Methodology transparency
The computational biology behind Genolux
Rosetta energy functions, AlphaFold2 structural predictions, multi-objective CDR design. No black box — every score is interpretable and traces to a specific energy term or database observation.
Methodology
Three computational stages
Stage 01 — Structural input pipeline
AlphaFold2 Fv prediction and RosettaDock interface modeling
Raw antibody sequences enter the pipeline as FASTA. We run each through an AlphaFold2 prediction sub-pipeline configured for the Fv region (VH + VL), producing predicted variable domain structures with full CDR loop resolution. Where the antigen structure is available as PDB, we use it directly. Where only the antigen sequence is known, a separate AlphaFold2 prediction run generates the antigen model.
The predicted antibody Fv is docked against the antigen using RosettaDock protocol 3.0 with local refinement. We sample 1,000 docking poses, apply energy filters, and select the lowest-energy pose for downstream analysis. Contact residues are identified by a 5 Å distance threshold with additional criteria on van der Waals interaction energy and solvation energy terms.
Stage 02 — CDR scoring model
Rosetta REF2015 energy function calibrated on SKEMPI2 affinity data
CDR variant scoring uses the Rosetta REF2015 energy function — a composite of van der Waals (Lennard-Jones), electrostatics (Coulombic with distance-dependent dielectric), solvation (Lazaridis-Karplus), hydrogen bonding, and torsion angle terms. The function is not modified; instead, a lightweight regression layer is trained to predict ΔΔG from Rosetta energy outputs calibrated on SKEMPI2 (version 2.0), which contains 7,085 single-point mutation measurements across 345 protein complexes, with antibody-specific entries weighted at 1.4×.
SAbDab structural data is used to supplement training: 3D structures with known affinities provide direct geometry-affinity pairings that improve calibration in the CDR loop geometry space specifically. Model performance: Pearson r = 0.71 on holdout, MAE < 0.4 kcal/mol.
Stage 03 — Multi-objective Pareto optimization
Selecting candidates that survive both affinity and developability objectives
A candidate that binds at 1 nM but aggregates at 10 mg/mL is not a development candidate — it is a hit with a problem. The Genolux scoring pipeline explicitly models the trade-off between affinity and developability by computing the Pareto front across three objectives: ΔΔG (lower is better), composite developability score (0–1, higher is better), and predicted CHO expression yield tier (low/medium/high). Candidates on the Pareto front — those for which no other candidate dominates all three objectives simultaneously — form the output shortlist.
Training data
Source databases and data vintage
The databases below train and calibrate the Genolux scoring models. Sizes and vintage are listed because the coverage limits of the training data define the coverage limits of the predictions.
SAbDab
4,800+
Structural Antibody Database. Curated antibody-antigen complex structures from PDB. Used for geometry-affinity calibration and CDR loop RMSD benchmarking. Vintage: SAbDab-2024Q1.
SKEMPI2
7,085
Protein–protein interaction free energy changes upon mutation. 7,085 single-point ΔΔG measurements across 345 complexes. Primary calibration source for Rosetta REF2015 regression layer.
OAS
1.5B+
Observed Antibody Space. 1.5 billion natural antibody sequences from B-cell repertoire sequencing studies. Used to train CDR loop length distributions and developability prior models.
Interpretability
Every prediction comes with a decomposition
A rank number is not a justification — it is an instruction to do more work. The Genolux output report is designed to give a scientist the decomposition they need to make a decision and defend it. That means: per-residue contribution scores identifying which CDR positions drive the predicted affinity change versus which are energetically neutral; energy decomposition broken out by term (van der Waals, electrostatics, solvation, hydrogen bonds); and 95% confidence intervals on ΔΔG derived from ensemble scoring across 10 docking poses sampled around the lowest-energy configuration.
The output schema for every top candidate includes: ddg_kcal_mol, ddg_ci_95, residue_scores[], energy_components (vdw, elec, solv, hbond), developability_flags[] with per-flag scores, and pareto_rank.
Known limitations
What we cannot predict yet
Every computational method has a performance envelope. These are the cases where Genolux predictions degrade — knowing them helps you use the tool correctly and plan your experimental follow-up accordingly.
CDR H3 loop prediction accuracy drops for loops > 14 residues
AlphaFold2 CDR H3 loop prediction has median RMSD of 1.4 Å for loops ≤ 12 residues, but degrades to > 3.5 Å for loops > 14 residues. If your antibody has a long H3 loop, treat structural predictions as approximate and prioritize developability scoring over interface geometry.
Affinity predictions have higher error for bispecific formats
The SKEMPI2 calibration set contains almost no bispecific constructs. ΔΔG predictions for bispecifics (CrossMab, knob-into-hole, DART formats) should be treated as directional, not quantitative. We recommend running bispecifics in relative mode (rank variants, do not interpret absolute values).
Developability flags do not replace full biophysical characterization
Genolux developability flags are predictive tools, not measurements. They are designed to eliminate obvious high-risk candidates before synthesis, not to certify that passing candidates are development-ready. Aggregation, viscosity, and expression yield must be confirmed experimentally — we narrow the search space, not the experimental program.
Conformational epitopes require antigen structure input
When an antibody binds a conformational epitope dependent on quaternary protein structure, antigen FASTA input alone is insufficient. The AlphaFold2 antigen prediction may not capture the correct binding-competent conformation. For targets with known quaternary structures (GPCRs, viral surface glycoproteins), submit the holo-state PDB directly.