HInt: Interaction-based homology discovery through accelerated genome-scale AlphaFold screening 1 Rouger et al. propose “interaction-based similarity” as a third axis for homology detection, complementary to sequence (e.g., BLAST) and structure (e.g., Foldseek), targeting cases where both signals have eroded but binding interfaces remain evolutionarily constrained. 2 The work introduces HInt (Homology by Interaction), an accelerated AlphaFold-based framework that makes proteome-scale protein–protein interaction (PPI) screening practical via biologically informed pre-filtering plus optimized high-throughput multimer modeling and scoring. 3 HInt formalizes two interaction-driven homology modes: (i) remote homologues inferred because each protein binds its (homologous) partner through a conserved interface, and (ii) interolog groups where multiple proteins in the same organism bind a common partner using the same binding site, revealing unannotated members. 4 Key engineering step: a pre-filtering stage that reduces the combinatorial one-vs-all search space using user-configurable biological priors (protein length, signal peptide via SignalP5, and subcellular localization via DeepLoc2/DeepLocPro), so expensive modeling is focused on plausible candidates. 5 To remove the MSA bottleneck, HInt reuses precomputed MSAs from the AlphaFold Database when UniProt IDs exist, and otherwise runs GPU-accelerated MMseqs2; it reports shallow alignments (<100 sequences) as potential confidence risks and caches MSAs for reuse across runs. 6 For structure prediction throughput, HInt combines multi-GPU parallelization with VRAM-aware batching (memory estimates based on complex size) and a dynamic job queue, yielding ~4.5x faster modeling for AF-Multimer and AF3 without changing output quality. 7 For interaction ranking, the authors benchmark multiple interface/complex quality metrics (pi-score, pDockQ/pDockQ2, ipSAE, LIS/iLIS, actifpTM, etc.) on 500 IntAct positives + 500 Negatome negatives; iQ-score performs best and becomes the default, with CPU-parallel scoring for scale. 8 End-to-end performance: compared to AlphaPulldown2 (Snakemake) on a one-vs-proteome screen, HInt runs ~5x faster without pre-filtering; with pre-filtering, overall acceleration reaches ~25x while returning the same set of modeled PPIs. 9 Biological validation in bacteria (remote homologue): in the F-plasmid type IV secretion system, long thought to lack core protein VirB5, HInt uses TraGB6 as bait and identifies TraH as a VirB5-like tip protein; TraH localizes to the pilus tip by western blot of purified pili and immunogold EM, despite <3% sequence identity and no meaningful monomer structural similarity (TM-score ~0.14). 10 Biological validation in yeast (interolog): screening 6,033 cytoplasm/nucleus proteins against Skp1, 20 of the top 22 hits are known F-box proteins (including several missing InterPro F-box annotations), and the unannotated Ydr249C ranks #2; NanoBiT confirms SCF incorporation dependent on its predicted F-box-like region (aa 71–118) and on Skp1, motivating the proposed name Fbl1. 💻Code: github.com/Qrouger/HInt 📜Paper: biorxiv.org/content/10.64898… #computationalbiology #bioinformatics #proteininteractions #alphafold #structuralbioinformatics #proteomics #genomeannotation #evolution #ppi #ubiquitin