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

FlyPredictome presents a comprehensive structural atlas of 1.5 million predicted protein-protein interactions in Drosophila, utilizing AlphaFold-Multimer to map binding interfaces, validate functional relevance through mutation analysis, and provide an open database for exploring the organism's interactome.

The Gist

Imagine the cell as a bustling, high-tech city. In this city, proteins are the workers, machines, and messengers. For the city to function, these workers need to talk to each other, shake hands, and work together in teams. These "handshakes" are called Protein-Protein Interactions (PPIs).

For a long time, scientists knew which workers in the fruit fly (Drosophila) city were supposed to work together, but they didn't know how they held hands. They didn't know the shape of their hands or which specific fingers touched. Without this 3D map, it's hard to understand exactly how the city works or why a broken handshake causes a traffic jam (disease).

Enter FlyPredictome, a new digital atlas created by researchers at Harvard. Here is how they built it, explained simply:

1. The Problem: The "Black Box" of Handshakes

Scientists have spent decades mapping the "who talks to whom" list for fruit flies. But knowing two people are friends doesn't tell you if they are holding hands, hugging, or just waving from across the room.

• The Old Way: They used experiments to find connections, but these were like looking at a blurry photo. You could see two people were close, but you couldn't see the details.
• The New Tool: They used a super-smart AI called AlphaFold-Multimer. Think of this AI as a master architect who can look at two blueprints (protein sequences) and instantly build a 3D model of how they fit together.

2. The Challenge: The "Wiggly" Workers

Here's the catch: Many proteins aren't stiff bricks; they are like wiggly noodles or floppy scarves. Standard AI tools often get confused by these floppy parts and say, "I can't predict how these fit together."

• The Solution: The researchers invented a new measuring stick called iLIS (integrated Local Interaction Score).
• The Analogy: Imagine trying to judge a handshake. A standard tool looks at the whole arm and says, "This arm is too floppy, I can't tell if they are shaking hands." The new iLIS tool zooms in specifically on the fingers. Even if the arms are wiggling, if the fingers are touching firmly, iLIS says, "Yes! That's a handshake!" This allowed them to find connections that other tools missed.

3. The Proof: The "Broken Handshake" Test

How do you know the AI isn't just making up pretty pictures? The researchers played a game of "Spot the Difference" using real-world data.

• The Test: They looked at fruit flies that had specific mutations (genetic typos) causing bad health or weird behaviors.
• The Discovery: They found that these "typos" almost always happened exactly where the AI predicted the proteins were shaking hands.
• The Metaphor: Imagine a car engine. If you break a specific bolt, the car stops. The researchers found that the "broken bolts" (mutations) were located exactly on the "handshake spots" predicted by the AI. This proved the AI's maps were real and useful.

4. The Result: A Giant Interactive Map

The team didn't just make a few maps; they built a structural atlas containing 1.5 million predicted interactions.

• The Database: They organized this into a giant, interactive network (like a massive subway map).
• The Clusters: They grouped these connections into neighborhoods. For example, they found a "Hippo Pathway" neighborhood where proteins work together to control cell growth, and an "Autophagy" neighborhood where proteins work together to clean up trash inside the cell.
• The Bonus: They found thousands of new connections that no one knew about before, especially involving those "wiggly noodle" proteins.

Why Does This Matter?

Think of FlyPredictome as the Google Maps for the fruit fly cell.

• Before: You knew two places were connected, but you didn't know the road conditions or the exact route.
• Now: You have a turn-by-turn, 3D GPS guide.

This is huge because fruit flies are a model for human biology. If we understand how these proteins fit together in flies, we can better understand how they work in humans. This helps scientists:

1. Design Better Drugs: If you know exactly where a protein handshake happens, you can design a tiny molecule to block it (like putting a gum in the lock) to stop a disease.
2. Fix Broken Engines: If a mutation breaks a handshake, we now know exactly which part of the protein to fix.

In short, FlyPredictome turns a blurry list of names into a crystal-clear, 3D instruction manual for how life works at the molecular level.

Technical Summary

1. Problem Statement

While Drosophila melanogaster is a foundational model for understanding cell signaling and development, the structural basis of its protein-protein interactions (PPIs) remains largely uncharacterized.

• Gap in Knowledge: Existing genetic and biochemical data (e.g., yeast two-hybrid screens, AP-MS) map functional relationships and co-complex associations but do not reveal direct physical binding interfaces, the specific domains mediating contact, or the residue-level details of the interaction.
• Limitations of Current Tools: Global confidence metrics used in AlphaFold-Multimer (AFM), such as ipTM, often fail to score interactions involving intrinsically disordered regions (IDRs) or flexible proteins correctly, leading to the dismissal of biologically relevant PPIs despite clear local interfaces visible in Predicted Aligned Error (PAE) heatmaps.
• Need: A proteome-scale structural interactome for Drosophila that provides residue-level resolution and robustly handles flexible proteins is missing.

2. Methodology

A. Data Generation

• Scale: The authors generated 1.5 million pairwise AFM predictions covering nearly all Drosophila proteins.
• Scope: The dataset includes experimentally reported PPIs, kinase interactomes, secreted ligand-receptor pairs, and homodimers.

B. Development of iLIS (Integrated Local Interaction Score)

To address the failure of global metrics on flexible proteins, the authors developed iLIS, a local confidence metric:

• Components:
  1. LIS (Local Interaction Score): Scores the confidence of the interaction domain defined by PAE ≤ 12 Å.
  2. cLIS (contact-filtered LIS): Further restricts scoring to residues in direct intermolecular contact (Cβ-Cβ distance ≤ 8 Å) within the PAE ≤ 12 Å domain.
• Calculation: $iLIS = \sqrt{LIS \times cLIS}$ (Geometric mean).
• Advantage: This formulation integrates domain-level and contact-level confidence, penalizing predictions where proteins are close in space but lack direct contacts, while robustly scoring interactions involving flexible/disordered regions that global metrics miss.

C. Benchmarking and Validation

• Metrics Comparison: iLIS was benchmarked against 774 positive PPIs (from yeast, fly, and human Y2H screens and ELM database) and 6,004 controls.
• Performance: iLIS showed the highest correlation with positive PPIs and significantly outperformed ipTM and other metrics, particularly for proteins with low pLDDT scores (0–50), which are characteristic of flexible regions.
• Thresholds: A threshold of iLIS ≥ 0.223 (at 10% False Positive Rate) was established for classifying positive PPIs.

D. Functional Validation via Allele Enrichment

• Hypothesis: If predicted interfaces are functionally relevant, missense mutations associated with phenotypes should be enriched at these sites.
• Analysis: 2,761 unique missense mutations from FlyBase were mapped to AFM models.
• Statistical Tests: The authors computed LIR (Local Interaction Residue) and cLIR (contact LIR) percentages. They used Cochran-Armitage trend tests and permutation tests to show that mutation rates increase monotonically with interaction frequency, even after controlling for evolutionary conservation (PhyloP scores).

E. Network Construction

• Evidence-Supported Network: A hierarchical network was built using positive predictions supported by either Drosophila literature or high-confidence interologs (DIOPT ≥ 4).
• Clustering: The Leiden algorithm was used to identify 31 functional clusters, which were further recursively sub-clustered to resolve specific protein complexes.

3. Key Results

• Recovery of Experimental PPIs:
  • iLIS recovered ~20% of FlyBi Y2H interactions, 16% of DPiM AP-MS, and 30% of literature-curated PPIs.
  • Crucially, iLIS rescued 47 well-established PPIs that fell below the ipTM threshold, including interactions involving flexible proteins (e.g., p53–MDM2, Emc–Da).
• Structural Models of Signaling Pathways:
  • The study provided residue-level structural models for the entire EGFR/Ras-MAPK cascade and 12 other major pathways (e.g., Hippo, Wnt, Notch, JAK-STAT).
  • Example: The predicted Chip–Apterous interface coincided with experimentally mapped biochemical regions.
• Mutation Enrichment:
  • Phenotype-associated missense mutations were 6.3-fold enriched in interaction domains and 2.1-fold enriched at direct contact interfaces compared to non-interacting residues.
  • This enrichment persisted even when controlling for evolutionary conservation, proving the functional constraint is specific to the interaction interface.
  • Case Studies: Mutations in the Fz–Dsh complex (Planar Cell Polarity) and Sce–Su(z)2 (PRC1 ubiquitylation) mapped precisely to the predicted binding interfaces, validating the structural accuracy.
• Hierarchical Network Organization:
  • The final network contains 21,657 PPIs involving 5,224 proteins.
  • Hierarchical clustering successfully resolved broad pathways into specific modules (e.g., separating the PI3K nucleation complex from the autophagosome assembly machinery in the autophagy cluster).

4. Key Contributions

1. FlyPredictome Database: The first proteome-scale structural interactome for Drosophila, containing 1.5 million predictions, available as an open resource (https://www.flyrnai.org/tools/fly_predictome).
2. iLIS Metric: A novel, robust confidence metric that significantly improves the detection of PPIs involving intrinsically disordered and flexible proteins, outperforming standard global metrics like ipTM.
3. Residue-Level Resolution: Provides specific binding interface predictions (residue-level) for thousands of interactions, enabling the design of interaction-disrupting mutations.
4. Genetics-Structure Integration: Demonstrates a powerful synergy between classical genetics and structural prediction by showing that phenotype-associated alleles cluster at predicted interfaces.
5. Functional Annotation: The hierarchical network assigns uncharacterized proteins to specific functional modules based on their interaction partners (e.g., predicting CG12645 as a membrane-associated Hippo regulator).

5. Significance

• Resource for the Community: FlyPredictome serves as a foundational structural atlas for the Drosophila research community, bridging the gap between genetic screens and structural biology.
• Discovery of Novel Interactions: The database identifies over 100,000 novel interactions lacking prior experimental support, offering new hypotheses for experimental validation.
• Mechanistic Insight: By visualizing the structural basis of signaling pathways, the study helps explain how specific mutations disrupt cellular function, aiding in the interpretation of genetic screens and disease models.
• Methodological Advancement: The success of iLIS suggests that local confidence metrics are essential for accurate structural interactome prediction, especially in organisms with high proportions of disordered proteins.

Limitations Noted:

• Predictions do not account for post-translational modifications, cofactors, or cellular compartmentalization.
• The network is limited to binary interactions and may miss higher-order assemblies or transient interactions dependent on specific cellular contexts.
• Recovery rates for AP-MS data are lower than Y2H, likely due to indirect associations in co-complex data.

In conclusion, FlyPredictome transforms the understanding of Drosophila biology by providing a structural framework that connects genetic phenotypes to physical molecular mechanisms, validated through rigorous statistical and evolutionary analysis.