ProtBERT Embeddings
ProtBERT embeddings produced the strongest and most balanced supervised performance. The best ProtBERT shallow MLP configuration reached about 0.704 F1 and 0.748 ROC-AUC, indicating moderate rather than strong discrimination. ProtBERT embe…
1 sources - 5 claims
ProtBERT embeddings produced the strongest and most balanced supervised performance. The best ProtBERT shallow MLP configuration reached about 0.704 F1 and 0.748 ROC-AUC, indicating moderate rather than strong discrimination. ProtBERT embeddings were extracted with Rostlab/prot_bert in inference mode without fine-tuning. Mean-pooled last hidden layer vectors produced 1024-dimensional fixed-length ProtBERT representations. Frozen ProtBERT limits conclusions about what fine-tuned protein language models might achieve.