Four modules, one pipeline

The Genolux Antibody Design Platform

Interface prediction, CDR scoring, developability assessment, candidate ranking — four modules running in sequence from FASTA input to a Pareto-optimal shortlist.

Module 01 — Interface Prediction

CDR–Antigen Contact Mapping

Submit an antibody Fv sequence and a target antigen (FASTA or PDB). Genolux runs AlphaFold2 variable-domain prediction on the antibody, then docks the predicted Fv against the antigen using RosettaDock. The result: per-residue contact maps identifying which CDR positions contribute to binding, including buried surface area, hydrogen bond geometry, and van der Waals contact energy at the paratope–epitope interface.

AlphaFold2 Fv prediction RosettaDock protocol Per-residue energy decomposition Buried surface area (BSA)

Module 02 — CDR Optimization

Affinity Scoring and Variant Ranking

Enumerate CDR loop sequence variants — single-point mutations, double mutants, or combinatorial CDR3 libraries up to 500 sequences per run. Each variant is scored for predicted binding affinity (ΔΔG) using a Rosetta REF2015 energy function trained on SAbDab structural data and SKEMPI2 binding affinity measurements (30,000+ experimentally measured interaction energies). Output: ranked variant list with per-variant Kd prediction, confidence interval, and residue-level energy contribution.

ΔΔG prediction (Rosetta REF2015) SAbDab + SKEMPI2 training data Kd MAE < 0.4 kcal/mol Up to 500 variants/run

Module 03 — Developability Assessment

Seven Physicochemical Flags

A candidate that binds well in silico can still fail in the clinic if it aggregates, has poor viscosity at high concentration, or misfolds during CHO expression. Genolux runs every candidate through a 7-flag developability screen covering the physicochemical properties that correlate with late-stage attrition. Each flag produces a binary pass/warn and a continuous score you can filter on.

Aggregation propensity (SAP score) Viscosity index Deamidation risk (NG/NS motifs in CDRs) Oxidation-prone sites (Met/Trp/His in CDRs) Charge patch analysis (polyreactivity) Predicted half-life (FcRn binding score) CHO expression yield prediction

Module 04 — Manufacturability Ranking

Pareto-Optimal Candidate Shortlist

Affinity and developability are competing objectives — optimizing one often degrades the other. Genolux computes the Pareto front across binding affinity (ΔΔG), developability composite score, and predicted CHO expression yield. The output is a Pareto-optimal shortlist of candidates that represent the true trade-off frontier: sequences that are optimal in a practical sense, not just highest-binding. Delivered as ranked CSV, annotated PDB files, and JSON API response.

Multi-objective Pareto optimization Ranked CSV + annotated PDB JSON API response Affinity × developability × yield

Inputs & Outputs

Accepted input formats and output deliverables

Accepted inputs

  • Antibody sequence in FASTA format (Fv or full-length)
  • Antigen sequence (FASTA) or 3D structure (PDB)
  • Homology model (.pdb) for antigen when crystal structure unavailable
  • Batch upload: up to 500 CDR variant sequences per run
  • Pre-computed complex structure (.pdb) for scoring-only mode

Output deliverables

  • Ranked variant CSV with ΔΔG, 7-flag scores, confidence intervals
  • Annotated PDB files for top 10 Pareto-optimal candidates
  • JSON API response (async, poll by job ID)
  • Per-residue energy decomposition report (PDF)
  • Interactive interface map visualization (embedded SVG)

Accuracy benchmarks

Validated against SAbDab holdout set

Performance evaluated on 418 antibody-antigen complexes held out from training. We report what the numbers actually are, including where accuracy degrades. Full methodology and evaluation code in Publications →

< 0.4

Kd prediction MAE

Mean absolute error in ΔΔG (kcal/mol) on SAbDab-2024 holdout set. Measured against SPR-confirmed affinities.

1.4 Å

CDR H3 loop RMSD

Median backbone RMSD for CDR H3 loops ≤ 12 residues. Accuracy drops for loops > 14 residues (see limitations).

86%

Developability flag precision

Precision of high-risk flag classification across 7 flags vs experimentally confirmed biophysical failure modes in test set.

Run your first antibody through the platform.

Early access is open. Submit a sequence and target antigen — results in under 10 minutes.