FlyPredictome: A structural atlas of predicted protein-protein interactions in Drosophila

Key results

1. 01

   ShownLocal interface metrics, iLIS included, outperformed ipTM and pDockQ in every pLDDT stratum; iLIS rescued 47 reference pairs that ipTM missed at matched specificity, tracked DockQ on post-cutoff complexes, and admitted non-contacting subunit pairs about three times less often than ipSAE or ipTM.

   Benchmark of 736 positive vs 5,909 control pairs. On 536 complexes deposited after the AlphaFold-Multimer training cutoff, iLIS vs DockQ Spearman ρ = 0.74 (Pearson r = 0.77); ρ = 0.62 (P = 3 × 10^-8) for 65 with < 30% identity to pre-cutoff PDB chains. At 10% false positive rate (FPR), non-contacting assembly subunit pairs passed iLIS 17% vs ipSAE 53%, ipTM 48%. · iLIS: a local confidence metric for identifying…; Prediction of higher-order protein complexes…; Benchmark dataset construction; Figure legends

2. 02

   ShownPhenotype-associated missense alleles concentrate at residues repeatedly used in predicted interfaces, with enrichment surviving controls for gene-level ascertainment, burial and conservation and seen in every severity class; curated phosphosites are also enriched at direct contacts.

   2,761 mutations in 995 proteins: rate rose from 0.11% (residues never in an interface domain) to 0.69% (top quartile), 6.3-fold; within domains, 0.40% to 0.83% across contact bins, 2.1-fold (within-gene permutation P < 10^-5 for both); burial-adjusted odds ratio 2.6 (P < 10^-100); PhyloP-adjusted P = 10^-59; phosphosites 37.1% vs 26.4% for matched Ser/Thr/Tyr at contacts. · Phenotype-associated alleles are enriched at…; Missense allele enrichment analysis; Phenotype severity and post-translational…

3. 03

   ShownLiterature- or interolog-supported positives form a network whose Leiden clusters and nested sub-clusters show distinct Gene Ontology enrichment and resolve Hippo, autophagy and double-strand break repair pathways into individual complexes, placing uncharacterized CG3448 beside DNA ligase IV.

   1,108,443 unique pairs (~1.1% of possible fly pairs); 144,796 heterodimers and 5,470 homodimers positive at iLIS ≥ 0.223; heterodimers: 15,386 fly-literature, 28,270 interolog (DIOPT ortholog score ≥ 2), 101,140 neither. Network (DIOPT ≥ 4): 5,224 proteins, 21,657 PPIs, 31 clusters, 257 sub-clusters, 407 sub-sub-clusters; CG3448-DNAlig4 iLIS = 0.448. · FlyPredictome database and evidence-supported…; Prediction-set composition; PPI network construction and clustering

4. 04

   ShownCo-folding sub-cluster members with AlphaFold3 largely reproduced deposited cryo-EM assemblies, with completeness declining as subunit and contact numbers grew, and proposed explicit interfaces for structurally unresolved partners (Misato inside the CCT/TRiC chaperonin cavity, CG12725 on eIF3).

   398 assemblies folded; of 202 with an evaluable cryo-EM reference, 178 (88%) recovered every deposited subunit contact (median recall 1.00), from 98% for three-subunit to 75% for eleven-or-more-subunit assemblies; origin recognition complex vs Drosophila cryo-EM (PDB 7JGR): 1.43 Å Cα RMSD, TM-score 0.959. · Prediction of higher-order protein complexes…; Higher-order complex modeling; Figure legends

5. 05

   ShownPositive rates for reported fly PPIs rose with the directness and replication of experimental support, iLIS gained most over ipTM in low-pLDDT pairs, and iLIS rescoring of an independent human predictome recovered nearly all high-confidence SPOC calls.

   Positive at iLIS ≥ 0.223: ~20% FlyBi (yeast two-hybrid), 16% DPiM and 10% DPiM2 (AP-MS), ~30% literature-curated pairs; 68 to 81% for direct-binding and structural assays (ELISA, SPR, X-ray, ITC); 17% (single record) to 75% (≥ 10 records); pLDDT 0 to 50: 22.7% iLIS vs 9.2% ipTM. Human predictome (1,614,047 pairs): 99.5% (16,479 of 16,557) of SPOC ≥ 0.86 calls recovered. · Benchmarking iLIS across experimental and…; Comparison with computational PPI prediction…; Experimental-evidence stratification; Figure legends