Summary
Subnetworks within protein language models encode specific protein families (1), and individual competencies are stored in distinct parts of the neural network. This might be why larger PLMs are able to cluster protein families are finer levels, as their capacity is simply larger.
Figures
Ref (1)
1.
Vinod R, Amini AP, Crawford L, Yang KK. Trainable subnetworks reveal insights into structure knowledge organization in protein language models. PLOS Computational Biology. 2026;22(2):e1013925. Available from: https://doi.org/10.1371/journal.pcbi.1013925