Evolutionary algorithms accelerate de novo design of potent Nectin-4-specific cancer biologics

By integrating an evolutionary genetic algorithm with AI-driven structural design, researchers successfully overcame the challenges of targeting Nectin-4 to rapidly generate highly potent, stable minibinders that function effectively as cancer detection reagents and T cell engagers.

The Gist

Imagine you are trying to design a custom-made key that fits perfectly into a very specific, tricky lock. In the world of medicine, this "lock" is a protein on the surface of cancer cells, and the "key" is a tiny, man-made protein designed to grab onto it, stop the cancer, or call in the body's immune system to destroy it.

This paper describes a breakthrough in how scientists design these keys, specifically for a difficult lock called Nectin-4, which is found on many types of cancer.

Here is the story of how they did it, broken down into simple concepts:

1. The Problem: The "AI" Was Stuck

Scientists have recently started using powerful Artificial Intelligence (AI) to design these protein keys from scratch. Usually, this works like magic: you tell the AI, "Here is the lock, make me a key," and it spits out thousands of designs.

However, when they tried to design a key for Nectin-4, the AI hit a wall.

• The Analogy: Imagine the AI is a master chef who can cook amazing meals. But Nectin-4 is a very picky eater with a weird, smooth texture that doesn't hold flavor well. The chef (AI) tried to cook 20,000 recipes, but almost none of them tasted good. The "lock" was just too slippery and boring for the AI to find a good grip.
• The Result: The AI produced very few promising designs. If they had kept going with just the AI, it would have taken them years and massive computer power to find a single working key.

2. The Solution: The "Evolutionary" Coach

Instead of giving up, the scientists added a second step. They took the few "okay" designs the AI made and ran them through a Genetic Algorithm (GA).

• The Analogy: Think of the AI designs as a group of young athletes. Most of them aren't Olympic champions yet. The Genetic Algorithm acts like a tough, smart coach who runs a training camp.
  • Selection: The coach picks the top 10% of athletes (the best designs) and sends the rest home.
  • Mutation & Mixing: The coach takes the best athletes and makes small changes to their training. Maybe they tweak their diet (change a few amino acids), or maybe they combine the best moves of two different athletes (recombination).
  • Repetition: They do this over and over again for 50 "generations." Just like in nature, the weak designs die out, and the strong ones get stronger.

3. The Two Training Methods

The scientists tried two different ways for the coach to train the athletes:

• Method A (Partial Diffusion): This is like taking a clay sculpture of a key and gently reshaping parts of it while keeping the main shape, then asking the AI to fill in the details.
• Method B (Direct Editing): This is like taking a text document of the key's instructions and using a computer program to swap out letters, delete words, or insert new sentences to see if the key works better.

The Result: This "AI + Coach" combo was a game-changer. It found high-quality keys 90 times faster than using the AI alone. They went from having almost no good keys to having hundreds of excellent candidates.

4. The Proof: Testing the Keys

The scientists didn't just trust the computer; they built the keys in the lab and tested them.

• The Flow Cytometry Test: They turned the keys into "glow-in-the-dark" flashlights. When they shined these flashlights on cancer cells, the cancer cells lit up brightly, but healthy cells stayed dark. This proved the keys only grabbed the cancer.
• The T-Cell Engager (The "Handshake"): This is the coolest part. They turned these tiny keys into a biological bridge.
  • One end of the bridge grabs the cancer cell.
  • The other end grabs a T-cell (a soldier from your immune system).
  • The Result: The bridge forces the immune soldier to shake hands with the cancer cell, triggering the soldier to attack and kill the cancer.

5. Why This Matters

• Speed: They found a working solution for a "difficult" target in weeks, not years.
• Versatility: This method works even when the AI is confused. It can take a bad start and turn it into a winner.
• Future Cures: These tiny protein keys are small, stable, and easy to manufacture. They could lead to new, highly specific cancer drugs that are better than current treatments because they don't hurt healthy cells.

In a Nutshell

The scientists realized that while AI is great at dreaming up ideas, it sometimes gets stuck on difficult problems. By adding a layer of evolutionary training (simulating natural selection in a computer), they were able to refine those rough ideas into perfect, life-saving tools. They turned a "bad lock" into a target that can be hit with pinpoint accuracy, potentially leading to new cures for bladder, lung, and breast cancers.

Technical Summary

1. Problem Statement

The authors address a critical bottleneck in AI-driven de novo protein design: target dependency. While AI tools like RFdiffusion have revolutionized structural biology, their success rates vary significantly depending on the physicochemical properties of the target protein's surface.

• Specific Challenge: The cancer antigen Nectin-4 (a cell adhesion molecule overexpressed in urothelial and other solid cancers) proved exceptionally difficult for standard RFdiffusion-based minibinder generation.
• Root Cause: Unlike structurally similar targets like PD-L1, which possess prominent hydrophobic patches that serve as clear binding "hotspots," Nectin-4 lacks these features. Consequently, RFdiffusion yielded a negligible number of high-quality candidates (only ~0.6% with pAE < 10 and 0% with pAE < 5 without predefined hotspots), requiring ~50-fold more computational time to generate a viable library compared to easier targets.
• Goal: To develop a strategy that overcomes these limitations, accelerates the discovery of high-affinity binders for "difficult" targets, and validates their utility as therapeutic biologics.

2. Methodology

The study introduces a hybrid pipeline combining AI-driven generative design with Evolutionary Genetic Algorithms (GAs) to refine and diversify initial candidates.

A. The AI-GA Pipeline

1. Seed Generation: Initial "seed" minibinders were generated using RFdiffusion targeting the Ig-like V-type domain of Nectin-4.
2. Genetic Algorithm (GA) Refinement:
   • Selection: A population of designs was scored based on interaction metrics (pAE interaction scores, ipTM). Low-fitness designs were pruned.
   • Diversification (Two Options):
     • Option I (Partial Diffusion): Used RFdiffusion's partial diffusion feature to introduce structural perturbations to the scaffold, followed by sequence redesign using ProteinMPNN.
     • Option II (Sequence Editing): Used the Chai-1 algorithm (with ESM2 embeddings) to introduce point mutations, insertions, deletions, and crossover operations directly into the amino acid sequences.
   • Iteration: The process ran for 50 generations over multiple independent runs to avoid local minima, effectively exploring the sequence-structure space.
3. Screening Strategy:
   • Mammalian Cell-Surface Display: A library of ~2,000 designs (seeds + GA offspring) was displayed on HEK293T cells.
   • FACS & Sequencing: Cells were stained with human and orthologous (mouse, rat, monkey) Nectin-4 Fc-fusion proteins. Positive populations were sorted, and DNA was extracted for Next-Generation Sequencing (NGS) to identify enriched binders.
4. Functional Engineering:
   • Quatrobinders: Top binders were engineered into tetravalent detection reagents by attaching biotinylated minibinders to streptavidin-AF647.
   • Bispecific T-Cell Engagers (TCEs): Minibinders were fused to an anti-CD3 scFv (OKT3) to create TCEs capable of recruiting T cells to Nectin-4+ tumor cells.

3. Key Contributions

• Novel Optimization Framework: Demonstrated that coupling AI generative models with evolutionary algorithms significantly improves success rates for targets lacking obvious binding hotspots.
• Computational Efficiency: The GA refinement process was found to be ~90-fold more compute-efficient than conventional RFdiffusion runs for generating high-quality candidates (pAE < 5) for Nectin-4.
• Cross-Species Reactivity: The pipeline successfully identified binders with varying degrees of cross-reactivity across human, mouse, rat, and monkey Nectin-4, facilitating preclinical modeling.
• Therapeutic Proof-of-Concept: Successfully converted de novo designed minibinders into functional biologics (flow cytometry reagents and potent T-cell engagers) without relying on natural antibody scaffolds.

4. Key Results

• In Silico Performance:
  • Initial RFdiffusion runs yielded only 8 candidates with pAE < 5 out of 20,000 designs.
  • After GA refinement (Option I), 92.4% of the resulting designs achieved pAE < 5.
  • The refined designs maintained structural diversity (helical bundles) while improving predicted binding affinity.
• Experimental Validation (Binding):
  • Success Rate: In pooled screening, 4.2% of the library was enriched for human Nectin-4 binding. Crucially, none of the original seed designs were recovered as functional binders; all successful hits emerged from GA refinement.
  • Affinity: Surface Plasmon Resonance (SPR) confirmed high affinity. Top binders (e.g., 3-8-16 and 3-24-26) achieved subnanomolar (0.5 nM) and low nanomolar (1.2 nM) dissociation constants ($K_D$).
  • Stability: All selected minibinders showed high thermal stability (melting temperatures > 80°C) and monomeric/oligomeric states without uncontrolled aggregation.
• Specificity and Functionality:
  • Flow Cytometry: Tetravalent "quatrobinders" specifically stained Nectin-4+ cell lines (RT4, HT-1376) but not Nectin-4-negative lines (T24) or CRISPR-generated Nectin-4 knockout cells.
  • T-Cell Engagers: Bispecific TCEs mediated target-specific T-cell activation (upregulation of CD69/CD25) and cytotoxicity against Nectin-4+ tumor cells.
  • Correlation: High-affinity binders (subnanomolar $K_D$) were strictly required for potent T-cell activation, particularly under stringent wash conditions where unbound TCEs were removed.

5. Significance

This work establishes a robust, generalizable strategy for accelerating therapeutic protein engineering for challenging targets.

• Overcoming AI Limitations: It proves that evolutionary algorithms can rescue "failed" AI design targets by efficiently navigating complex sequence-structure landscapes that pure generative models struggle with.
• Therapeutic Relevance: Nectin-4 is a validated target for antibody-drug conjugates (e.g., enfortumab vedotin). This study demonstrates that de novo designed minibinders can match or exceed the performance of conventional antibodies in terms of affinity and specificity, while offering advantages in size, stability, and ease of engineering into multi-specific formats.
• Future Impact: The framework is agnostic to the upstream design algorithm (compatible with RFdiffusion, Chai-1, etc.) and can be adapted for other protein-protein or protein-small molecule interactions, potentially democratizing the development of next-generation cancer biologics.