A Critical Commentary on "Atomically accurate de novo design of antibodies with RFdiffusion" by Bennett et al., Nature 2025; doi: 10.1038/s41586-025-09721-5

This repository contains a comprehensive, figure-by-figure scientific commentary on the article “Atomically accurate de novo design of antibodies with RFdiffusion” by Bennett et al. (Nature, 2025; doi:10.1038/s41586-025-09721-5). The commentary, authored by Mengxi Zhu and Shu-Feng Zhou, provides a critical and systematic evaluation of the methodology, data interpretation, experimental validation, and claims presented in the original publication. The purpose of this work is to promote transparent, rigorous, and evidence-driven scientific discourse in the field of computational protein design and antibody engineering.

The commentary examines every main figure, Extended Data Figure, and Supplementary Figure in the Bennett et al. paper. Particular attention is given to the analytical methods, the validity of structural predictions, the robustness of experimental assays, and the reproducibility claims underlying RFdiffusion-based antibody design. Although the original paper reports “atomically accurate” de novo generated antibodies, our in-depth analysis uncovers substantial methodological gaps, over-interpretation of in silico outputs, limited experimental depth, selective reporting of favorable results, and insufficient benchmarking against contemporary protein-design methodologies.

Major critique themes covered in this repository include:

1. Structural Accuracy Claims — A detailed reassessment of backbone and side-chain alignment metrics, highlighting inadequacies in RMSD-focused evaluation, ambiguity in the superposition strategy, and the absence of rotamer-level validation.

2. Biophysical and Functional Validation — Systematic discussion of SPR/BLI binding assays, expression and stability measurements, and structural characterization. The commentary notes the extremely small experimental sample size, absence of replicates, and lack of raw sensorgram and chromatographic data.

3. Generative Model Transparency and Reproducibility — Examination of RFdiffusion’s training data, architectural modifications, hyperparameters, and the undisclosed post-processing pipelines involving Rosetta and AlphaFold2. The lack of public model weights, inference scripts, and training logs raises concerns about reproducibility.

4. Biological Relevance and Antigen Selection — Critical analysis of antigen simplifications, missing glycosylation, oligomerization effects, and epitope validation. The commentary emphasizes the gap between computational designs and physiologically relevant antigen contexts.

5. Diversity, Generalization, and Failure Modes — Investigation of sampling diversity, CDR-H3 length distributions, homology searches, and structural clustering. The commentary highlights evidence of mode collapse, overfitting to training data, and the absence of reported failed designs.

6. Figure Integrity and Data Interpretation — Close inspection of consistency across structural overlays, interface analyses, and sequence alignments, noting unusual uniformity and potential redundancy in several Supplementary Figures.

This commentary aims to assist researchers, reviewers, and developers seeking a deeper, more objective understanding of the capabilities and limitations of generative diffusion models in antibody design. The included analysis prioritizes scientific rigor, transparent critique, and constructive evaluation of emerging computational protein-design technologies. All content in this repository is original, independently produced, and based solely on publicly available information from the Bennett et al. article.