High-throughput prediction of protein-protein interactions uncovers hidden molecular networks in biosynthetic gene clusters

This study presents a high-throughput pipeline that leverages AlphaFold3 and MMSeqs2 to systematically predict protein-protein interactions across nearly 500,000 pairs in biosynthetic gene clusters, successfully uncovering hidden molecular networks and functional enzyme complexes involving previously uncharacterized proteins.

The Gist

The Big Picture: Solving the "Lego" Mystery

Imagine you have a massive box of Lego bricks. Inside, there are instructions for building thousands of different, incredibly complex machines (like a working car, a robot, or a castle). These instructions are written in a secret code called Biosynthetic Gene Clusters (BGCs).

Nature uses these "instruction manuals" to build secondary metabolites—special chemicals that bacteria and fungi make. Some of these chemicals are life-saving antibiotics, while others are dangerous poisons.

The Problem:
For a long time, scientists could read the code (the DNA) and see the list of parts (the proteins). But they didn't know how the parts fit together. It was like having a list of Lego bricks but no idea which ones snap together to make the wheels, the engine, or the roof. Many of the instructions just said "Unknown Part" or "Hypothetical Brick," leaving huge gaps in our understanding of how these biological machines work.

The Solution: A High-Speed "Snap-Together" Simulator

The researchers in this paper built a super-fast computer program to solve this puzzle. They used a powerful AI tool called AlphaFold 3 (think of it as a digital wizard that can guess how two Lego bricks will snap together just by looking at their shapes).

However, the standard version of this wizard is slow and has strict limits on how many times you can ask it to work. To fix this, the team built a "turbo-charged" version of the pipeline. They swapped out the slow part of the process for a faster one (using a tool called MMSeqs2), allowing them to run millions of simulations in record time.

What they did:
They took 2,437 different "instruction manuals" (gene clusters) from a massive database and asked the computer to check every single pair of proteins within them. They asked: "If Protein A meets Protein B, do they snap together to form a machine?"

The Results: Finding Hidden Connections

After crunching the numbers on nearly 488,000 protein pairs, they found some amazing things:

1. The "Unknowns" aren't useless: They found that many proteins previously labeled as "Unknown" or "Broken" actually snap together perfectly with other proteins to form working machines.

   • Analogy: It's like finding a weird, jagged Lego piece that you thought was trash, only to realize it's the secret key that locks the engine block together.

2. The "Twin" Trick: They discovered many pairs of proteins that look almost identical (like twins) but work together as a team.

   • Analogy: Imagine two identical twins. One is the driver, and the other is the navigator. Even though they look the same, they have different jobs. The computer figured out that in these gene clusters, these "twins" often team up to do a specific job that neither could do alone.

3. The "Helper" Proteins: Some proteins can't do the work alone; they need a partner to become active.

   • Example: In one case, they found a protein that looked like it was broken (missing a key part). But when the computer simulated it snapping together with its partner, the "broken" part was fixed by the partner's shape. It turned out the "broken" protein was actually a helper that stabilizes the main worker.

Why This Matters

Before this study, scientists had to guess how these biological machines worked, often by trial and error in a lab. This new method acts like a blueprint generator.

• For Drug Discovery: If we know exactly how the "machines" are built, we can engineer them to make new, better antibiotics or stop them from making toxins.
• For Biology: It helps us understand how nature builds such complex chemistry.

The "Caveats" (Where the Wizard Stumbles)

The authors are honest about the limits of their tool:

• The "Ghost" Interactions: Some proteins only touch each other for a split second (like a handshake) before moving on. The computer is great at finding things that are glued together, but it sometimes misses these quick, fleeting handshakes.
• The "Big" Bricks: The tool couldn't handle the massive proteins (which are like huge, pre-assembled Lego sets) because they were too big for the computer's memory.
• Shape-Shifting: Some proteins change their shape when they meet a partner. The computer sometimes predicts the "before" shape instead of the "after" shape, leading to a slightly wrong guess.

The Takeaway

The researchers have created a public map (a website) where anyone can look up these gene clusters and see the predicted connections between the proteins. They have turned a massive, confusing puzzle of "unknown parts" into a clear map of who works with whom.

In short: They built a high-speed simulator that tells us which biological parts snap together, revealing hidden teamwork in nature's chemical factories and giving us a better chance to harness these processes for medicine and science.

Technical Summary

1. Problem Statement

Biosynthetic Gene Clusters (BGCs) encode the machinery for producing diverse natural products (secondary metabolites) such as antibiotics and mycotoxins. While tools like antiSMASH can identify BGCs, a significant portion of the proteins within these clusters are annotated as "hypothetical," "uncharacterized," or "potentially dysfunctional."

• Limitation of Current Tools: Conventional sequence-based bioinformatics tools fail to predict the function of these proteins without known homologs.
• The Gap: Protein-protein interactions (PPIs) are crucial for intermediate transfer, enzymatic regulation, and stability in BGCs. However, experimentally resolving these interactions for the vast number of uncharacterized proteins is slow and resource-intensive.
• Challenge: Existing high-accuracy structure prediction tools (like AlphaFold3) are computationally expensive and limited in throughput, making systematic screening of all pairwise interactions within thousands of BGCs infeasible.

2. Methodology

The authors developed a high-throughput computational pipeline to predict protein complexes within BGCs using AlphaFold3 (AF3).

• Data Source: The study focused on 2,437 "active" BGCs from the MIBiG database (version 4.0).
• Filtering: Proteins longer than 1,950 amino acids (typically modular PKS/NRPS proteins) were excluded due to computational constraints.
• Workflow Optimization:
  • MSA Generation: To overcome the speed bottleneck of AF3's default MSA generation (HMMER), the authors replaced it with MMseqs2 via LocalColabFold. They generated both unpaired MSAs (for individual proteins) and paired MSAs (concatenating sequences from the same organism) to improve complex prediction accuracy.
  • Prediction Scale: They performed exhaustive pairwise predictions ($N(N+1)/2$) for all proteins in each BGC, covering 487,828 protein pairs.
  • Hardware: Predictions were distributed across GPU-equipped servers (V100, RTX4090, H100) based on complex size.
• Scoring Metrics:
  • ipTM (interface predicted TM-score): Measures the confidence of the relative arrangement of subunits.
  • ipSAE (interaction prediction score from aligned errors): A refined metric focusing on inter-chain residue pairs with reliable PAE scores, designed to distinguish true interactions from artifacts, especially in cases with multiple homologous proteins.
• Validation: Predicted models were validated against experimentally determined structures in the Protein Data Bank (PDB) with >95% sequence identity to MIBiG entries.

3. Key Contributions

1. High-Throughput Pipeline: Established a scalable workflow replacing slow MSA generation with MMseqs2, enabling the prediction of nearly half a million protein pairs.
2. Comprehensive Interaction Map: Generated a reusable dataset of predicted PPIs for 2,437 BGCs, visualized via a web interface (VIVACE).
3. Metric Refinement: Demonstrated that ipSAE is superior to ipTM for distinguishing correct heterocomplex pairs when multiple structurally similar proteins exist within a cluster (e.g., distinguishing specific KS-CLF pairings).
4. Discovery of "Hidden" Complexes: Identified functional mechanisms for proteins previously labeled as hypothetical or dysfunctional, revealing that many require complex formation to function.

4. Key Results

• Prediction Statistics:
  • Predicted 15,438 heteromeric interactions with ipTM ≥ 0.6.
  • Identified 1,390 pairs with high structural homology (RMSD ≤ 2.0 Å).
• Validation Performance:
  • Homodimers: 87.7% of known homodimers were predicted with ipTM ≥ 0.6.
  • Heterodimers: 26 out of 30 known heterodimers were successfully predicted (ipTM ≥ 0.6).
  • Limitations: The pipeline struggled with transient complexes (e.g., Carrier Proteins like ACPs) and those induced by artificial tags or cross-linking reagents in crystal structures.
• Novel Biological Insights:
  • Hypothetical Proteins: Identified a heterodimer (cae45685.1/cae45686.1) in the borrelidin cluster that mimics the Naa30 subunit of the NatC complex, suggesting a role in N-terminal acetylation of borrelidin B.
  • DUF Proteins: Predicted a complex between a DUF4399 protein and a DoxX-family protein in the tropodithietic acid cluster, suggesting a membrane-bound enzymatic function.
  • Structural Homologs: Confirmed that AF3 can correctly identify specific heterodimeric pairings (e.g., KS-CLF in Type II PKS) even when the monomers share high structural similarity, a task where homodimer prediction often fails.
  • Auxiliary Proteins: In the armeniaspirol cluster, predicted that a "dysfunctional" halogenase-like protein (Ams8) lacks catalytic residues but forms a heterodimer with Ams9 to facilitate trichlorination, acting as an essential auxiliary protein.

5. Significance

• Functional Annotation: This study provides a "bottom-up" approach to annotate uncharacterized proteins in BGCs by inferring function from structural interaction networks rather than sequence homology alone.
• Metabolic Engineering: By identifying which proteins must be co-expressed to form functional complexes (e.g., MonBI-MonBII), the data guides the rational engineering of enzymes and the reconstruction of biosynthetic pathways in heterologous hosts.
• Resource Availability: The authors released a web server for visualizing PPI networks and a Zenodo repository containing all JSON files and Python scripts, enabling the broader research community to validate and utilize these predictions for experimental design.
• Evolutionary Insight: The prevalence of structurally homologous heterocomplexes suggests that gene duplication in BGCs often leads to specialized regulatory or auxiliary roles rather than simple redundancy, challenging the "genome streamlining" theory in the context of secondary metabolism.

In conclusion, this paper bridges the gap between genomic data and structural biology, offering a powerful computational framework to decode the "dark matter" of biosynthetic gene clusters and accelerating the discovery of novel natural products.