LIVIA Atlas: AlphaFold-Multimer protein interaction screens resolved to residues
Abstract
Data behind LIVIA Atlas: AlphaFold-Multimer protein interaction screens scored with iLIS (AFM-LIS), keeping the interface and contact residues of every prediction. livia-atlas-human-proteome.zip: human proteome-wide screen. Schmid et al. (2025), doi:10.1101/2025.11.10.687652. livia-atlas-human-kinase-tf.zip: human kinase–transcription factor screen. Kim et al. (2025), doi:10.1101/2025.10.10.681672. livia-atlas-flypredictome.zip: FlyPredictome, Drosophila melanogaster. Kim et al. (2026), doi:10.64898/2026.04.14.718529. Keyed by FlyBase gene; identity.tsv maps every construct name to its gene, and sets.json marks subsets such as the fly kinase–TF screen (Kim et al. 2025). livia-atlas-{human,zebrafish,yeast,worm}-kinase-kinase.zip: kinase–kinase screens in human, zebrafish, yeast and C. elegans. livia-atlas-viral-dimers-afdb.zip: protein pairs of 2,765 viruses (1,699,288 predictions: every pair within each viral proteome, homodimers included, never across viruses) from the AlphaFold Database viral protein complexes release (EMBL-EBI, Google DeepMind, NVIDIA and collaborators, 2026; CC BY 4.0), one AlphaFold-Multimer model per pair, rescored with lis.py. livia-atlas-afdb-heterodimers-part1.zip to part6.zip: heterodimer screens from the AlphaFold Database protein complexes released with NVIDIA (EMBL-EBI, NVIDIA and collaborators; CC BY 4.0; ftp.ebi.ac.uk/pub/databases/alphafold/collaborations/nvda/heterodimers/), one screen per species (7,573,187 protein pairs, 312,636 proteins), rescored with lis.py. The screens are grouped into six zips because a Zenodo record holds at most 100 files.livia-atlas-afdb-heterodimers-structure-index.zip: where each heterodimer model and its PAE sit in the release archive at EBI, so a single model can be read by byte range. livia-atlas-heterodimers-lis-part01.csv.zst and part02.csv.zst: every heterodimer pair as lis.py scored it (all columns), with the address of its model in the EBI archive. Zstandard-compressed; README_heterodimers_lis_table.md explains how to read the table and open one structure. Each zip is uncompressed, so the website reads single files by byte range. Inside: manifest.json, proteins.json, edges.tsv (pairs past the 10% FPR cutoff), and one bundle per protein in b/ (lis.py rows with residues, and a FASTA). Please cite the source of each screen. Version 0.1.1: FlyPredictome constructs carry the sequences that were folded and are placed on their genes by sequence; kinase–kinase screens added. Version 0.1.2: a gene whose isoforms were folded separately keeps one bundle per isoform, so a page can read the reference first: b/ .zip holds the reference and an isoforms.json listing the others, each in b/ ~ .zip. Version 0.1.3: the viral screen is added. In the existing screens, scores are cut, not rounded (4 decimals in bundles, 3 in edge lists); 3,192 residue lists that the old scorer had shortened are emptied and flagged; missing scores are null; a sequence pair folded more than once is counted once, by its highest-iLIS prediction, while each screen keeps its own; and the fly kinase–kinase screen is complete (34,271 pairs; one batch had been filed with the kinase–TF screen). Version 0.1.4: the AlphaFold Database heterodimer screens are added, with the full lis.py tables and a structure index. The files of 0.1.3 are unchanged (same checksums). Versions were renumbered from 1.x to 0.1.x on 27 September 2026 (1.0 is 0.1.0, 1.1 is 0.1.1, 1.2 is 0.1.2); their files and DOIs are unchanged.
