Consolidated Dataset from the RNS Study: Protein Embedding RNS Scores, Jensen–Shannon Divergence, and Sequence Alignment Matches

Embeddings produced by language models (LMs) are widely used as numerical representations of natural language sentences and structured data. However, using embeddings without accounting for model confidence is a critical limitation. The <b>Random Neighbor Score (RNS)</b> provides a model- and task-agnostic measure of embedding uncertainty.<b>Associated Preprint</b>: https://www.biorxiv.org/content/10.1101/2025.04.30.651545v1<b>Files included</b>:<b>STable_consolidated_RNSscores_Rev1_v0.tsv.gz</b>: Consolidated sheet containing RNS scores, sequences, alignment results, and Jensen–Shannon divergence values.Columns labeled <code>RNS_IS_*</code> and <code>RNS_ISb_*</code> correspond to RNS values computed at different <i>k</i> settings using <b>Astral40R</b> and <b>Proteome4R</b> as random sets, respectively.<b>Astral40.fasta</b>: Sequences of selected Astral40 domains (<b>Astral40</b>).<b>Astral40_Rshuffled.fasta</b>: Sequences of "synthetic" / "random" set (<b>Astral40R</b>) - replicates AA composition of Astral40.<b>STable_Perf_esm2_t36_3B_UR50D_Astral40_Rev1_v0.tsv </b>: Embedding's RNS and contact prediction accuracy of ESM for Astral40 domains.<b>STable_Perf_esm2_t36_3B_UR50D_PDB23to24_Rev1_v0.tsv</b><b> </b>: Embedding's RNS and contact prediction accuracy of ESM for PDB23to24 structures.<b>STable_Perf_prot_t5_xl_u50_Astral40_Rev1_v0.tsv</b> : Embedding's RNS and Sec . Str. prediction accuracy of ProtT5 for Astral40 domains.<b>STable_Perf_prot_t5_xl_u50_PDB23to24_Rev1_v0.tsv</b><b> </b>: Embedding's RNS and Sec . Str. prediction accuracy of ProtT5 for PDB23to24 structures.<b>Sequence sources</b> (see manuscript for details):ASTRAL 40: https://scop.berkeley.edu/astral/Novel Meta set / Orphan set: https://10.0.23.196/m9.figshare.c.6737127Novel Hallucination set: https://www.nature.com/articles/s41586-021-04184-w#data-availability → https://files.ipd.uw.edu/pub/trRosetta/hallucinations2K.tar.gzIntrinsically Disordered Proteins (IDP) &amp; Intrinsically Disordered Regions (IDR): https://disprot.org/download (version 2024_12)<b>Related repos:</b> Sample sequence embeddings: https://doi.org/10.6084/m9.figshare.30179413.v1Code to compute RNS scores: https://bitbucket.org/bromberglab/rns/src/main/<br><br>