ProFam: Open-Source Protein Family Language Modeling for Fitness Prediction and Design - The ProFam Atlas Dataset

ProFam-Atlas Dataset The ProFam-Atlas is a curated, large-scale dataset of protein sequences grouped by varying definitions of "protein family". It was developed to train the ProFam-1 protein family language model (pfLM) and supports research in protein design, fitness prediction, and evolutionary biology. Full details of dataset creation can be found in: ProFam: Open-Source Protein Family Language Modeling for Fitness Prediction and Design (Wells et al., 2025.) Dataset Composition The dataset aggregates protein family definitions from multiple sources, ranging from structure-centric clusters to deep multiple sequence alignments (MSAs) and functional groupings. ### 1. FoldSeek AlphaFold Database Clusters - **Source:** FoldSeek AlphaFold Database (AFDB) clusters. - **Definition:** Structure-centric families formed by first clustering sequences (50% identity, 90% coverage) and then clustering structures (FoldSeek E-value ≤ 0.01). - **Statistics:** Approximately **2.3 million families** comprising **30 million sequences**. ### 2. OpenProteinSet MSAs (Re-clustered) - **Source:** OpenFold OpenProteinSet (subset of 270k maximally diverse MSAs). - **Definition:** Re-processed to ensure length consistency and cleaner alignments. Aligned sequences were split at contiguous gaps (>10 residues), filtered for length (≥90 residues), and re-clustered using MMseqs2 (30% sequence identity, 70% coverage). - **Statistics:** **37 million families** comprising **246 million sequences**. ### 3. TED FunFams - **Source:** The Encyclopedia of Domains (TED) and CATH Functional Families (FunFams). - **Definition:** Continuous domains from TED mapped to CATH FunFam HMM profiles. Matches were redundancy-reduced (50% identity) and further clustered (30% identity, 80% overlap) to ensure sequence and length consistency. - **Statistics:** **38,000 families** comprising **765,000 sequences**. ### 4. UniRef90 Single Sequences - **Source:** UniRef90 (release 2025_01). - **Definition:** Single sequences included to cover regions of the protein universe not captured by the family-based groupings. - **Statistics:** **205 million single sequences**. File Structure and Format The dataset is distributed as a collection of paired `.mapping` and `.sequences` files. This structure allows for efficient random access to specific families without loading entire files into memory. Mapping Files (`*.mapping`) Contains the index information for each family. - **Format:** - Line 1: Family Identifier (prefixed with `>`). - Line 2: The name of the corresponding `.sequences` file, followed by a colon `:`, and a comma-separated list of byte offsets (or entry indices) for the sequences belonging to that family. **Example:** ```text >1.25.40.900_ted.fasta.0 train_001.sequences:0,2,4,5,6,7,8... ``` Sequence Files (`*.sequences`) Contains the actual protein sequences in a FASTA-like format. - **Format:** - Header line starting with `>` containing the sequence identifier (and potentially metadata). - Sequence line containing the amino acid sequence. **Example:** ```text >AF-A0A1H4DYA3-F1-model_v4_TED03/0.0 EWAKAETAANAVINSGIYTLNKDLNLRLAEQYLIRAEARAQLGNLAGAVADVDSIRSKAGLPQLDNSITQPALLLAIEKERKAELFGEWGHRWFDLKRTPAVAGGGKTRAD >AF-A0A1W1ZJU1-F1-model_v4_TED01/1.1 LLMQAEAENEVNGPTQVAYNAVNEVRHRAGLPDLTPGLAKEAFFNALVDERAHELCFEGFRKWDLIRWNMLGAKIRATQTALKAYRANFPYVAGDNF ``` Citation If you use the ProFam-Atlas dataset in your research, please cite the following preprint: ProFam: Open-Source Protein Family Language Modeling for Fitness Prediction and Design Wells et al., 2025.