Hi authors,
Thanks for your earlier reply. A few quick follow-ups:
-
After using multiple sequences and y_test_esm2_dataset.csv as the single y file, the pipeline now works properly.
-
When all placeholder values are identical (e.g., all 0), construct_dataset.py crashes due to duplicate dict keys; mixing in a different value (e.g., 1) resolves it. Is this expected behavior?
-
On my test set, the predictions appear inverted: three PETase variants labeled soluble (1) received low scores (~0.17–0.27), while two sequences labeled insoluble (0)—one designed by DeepSeek as an insoluble variant, and CsoS1E_0, which is experimentally insoluble yet highly hydrophilic—received high scores (0.52, 0.83). ISPETase is known to have poor solubility, so the low scores may be correct. These confusing results may offer some clues for investigating the unexpected prediction outcomes of ProtSATT.
Thanks!
test.fasta.txt
Hi authors,
Thanks for your earlier reply. A few quick follow-ups:
After using multiple sequences and y_test_esm2_dataset.csv as the single y file, the pipeline now works properly.
When all placeholder values are identical (e.g., all 0), construct_dataset.py crashes due to duplicate dict keys; mixing in a different value (e.g., 1) resolves it. Is this expected behavior?
On my test set, the predictions appear inverted: three PETase variants labeled soluble (1) received low scores (~0.17–0.27), while two sequences labeled insoluble (0)—one designed by DeepSeek as an insoluble variant, and CsoS1E_0, which is experimentally insoluble yet highly hydrophilic—received high scores (0.52, 0.83). ISPETase is known to have poor solubility, so the low scores may be correct. These confusing results may offer some clues for investigating the unexpected prediction outcomes of ProtSATT.
Thanks!
test.fasta.txt