Summary
Generic PLMs outperform antibody LMs on zero-shot prediction of affinity changes. This was observed using ESM (1) and other BERT-based models (2), as well as the autoregressive model like ProGen (3). However, they are worse at predicting intra-family thermostability (4).
Details
Hie et al. concluded this by comparing experimental binding affinity data collected on clones picked using ESM to zero-shot predictions by AbLang and SAPIENS. (1)
The failure of antibody LMs to predict residues with high affinity may be due to their bias toward germline residues.
1.
Hie BL, Shanker VR, Xu D, Bruun TUJ, Weidenbacher PA, Tang S, et al. Efficient evolution of human antibodies from general protein language models. Nature Biotechnology. 2023;42(2):275–83. Available from: https://doi.org/10.1038/s41587-023-01763-2
2.
Li L, Gupta E, Spaeth J, Shing L, Jaimes R, Engelhart E, et al. Machine learning optimization of candidate antibody yields highly diverse sub-nanomolar affinity antibody libraries. Nature Communications. 2023;14(1). Available from: https://doi.org/10.1038/s41467-023-39022-2
3.
Nijkamp E, Ruffolo JA, Weinstein EN, Naik N, Madani A. ProGen2: Exploring the boundaries of protein language models. Cell Systems. 2023;14(11):968-978.e3. Available from: https://doi.org/10.1016/j.cels.2023.10.002
4.
Chungyoun M, Ruffolo J, Gray J. FLAb: Benchmarking deep learning methods for antibody fitness prediction. openRxiv; 2024. Available from: https://doi.org/10.1101/2024.01.13.575504