bioRxiv preprint
November 2025
Multi-Objective Pareto Optimization of CDR Sequences for Simultaneous Affinity and Developability: A Computational Framework
Voss, E., Chen, M., Nair, P., Richter, T., Petrova, A.
We present a computational framework for simultaneously optimizing antibody CDR sequences across competing objectives: binding affinity (ΔΔG), developability flags, and predicted expression yield. Applying Pareto-front selection to candidate pools of up to 500 variants demonstrates that affinity-only optimization systematically produces candidates with elevated aggregation propensity and deamidation risk. The Pareto approach recovers candidates with > 85% of the affinity improvement while reducing developability flag rates by 3.1-fold on a benchmark set of 120 antibody-antigen pairs. Note: this is a synthetic preprint for demonstration purposes; results are not experimentally validated.
View preprint
bioRxiv preprint
August 2025
Benchmarking CDR Loop Structure Prediction on the SAbDab-2024 Dataset: AlphaFold2 Performance Across Loop Length Classes
Petrova, A., Chen, M., Voss, E.
Systematic evaluation of AlphaFold2 CDR loop prediction accuracy on 4,800+ antibody-antigen complex structures from the Structural Antibody Database (SAbDab-2024). We report backbone RMSD stratified by loop type (H1, H2, H3, L1, L2, L3) and loop length, identifying a performance inflection point at CDR H3 length = 14 residues beyond which median RMSD exceeds 3.5 Å. Benchmark code and evaluation datasets released at [placeholder GitHub]. Note: this is a synthetic preprint for demonstration purposes.
View preprint
bioRxiv preprint
May 2025
Physicochemical Developability Flags as Early-Stage Attrition Predictors: A Retrospective Analysis of 200 Antibody Drug Candidates
Nair, P., Voss, E., Richter, T.
Retrospective analysis of 200 synthetic antibody drug candidate profiles using seven computational developability flags (aggregation, viscosity, deamidation, oxidation, charge patches, FcRn half-life, CHO yield). We demonstrate that sequences with two or more high-risk flags have a 3.7-fold higher rate of late-stage biophysical failure in a held-out validation set. Aggregation propensity and CDR deamidation risk are the two most predictive individual flags. Note: this is a synthetic preprint based on simulated data.
View preprint
bioRxiv preprint
February 2025
Calibrating Rosetta REF2015 ΔΔG Predictions for Antibody–Antigen Interfaces Using SKEMPI2 Antibody Subsets
Chen, M., Voss, E., Petrova, A.
The Rosetta REF2015 energy function applied directly to protein-protein interactions yields ΔΔG predictions with MAE of 0.92 kcal/mol across SKEMPI2. We demonstrate that training a lightweight regression correction layer specifically on antibody-specific SKEMPI2 entries (n=1,240) reduces MAE to 0.38 kcal/mol on a held-out antibody-antigen subset. The calibration method, pre-trained weights, and evaluation scripts are released openly. Note: this is a synthetic preprint for demonstration purposes.
View preprint