Benchmarking antibody discovery fidelity and reproducibility with an assay-locked residue fidelity index

Reproducibility in antibody discovery is undermined by dropout, paratope degeneracy, and immunogen imprinting. Although guidelines exist, the field lacks a shared, assay-locked metric to benchmark these liabilities across discovery pipelines. We propose the Residue Fidelity Index (RFI), a discovery-stage, within-assay comparative framework that consolidates fidelity liabilities into a single score. Assay-locked means RFI is comparable only within a fully specified assay configuration, uses fixed normalization and weights, and requires co-reporting of component metrics to expose drivers of the composite. RFI is not intended as a universal or optimal standard, but as one implementable example meeting basic benchmarking requirements: defined components, fixed assay context, pre-specified weights, and primary-data disclosure. RFI is reported alongside its components (D, G, I) to visualize reproducibility in parallel with affinity, humanization, and yield. To demonstrate feasibility, we applied RFI in a simulated Epivolve testbed derived from multiple studies, embedding MILKSHAKE (context retention) and Sundae (residue discrimination) validation modules into discovery. Across three campaigns (n = 25 clones), antibodies meeting provisional thresholds (dropout ≤10%, degeneracy ≤0.10, no imprinting) yielded RFI values from 0.610 to 0.982 (mean ± SD 0.922 ± 0.103; I = 1 for clones C04 and C20). These results show how fidelity can be consolidated to prioritize candidates before application-specific cellular validation. Despite limited scope and proprietary immunogens, this work provides a proof of concept for residue-level benchmarking, with per-clone primary values and RFI reported in Supplementary Table S1 for auditability.
Residue Fidelity Index (RFI) consolidates dropout (D), degeneracy (G), and immunogen imprinting (I) comparative fidelity score. We illustrate the framework using simulated data from an Epivolve testbed, which integrates MILKSHAKE and Sundae assays into discovery, co-reporting D, G, I and the fixed weight set used to compute RFI. We report means; replicate-level CIs are not reported for this simulated cohort.