforms final layer embeddings by +32% ARI for functional classification. The dataset includes 1,255 kinase domain sequences across 8 major families (AGC, Atypical, CAMK, CK1, CMGC, STE, TK, TKL), along with ESM-2 embeddings (1280-dimensional), homology-aware splits, trained models, clustering results, supervised classification results, and publication-quality figures.