DiffAb Pipeline: Redesigning CDR Loops with Diffusion Models

Antibody engineering has long relied on labor-intensive wet-lab iteration cycles. Foldexa's DiffAb pipeline changes that — enabling in silico CDR redesign that produces structurally validated candidates in minutes, not months.

What Are CDRs and Why Do They Matter?

Complementarity-Determining Regions (CDRs) are the six hypervariable loops — H1, H2, H3 on the heavy chain; L1, L2, L3 on the light chain — that form the antigen-binding site of an antibody. The H3 loop alone accounts for over 60% of antigen contacts in most antibody–antigen complexes, making it the primary target for affinity maturation and specificity engineering.

Traditional CDR engineering approaches (phage display, directed evolution, rational mutagenesis) are powerful but slow. A single round of phage selection takes 2–3 weeks; affinity maturation campaigns can span 6–12 months. Computational approaches that pre-select high-confidence candidates dramatically compress this timeline.

DiffAb: Diffusion for Antibody Design

DiffAb is an SE(3)-equivariant diffusion model trained on the Protein Data Bank's antibody–antigen complexes. Unlike sequence-only models, DiffAb operates directly in 3D coordinate space — simultaneously denoising both backbone coordinates and amino acid identities. This structure-centric approach enforces geometric consistency from the first generation step.

The model conditions CDR generation on the antigen surface and the fixed antibody framework (Fv region excluding CDRs). Given these constraints, it samples CDR conformations from the learned distribution of natural antibody–antigen interfaces.

How Foldexa Integrates DiffAb

Our Pipeline 1 wraps DiffAb in a fully managed workflow with automatic pre- and post-processing, parallelized generation, and AlphaFold2 validation. You upload a PDB, configure generation parameters, and receive ranked, validated candidates.

1. 1Upload antigen PDB (or fetch directly from RCSB by accession code)
2. 2Optionally provide an antibody framework; otherwise Foldexa selects an appropriate Fv template
3. 3Select CDR loops to redesign (any combination of H1–H3, L1–L3)
4. 4Set generation parameters: num_designs (default 50), sampling temperature (default 1.0)
5. 5Job is dispatched to GPU cluster — typical runtime: 4–8 minutes for 50 designs
6. 6Each design is validated with AlphaFold2 Multimer; scored on pLDDT, iptm, and PAE
7. 7Results ranked and returned with downloadable PDB files

curl -X POST https://api.foldexa.com/v1/jobs \
  -H "Authorization: Bearer $FOLDEXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pipeline": "diffab-cdr-redesign",
    "antigen_pdb": "7KMG",
    "cdrs": ["H1", "H2", "H3"],
    "num_designs": 50,
    "temperature": 1.0
  }'

{
  "job_id": "job_9f3a2b",
  "status": "queued",
  "pipeline": "diffab-cdr-redesign",
  "estimated_runtime_minutes": 6,
  "designs_requested": 50,
  "created_at": "2026-03-18T09:14:32Z"
}

Early Benchmark Results

We benchmarked our DiffAb pipeline against 12 antibody–antigen complexes from the SAbDab database, held out from the model's training set. For each complex, we masked the CDR H3 loop and asked the model to redesign it from scratch, then compared the top-ranked design to the experimentally determined structure.

Interpreting Your Results

Foldexa ranks designs using a composite Foldexa Score that integrates three AlphaFold2 metrics:

Next Steps

The DiffAb pipeline is available to all Foldexa users today. Try it in the platform dashboard or via the REST API. In Q2 2026, we will release multi-target optimization — simultaneously designing CDRs that bind two different antigens — and integration with experimental binding affinity data from our KAIST partnership.

“We designed 50 CDR variants in 6 minutes. Three of them passed our binding assay. That's 3× our hit rate from phage display, at 1% of the cost.”

— KAIST Bioengineering Lab collaborator