ProteinGym TabPFN3-ESMC600M Supervised Substitutions Feature Matrices

This record contains derived ESMC600M feature matrices used to reproduce the TabPFN3-ESMC600M supervised ProteinGym DMS substitutions baseline. The archive contains float16 memmap feature matrices for 217 ProteinGym DMS substitution assays. The features were generated from ESMC600M-derived protein representations and associated mutation descriptors, then used as tabular inputs to TabPFN3 for the three official ProteinGym supervised substitution split schemes: fold_random_5, fold_modulo_5, and fold_contiguous_5. The archive contains derived feature matrices and row metadata only. It does not contain raw ProteinGym datasets, raw DMS labels, TabPFN3 checkpoints, or ESMC model weights. Reproduction requires the official ProteinGym supervised DMS substitution CSV files and authorized access to the TabPFN3 model weights through the official Prior Labs / TabPFN license flow. Associated code and ProteinGym PR materials are provided separately in the ProteinGym pull request and issue for the TabPFN3-ESMC600M baseline.