DiffAbXL

The official implementation behind Exploring Log-Likelihood Scores for Ranking Antibody Sequence Designs, this repo runs a benchmark that ranks antibody sequence designs and publishes a leaderboard of how models score.

DiffAbXL is the official implementation benchmarked in the paper Exploring Log-Likelihood Scores for Ranking Antibody Sequence Designs. The paper went by a different name at first, Benchmarking Generative Models for Antibody Design, and the change was meant to highlight what the work actually contributes. DiffAbXL itself is a re-implementation of the earlier DiffAb work, and the repo links to both the paper and the original code.

A leaderboard ranks models by average Spearman correlation across five target datasets. A model that does not show statistically significant correlation gets a score of zero. Variants such as DiffAbXL-A-DN and DiffAbXL-H3-DN appear with correlation values on datasets named Absci HER2, Nature, AZ Target-2, and HEL, and the average is simply the mean over the five sets.

To make benchmarking easier, the repo recommends a model interface written as a Python method inside a class that fits into the evaluation pipeline. The method takes inputs like antibody sequences and returns a dictionary with log-likelihood scores, which is how sequences get ranked by predicted binding affinity. Other metrics, such as RMSD or pAE, can be returned alongside.

Training is driven by a single configuration file, sabdab.yaml, which is where parameters get changed. Weights and Biases can track experiments, though it is off by default and only turns on when an option in the config file is changed, so people who do not need it can leave it alone. Benchmarking datasets and their licenses live in the benchmarking datasets folder, and the original Absci sets are linked from the Absci IgDesign and Absci Her2 repositories.

• Source 1 · Project repository
• Source 2 · README