ProtT5-derived data for the Human Proteome generated with the VESPA pipeline

This dataset contains various ProtT5-derived outputs for the human proteome, with all sequences truncated to a maximum length of 1024 amino acids. The data was generated using a combination of the original VESPA framework and a custom pipeline developed for a bachelor's thesis at the Technical University of Munich's Rostlab. All outputs, except for the SAV predictions, are stored as PyTorch tensors in HDF5 format, while SAV predictions are provided as multiple CSV files. Included Folders: fasta_sequences: Raw input sequences without truncation. UniProd identifier is not available and would need to be mapped using sequences. embeddings: ProtT5 embeddings generated using the original VESPA framework. logits: Raw per-residue ProtT5 logits. reconstruction_probabilities: Per-residue reconstruction probabilities derived from the logits via softmax. log_likelihoods: Per-residue, per-mutation log-odds ratios in the shape (L × 20). These were computed by normalizing log-probabilities by the log-probability of the wild-type residue. conservation_scores: 9-state per-residue conservation predictions. sav_predictions: Final VESPA mutation effect predictions for all single-residue substitutions, based on sequences with applied truncation, conservation scores, and log-likelihoods. Generation Details: The logits, reconstruction probabilities, and log-likelihoods were generated using a custom pipeline developed for this thesis. This pipeline utilized Dockerized runs on the LRZ cluster, processed across multiple GPUs using SLURM job arrays (see GitHub repository: https://github.com/Salzemann/fast_loglikelihoods). The other outputs (embeddings, conservation scores, SAV predictions) were generated using the original VESPA framework (https://github.com/Rostlab/VESPA) on a local machine with an RTX 3060 12GB VRAM GPU and an AMD Ryzen 9 7950X3D 16-Core CPU.