-
|
First of all, congratulations on your great work and thank you very much for sharing your models with the research community. I have a question about the performance of ESM-1b and ESM-1v models. In the Meier et al. 2021 preprint, the predictive performance is better using UniRef90 compared with UniRef50. However, in your github "Comparison to related works" table, ESM-1b using UniRef 50 is better than ESM-1v using UniRef 90. Could you advise which model is better if we would like to use your models for downstream predictions. Thank you! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hi @fancy1228, we've found that performance depends on the downstream task. If you're interested in function prediction, use ESM-1v (trained on Uniref90). If you're interested in structure prediction, use ESM-1b (trained on Uniref50). For other applications, you may want to experiment with representations from both models. |
Beta Was this translation helpful? Give feedback.
Hi @fancy1228, we've found that performance depends on the downstream task. If you're interested in function prediction, use ESM-1v (trained on Uniref90). If you're interested in structure prediction, use ESM-1b (trained on Uniref50). For other applications, you may want to experiment with representations from both models.