Efficient exploration of peptide libraries using active learning with AlphaFold-based screening

This paper demonstrates that an active learning strategy based on Thompson sampling significantly improves the efficiency of screening peptide libraries for BET protein binders using AlphaFold2, recovering 50% of known binders with only 15% of the queries required for exhaustive sampling.

The Gist

The Big Problem: Finding a Needle in a Cosmic Haystack

Imagine you are a detective trying to find a specific type of key (a peptide) that fits into a very specific lock (a protein called BRD3). These keys are crucial because they can turn off bad signals in the body that cause diseases.

The problem? You have a library containing 142,000 different keys. Some of them fit the lock perfectly, but most are useless junk.

In the past, scientists tried to find the good keys by testing every single one of them. This is like trying every key in a giant ring to see which one opens a door. It works, but it takes forever and costs a fortune in computer time. If you tried to do this for every protein in a virus, it would be impossible.

The Old Way vs. The New Way

The Old Way (Exhaustive Search):
Imagine walking into a massive library with 142,000 books. You want to find the 3,393 books that are about "cooking." The old method says: "Pick up every single book, read the first page, and check if it's a cookbook." You will eventually find them all, but you will be exhausted and have wasted time reading thousands of books about gardening or history.

The New Way (Active Learning with AlphaFold):
The researchers used a super-smart AI (called AlphaFold) that can predict if a key fits a lock just by looking at a picture of it. But even with this AI, checking 142,000 keys is too slow.

So, they invented a smarter strategy called Thompson Sampling. Think of this as a "Smart Detective" who knows how to gamble.

The Analogy: The Casino Slot Machine

To understand how their "Smart Detective" works, imagine a casino with 1,000 slot machines (these are your groups of keys, or "clusters").

• You don't know which machines pay out (have the good keys).
• Some machines are "hot" (they pay out often).
• Some machines are "cold" (they never pay out).
• Some machines are "unknown" (you haven't played them yet).

The Goal: You have a limited number of coins (computer time). You want to win as many jackpots (find the binding keys) as possible without running out of coins.

How the "Smart Detective" Plays:

1. The Guess: The detective looks at the machines. For the ones they haven't played, they imagine a "ghost" version of the machine that might be a jackpot winner.
2. The Gamble: They pick the machine that might be the best winner right now.
3. The Test: They pull the lever (run the AI check).
   • If they win: They keep playing that machine because it's clearly a "hot" machine.
   • If they lose: They stop playing that machine for a while and try a different one.
4. The Balance: The detective is smart enough to keep trying the "unknown" machines just in case they are actually the best ones, but they mostly stick to the ones that are already showing signs of paying out.

What Happened in the Experiment?

The researchers applied this "Casino Strategy" to their 142,000 peptide keys.

• The Result: They managed to find 50% of all the good keys by only checking 15% of the total library.
• The Comparison: If they had just picked keys randomly (like a drunk person pulling slot machine levers), they would have needed to check 3.3 times more keys to find the same amount of good ones.
• The Speed: They found the most famous, biologically important keys (the ones scientists already knew existed) much faster than the random method.

Why Did It Work So Well?

The secret sauce was grouping.

Instead of treating every single key as a separate machine, they grouped similar keys together into "clusters" (like grouping all red keys, all blue keys, etc.).

• They discovered that the "good" keys were crowded into just a few specific groups.
• Once the "Smart Detective" found a winning group, it focused all its energy there, ignoring the groups that were full of junk.

The Bigger Picture

This isn't just about finding keys for one specific lock. The method is a universal tool.

• For Medicine: It can help find new drugs faster by screening millions of possibilities without needing a supercomputer for years.
• For Viruses: It can quickly scan a virus's entire genetic code to see which parts might attack human cells.
• For Other Properties: You can use this same "Smart Detective" to find peptides that are soluble (dissolve well in water) or don't clump together, which is vital for making stable medicines.

The Takeaway

The paper shows that we don't need to check everything to find the best things. By using a smart, adaptive strategy (Thompson Sampling) combined with powerful AI (AlphaFold), we can explore the vast universe of biological molecules efficiently, saving time, money, and computing power. It's the difference between searching a haystack by hand versus using a metal detector that learns where the needles are as you go.

Technical Summary

1. Problem Statement

• The Challenge: Protein-peptide interactions are critical for cellular processes and therapeutic targeting, but the vast size of peptide sequence space (e.g., $20^{12}$ for a 12-residue peptide) makes exhaustive screening computationally infeasible.
• Limitations of Current Methods:
  • Classical Docking: Struggles with intrinsically disordered peptides that fold upon binding, leading to unreliable scoring.
  • AlphaFold2 (AF2) Screening: While AF2 has improved the ability to model peptide-protein complexes (specifically via the AlphaFold Competitive Binding Assay or AF-CBA), running exhaustive AF2 predictions on large libraries (e.g., viral proteomes or thousands of pull-down candidates) is computationally prohibitive.
  • Random Sampling: Uniformly random selection of peptides is inefficient, often requiring a massive number of queries to find a significant fraction of binders.
• The Goal: Develop a strategy to efficiently explore peptide sequence space to identify binding epitopes with a limited computational budget, prioritizing candidates that yield the most informative results.

2. Methodology

The authors propose an Active Learning framework based on Thompson Sampling (TS), a Bayesian multi-armed bandit algorithm, to optimize the search for peptide binders.

A. Data Generation & Labeling

• Source Data: Derived from a previous study involving pull-down experiments with the BRD3 protein (targeting the Extraterminal/ET domain of BET proteins).
• Library Construction: 318 human proteins were fragmented into overlapping 25-amino acid peptides (sliding window, 1-AA step), generating 142,338 unique peptides.
• Labeling (Ground Truth): Each peptide was screened against the BRD3-ET domain using AF2 (ColabFold).
  • Criteria: A peptide is a "binder" (label = 1) if at least 4 out of 5 predicted models satisfy:
    1. Mean peptide pLDDT score $\ge$ 70.
    2. Average $C_\alpha-C_\alpha$ distance to key binding residues (I42, E43, I44) $<$ 20 Å.
  • Result: 3,393 binders and 138,945 non-binders were identified.

B. Clustering Strategy

To reduce the search space, the peptide library was partitioned into clusters (acting as "arms" in the bandit problem) using sequence similarity:

• Algorithms: CD-HIT, MMseqs2 (LINCLUST), and MMseqs2 (LINCLUST/MMseqs2).
• Parameters: Tested at sequence identity thresholds of 0.4, 0.5, 0.7, and 0.9.
• Rationale: Clusters with high sequence similarity are assumed to share similar binding properties. This allows the algorithm to infer the binding probability of a whole cluster based on a few sampled peptides.

C. Thompson Sampling Workflow

1. Initialization: Each cluster $c$ is assigned a Beta distribution prior, $\theta_c \sim \text{Beta}(\alpha_0, \beta_0)$, representing the estimated probability of finding a binder. Priors were initialized based on the global hit rate (~2.4%) or a conservative bias ($\beta_0 > \alpha_0$).
2. Seed Set: A random seed set of peptides is queried first to update the initial priors.
3. Iterative Selection (The Loop):
   • Sampling: For each iteration, a plausible binder rate $\tilde{\theta}_c$ is sampled from the posterior Beta distribution of every non-exhausted cluster.
   • Selection: The top $k$ clusters with the highest sampled $\tilde{\theta}_c$ are selected.
   • Allocation: A fixed batch size (50 peptides) is distributed among selected clusters (using proportional or equal allocation).
   • Update: The labels of the queried peptides are revealed. The Beta parameters for the touched clusters are updated: $\alpha_c = \alpha_0 + \text{wins}$, $\beta_c = \beta_0 + \text{losses}$.
4. Termination: The process repeats until the query budget is exhausted.

3. Key Contributions

• Novel Application: First application of Thompson Sampling to large-scale peptide sequence space exploration using AF2-based binary labels.
• Efficiency: Demonstrated that active learning can recover a substantial fraction of binders with significantly fewer queries than random sampling.
• Transferability: Showed the method is not limited to binding affinity but applies to any peptide property with binary labels (e.g., solubility).
• Biological Insight: Proved that the method prioritizes known, biologically validated binding epitopes earlier in the search process.

4. Key Results

• Performance vs. Random Sampling:
  • Using a 0.5 sequence identity threshold (CD-HIT), Thompson Sampling recovered 50% of all binders using only 15% of the total queries required for exhaustive screening.
  • This represents a 3.3-fold improvement in efficiency over uniform random sampling.
  • At specific checkpoints (30k, 50k, 70k queries), TS was 2.9x, 2.2x, and 1.78x more efficient, respectively.
• Impact of Clustering Granularity:
  • TS performed best at moderate identity thresholds (0.4–0.7).
  • At 0.9 identity, performance dropped because binders were spread across too many sparse clusters, making it difficult for the algorithm to identify "enriched" regions.
  • At 0.5 identity, 50% of all binders were concentrated in only 151 clusters, allowing TS to quickly identify and exploit these high-value regions.
• Recovery of Known Epitopes:
  • TS successfully identified experimentally known binders (BRG1, INO80B, CHD4, NSD3, BICRA) much faster than random sampling.
  • For example, BRG1, INO80B, and CHD4 were recovered in >90% of 100 TS runs within the first 30k queries (20% of the dataset).
• Mechanism of Action:
  • Analysis of Beta distributions showed that TS rapidly shifts probability mass toward clusters containing binders (e.g., Cluster 1105 containing INO80B) while deprioritizing clusters with only non-binders (e.g., Cluster 1).
  • The algorithm effectively balances exploration (testing uncertain clusters) and exploitation (focusing on clusters with high sampled success rates).
• Generalizability:
  • When applied to a solubility prediction task (using NetSolP labels), TS showed similar enrichment performance, confirming its utility for general peptide property optimization.

5. Significance

• Scalability: This approach makes it feasible to screen massive datasets (such as entire viral proteomes or large-scale proteomics data) for peptide interactions, which was previously impossible due to the computational cost of running AlphaFold2 millions of times.
• Resource Optimization: By reducing the number of required AF2 calculations by a factor of 3–4 (or more depending on the target fraction), the method significantly lowers computational costs and time-to-discovery.
• Framework for Discovery: The study establishes a robust pipeline for "AI-driven experimental design," where computational models guide the selection of candidates for high-cost validation (either further computation or wet-lab experiments).
• Broad Applicability: The framework is agnostic to the specific biological target, provided a method exists to generate binary labels (binder/non-binder, soluble/insoluble, etc.), making it a versatile tool for peptide drug discovery and protein engineering.