ProtCC Latent Browser: data release for a sparse crosscoder over the ProtT5 encoder

Precomputed data for the ProtCC Latent Browser. The browser is a static web dashboard over a sparse crosscoder that reads the residual stream of all 24 encoder layers of ProtT5 (Rostlab/prot_t5_xl_uniref50) at once. The code that builds and serves the dashboard is at github.com/SohrabTa/protcc. Its folder dashboard/deploy downloads this record and serves the site with docker compose. Contents. 33 gzipped tar archives unpack into one data tree of 423,258 files. The file data-release.json gives the sha256, the size and the file count of each archive. indexes.tar.gz: the index files that the site reads first (concepts, latents, depth profiles, locality, protein lookup) and 208 protein bundles. feature.tar.gz: one ranking file for each of the 8,128 live latents. A ranking holds every protein that the latent fires on, with the peak activation and the share of the protein that the latent covers. tracks-00.tar.gz to tracks-18.tar.gz: the per-residue activations of all latents, one file for each of the 207,463 proteins, in groups of two-character subfolders. structures-00.tar.gz to structures-11.tar.gz: an AlphaFold model for 207,459 of the proteins, as gzipped mmCIF, grouped in the same way. Model and proteins. The crosscoder is crosscoder_l8192_k32_bs512_full_uniref50 (8,192 latents, k = 32, JumpReLU at inference). The proteins are the reviewed UniProtKB/Swiss-Prot entries with UniProt annotation score 3 to 5, an AlphaFold model, and a length of at most 512 residues. Sources and changes. The concept annotations come from UniProtKB/Swiss-Prot (The UniProt Consortium, CC BY 4.0). The structures come from the AlphaFold Protein Structure Database (AlphaFold Data Copyright (2021) DeepMind Technologies Limited, CC BY 4.0; Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature (2021)). We changed the structures as follows: most were decoded from the Foldcomp database afdb_swissprot (Kim, Mirdita and Steinegger, Bioinformatics 2023), which stores the models with a lossy torsion-angle encoding, and 5,353 were downloaded from EBI. All of them were reduced to the four backbone atoms N, CA, C and O of each residue. The B-factor column keeps the pLDDT. The predictions have varying levels of confidence.