FoldCraft: An Open Framework for De Novo Design of EGFR-Targeting Nanobodies

1. Overview

This project aims to experimentally validate nanobody binders against EGFR, designed entirely from scratch using our AI-powered platform, FoldCraft. FoldCraft uses AlphaFold2-Multimer as a generative engine and has already demonstrated performance superior to state-of-the-art (SOTA) models like RFdiffusion and RFAntibody in in silico benchmarks.

To test these designs in real-world settings, we will submit our top 50 nanobody candidates to the Adaptyv Bio binding screening platform, a high-throughput experimental assay that rapidly evaluates the binding ability of synthetic proteins.

This work builds on a strong foundation: FoldCraft placed 2nd in Round 1 of Adaptyv Bio’s EGFR binder design competition, showing its potential to generate high-quality miniproteins binders with therapeutic relevance.

The project will be completed in 6 months, with no additional infrastructure or ethics approvals required. Most importantly, we commit to open-sourcing all binder sequences and binding data — making this a fully transparent and reusable resource for the broader scientific community.

2. Introduction

The Challenge of De Novo Binder Design

Designing proteins that can selectively bind to molecular targets — like antibodies that block cancer pathways or neutralize pathogens — is one of the grand challenges in bioengineering. Traditional approaches are often slow, expensive, and dependent on animal immunization or brute-force screening of vast experimental libraries.

But what if we could design these proteins directly from code?

Recent advances in AI, particularly models like AlphaFold, RFdiffusion, and RFAntibody, have brought us closer to this vision. However, most existing tools still struggle to control the 3D fold of the resulting protein, limiting their ability to produce compact, stable, and therapeutically relevant scaffolds such as VHH nanobodies.

This is particularly important when targeting proteins like EGFR (Epidermal Growth Factor Receptor) — a well-characterized cancer target implicated in multiple tumor types. Its clinical relevance makes EGFR an ideal benchmark for evaluating the next generation of AI-driven protein design tools.

Enter FoldCraft

FoldCraft introduces a new approach called fold-conditioned hallucination, which combines AlphaFold2-Multimer’s generative capabilities with a single, interpretable contact map loss to guide structure formation. Rather than producing arbitrary shapes, FoldCraft steers sequence design toward antibody-like folds, enabling precise control over structural topology from the start.

In benchmarking studies, this simple yet powerful method leads to dramatic improvements. FoldCraft has outperformed state-of-the-art models like RFdiffusion and RFAntibody across multiple targets, achieving up to 19.5% success rates in generating valid binders against EGFR, PD-1, and PD-L1 — compared to less than 1% in many competing models.

Why This Project Matters

We believe FoldCraft’s nanobody designs are ready for experimental testing. This project will determine:

· Whether FoldCraft’s in silico success translates to real-world binding.

· Whether AlphaFold3 confidence scores predict actual binding.

· Whether we can publicly release a dataset of fully de novo nanobody sequences and measured affinities — something rarely done in the field.

Key Hypotheses

H1: FoldCraft-designed nanobodies will show measurable binding to EGFR in wet-lab screening.

H2: AlphaFold3 metrics (ipTM, pLDDT, iPAE) correlate with binding strength.

H3: Optimization of only non-interface regions with SolMPNN will improve experimental binding compared to existing strategies.

3. Methods

Sample

We will generate a large initial set of at least 500 nanobody sequences against EGFR using FoldCraft, focusing on structural designs that adhere to a canonical VHH fold. All sequences will undergo interface-preserving optimization using SolMPNN, and only those passing strict structural filters (e.g., AlphaFold3 plDDT > 70, ipAE < 10, ipTM > 0.5) will be retained.

From this filtered pool, we will select the top 50 designs based on AlphaFold3 confidence metrics and sequence/epitope diversity for submission to experimental screening.

In addition to high-throughput binding screens, we will order the top-ranked (Top-1) FoldCraft design and perform a functional validation in EGFR-positive cancer cell lines using standard cell-viability assays to assess antiproliferative effects. These experiments will be conducted with existing institutional resources and/or in-kind support without altering the $5,000 allocation to Adaptyv Bio.

Procedures1. Step 1: In Silico Design

· Generate nanobody scaffolds conditioned on the VHH fold.

· Define contact regions based on known EGFR epitopes.

· Optimize sequences using SolMPNN.

2. Step 2: Selection and Submission:

· Rank designs by AlphaFold3 metrics.

· Cluster for diversity and select 50 designs.

3. Step 3: Experimental Screening (Adaptyv Bio):

· Submit all 50 sequences to Adaptyv Bio, who will synthesize, express, and test them against EGFR.

· Results will include quantitative binding scores.

4. Step 4: Data Analysis:

· Compute correlations between AlphaFold3 predictions and experimental scores.

· Visualize all structures, docking geometries, and experimental rankings.

Ethics and Data Management

Cell-line work will follow institutional biosafety policies; authenticated lines and mycoplasma-free status will be documented. All raw viability data, analysis scripts, and the Top-1 sequence will be open-sourced alongside binding data. All sequences, AlphaFold3 models, and binding data will be:

· Uploaded to GitHub (FoldCraft repo)

· Published on Zenodo with open CC-BY licensing

· Included in a future preprint or publication

4. Pilot Data

We have already validated FoldCraft computationally and published a preprint:

· Designed over 160 EGFR nanobody binders.

· Achieved >19% success under AlphaFold3 evaluation thresholds.

· Outperformed RFAntibody (SOTA model), which failed to generate binders passing ipAE or ipTM cutoffs on diverse set of therapeutically relevant targets (PD-L1, PD1, EGFR, IFNAR).

Additionally, our pipeline placed 2nd out of 74 submissions in the Adaptyv Bio EGFR binder design competition, by designing de novo EGFR miniproteins binder with KD of 4.77*10-6 M.

This proposal builds directly on that momentum.

5. Budget

All funds will be used to run 50 binding screens at Adaptyv Bio.

Item	Description	Cost (USD)
Adaptyv Bio binder screening	50 de novo nanobody sequences vs. EGFR	$5000
Total		$5000

Computational modeling is performed on existing local GPU infrastructure (RTX 3090/4080/4090).

6. References

1. Rustamov, K.R., and Baev, A.Y. Fold-Conditioned De Novo Binder Design via AlphaFold2-Multimer Hallucination. bioRxiv, 2025.

2. Cotet, T.-S., et al. Crowdsourced Protein Design: Lessons From the Adaptyv EGFR Binder Competition. bioRxiv, 2025.

3. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature, 2021.

4. Dauparas, J. et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science, 2022.

5. Bennett, N.R. et al. Atomically accurate de novo design of antibodies with RFdiffusion. bioRxiv, 2025.

6. Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold3. Nature, 2024.

FoldCraft: An Open Framework for De Novo Design of EGFR-Targeting Nanobodies

FoldCraft: An Open Framework for De Novo Design of EGFR-Targeting Nanobodies

Funding Progress

Funders

Related Funding

Request for Proposals - Bioinformatics & Genomics Research

Nonprofit Recipient

Topics

DOI