Towards reconstruction of the human interactome from positive and negative experimental evidence

Key results

1. ShownReconstructing each study's prey search space recovers tested-but-undetected pairs at scale. It gives a small set of high-replication interactions (HRIs) and a very large set of high-replication non-interactions (HRNIs) from AP-MS and yeast two-hybrid (Y2H) data.

   491,601 interactions from 5,530 IntAct studies map to 97 M unique tested pairs over 182 M tests; most pairs were tested once or twice, the most-tested 69 times. HRIs (beta-binomial Q2.5 > 0.29, e.g. 3 of 3 detections): 6,221 pairs, mean detection ratio 0.934. HRNIs (≥3 tests, none detected): 18,982,508 pairs, mean 3.9 tests, max 20. · Mapping the search space; 1. Estimation of pair-wise probability of…; Figure legends

2. ShownHRIs share Gene Ontology biological-process terms, subcellular localisation, high co-expression and loss-of-function intolerance more often than HRNIs. The signal is clearest for AP-MS and weaker for Y2H.

   Shared-annotation OR > 1 (q < 0.05, Benjamini-Hochberg) for 24 of 25 terms in the MS and combined sets vs 9 of 24 for Y2H (nucleoplasm the only enriched Y2H localisation). High ARCHS4 co-expression OR 1.36 to 1.56 across datasets. Both-LoF-intolerant pairs (gnomAD LOEUF): OR 1.18 (95% CI 1.11-1.25) MS, 1.19 (1.12-1.26) combined, 1.08 (0.96-1.21, p=0.16) Y2H. · Validating non-interacting pairs; 2. Co-annotation analysis HRNI vs HRI; Figure legends

3. ShownRepeatedly tested non-interactions give false discovery rate (FDR) estimates for held-out experimental screens and for RF2-PPI structure-based predictions. High-confidence RF2-PPI calls fall among HRNIs far less often, and HRNIs with 3 tests behave much like those with 7.

   One large held-out study per assay: FDR 21% (Y2H) and 40% (MS) when any test counts as negative, 5% and 16% when ≥3 tests are required. RF2-PPI (43.6 M pairs) vs 351,552 reported interactions: FDR 0.886/0.651/0.320 at p > 0.5 and 0.153/0.038/0.010 at p > 0.99 for HRNI tiers ≥3/≥5/≥7 tests. Size-adjusted enrichment was 2.31 to 2.33 and 99.54 to 110.25, nearly the same across tiers. · Experimental negatives as a benchmarking tool; Figure legends; Tables

4. SuggestedD measures how much a protein's detection depends on its partner. Known Y2H auto-activators and CRAPome AP-MS contaminants tend towards low D, and those with the highest detection rates sit near partner-independent detection (D ≈ 1), suggesting D as a stickiness score for unlabelled proteins.

   D is the mean Pearson chi-square contribution against the protein's global detection ratio in a given role. Known problematic proteins with high detection rates (71 curated Y2H auto-activators; 576 CRAPome 1.0 proteins with mean spectral count > 1) cluster at D ≈ 1, while most proteins with similar detection rates spread to higher D. The comparison is visual; no test statistic is reported. · Role-specific binding variance ranks…; Figure legends

5. SuggestedTraining sequence-based classifiers on HRNIs instead of randomly sampled non-observed pairs did not improve accuracy, but it did change which pairs were predicted as non-interacting.

   ESM-2 + XGBoost, node-disjoint degree-balanced splits; accuracy similarly low with either negative type. Jaccard agreement on HRNI-set predictions: median 0.59 to 0.74 for HRNI- vs non-observed-trained models, 0.61 to 0.76 among non-observed-trained ones. HRNI-trained models agreed with each other less than non-observed-trained ones, despite training-set overlap of Jaccard 0.9 to 0.95 vs near zero. · Training and evaluating machine learning of PPIs; Discussion; 4. ML-splits strategy and model setup; Classification setup and protein-embeddings