FlyPredictome is a structurally resolved atlas of the Drosophila protein interactome. Although Drosophila melanogaster has been a foundational model for dissecting conserved signaling pathways, the structural basis for most of its protein–protein interactions (PPIs) remains uncharacterized. Existing large-scale proteomics catalogs remain incomplete in coverage and provide no information about interaction interfaces.
To address this gap, we used AlphaFold-Multimer (AFM) to generate pairwise structural predictions at proteome scale, including the majority of reported PPIs from fly literature, kinase pairs, secreted ligand–receptor extracellular domain pairs, and many additional candidate interactions. Every prediction is made available with its structural model, PAE heatmap, residue-level interface annotations, and cross-references to experimental evidence in MIST (Hu et al., 2018).
| Metric | FPR 10 | FPR 5 | FPR 1 |
|---|---|---|---|
| iLIS_best | 0.223 | 0.339 | 0.551 |
| iLIS_avg | 0.072 | 0.120 | 0.275 |
| ipSAE_best | 0.164 | 0.363 | 0.615 |
| ipSAE_avg | 0.039 | 0.085 | 0.241 |
| actifpTM_best | 0.745 | 0.880 | 0.963 |
| actifpTM_avg | 0.419 | 0.504 | 0.731 |
| ipTM_best | 0.480 | 0.590 | 0.720 |
| ipTM_avg | 0.292 | 0.336 | 0.442 |
| iLISA_best | 177.1 | 465.7 | 1317.0 |
| iLISA_avg | 57.1 | 135.92 | 479.6 |
| Composite Score | +1 | +2 | +3 |
Global confidence metrics such as ipTM are known to underperform for proteins with low pLDDT, which often reflects intrinsic disorder or flexibility — hallmarks of many signaling components. We therefore developed the integrated Local Interaction Score (iLIS), a local-confidence metric built on the LIS framework (Kim et al., 2024).
iLIS is the geometric mean of two component scores:
Across 774 positive PPIs and 6,004 controls, iLIS outperformed global metrics in every pLDDT subgroup and uniquely rescued 47 well-established PPIs missed by ipTM. We use a 10% false positive rate threshold (iLIS ≥ 0.223) for downstream classification; users can adjust the threshold for their specific dataset.
For each protein, FlyPredictome reports two frequency profiles along its sequence: the interaction domain residue (Local Interaction Residue, LIR) profile shows how often each residue falls within a predicted interaction domain across all partners, and the contact interface residue (contact Local Interaction Residue, cLIR) profile restricts this to residues in direct intermolecular contact.
FlyPredictome also hosts an evidence-supported PPI network: iLIS-positive predictions supported by either fly literature (physical or genetic interaction) or conserved interologs at DIOPT ≥ 4. The network is partitioned by Leiden community detection into clusters, each hierarchically decomposed into sub-clusters and sub-sub-clusters corresponding to coherent functional modules — from entire signaling pathways down to individual protein complexes. An interactive network viewer is available at https://flyark.github.io/FlyPredictome-network/.
For the original LIS framework: Kim, A.-R., Hu, Y., Comjean, A., Rodiger, J., Mohr, S.E., Perrimon, N. (2024). Enhanced protein–protein interaction discovery via AlphaFold-Multimer. bioRxiv.
The iLIS pipeline (AFM-LIS) is available at https://github.com/flyark/AFM-LIS.