Multi-Scale Machine Learning for Antibody-Antigen Binding Affinity Prediction
One-line Summary
A reported multi-scale ML framework for antibody-antigen binding affinity prediction using deep mutational scanning and structural features. Detailed metrics should remain to verify until the paper is read directly.
Why It Matters
This paper may be relevant to Structure-based modeling and binding affinity. Its useful angle is not only model performance, but how antibody-antigen binding evaluation handles mutation units, pathogen or antigen holdouts, confidence filtering, and theoretical performance ceilings.
Key Points
- Reported multi-scale descriptor set spanning sequence, structure, physicochemical, and learned representation features
- Reported use of deep mutational scanning and structural features
- Reported confidence-stratified prediction; exact coverage/metric values to verify
- Reported cross-pathogen or cross-antigen transfer limitations; split protocol to verify
- Boltzmann ceiling analysis appears to be a central evaluation idea; formula and assumptions need direct reading
Reading Questions
- What is the variance across pathogens in the MCC distribution?
- Does the confidence stratification primarily reduce false positives or increase true positives?
- How does the ESM-2 embedding modality compare to structural features alone?
- What would be needed to close the gap from 79.1% to 100% of the Boltzmann ceiling?
- What exactly is the prediction unit: mutation, antibody-antigen pair, antigen family, or assay row?
- Are all measurements from one DMS experiment kept on the same side of each split?
Artifact Availability
| Artifact | Status | Notes |
|---|---|---|
| Paper | found | bioRxiv preprint page |
| Code | to verify | Search linked repository or supplementary materials |
| Data | to verify | AbAgym/DMS source and filtering rules |
| Splits | to verify | Need LOCO-DMS or holdout definition |
| Weights | not applicable | Classical ML or ensemble status to verify |
| Environment | to verify | Feature extraction versions unknown |
Related Notes
- Binding affinity
- Protein representation
- Protein family split
- Boltzmann ceiling analysis
- Selective prediction
- Scoring function
- Artifact availability
Metadata
- DOI: verified on bioRxiv
- Venue: bioRxiv
- Posted: 2026-06-23
- Authors: S. Sivasubramani