Hi there,
I recently used transfer learning to improve my spectral libraries for ArgC peptides, which increased identifications by >50%. I went on to also try transfer learning for AspN and GluC peptides, but the resulting IDs were overly inflated; results for GluC exceeded more than 250k total identified peptides, which would be lovely, but not realistic. My colleague pointed me to the figures from the "/transfer/quant/{file_name}/figures" folder that seem to suggest some over-fitting for AspN and especially for GluC. The attached files are for the seemingly successful ArgC transfer learning (frd_17.pdf) and the failed ones for AspN (fdr_20.pdf) and GluC (fdr_21.pdf). Given the success for ArgC, I would like to try the other proteases again if there is a specific bug responsible for the over-fitting behavior. I am happy to provide the complete results folder if helpful.
fdr_21.pdf
fdr_20.pdf
fdr_17.pdf
Hi there,
I recently used transfer learning to improve my spectral libraries for ArgC peptides, which increased identifications by >50%. I went on to also try transfer learning for AspN and GluC peptides, but the resulting IDs were overly inflated; results for GluC exceeded more than 250k total identified peptides, which would be lovely, but not realistic. My colleague pointed me to the figures from the "/transfer/quant/{file_name}/figures" folder that seem to suggest some over-fitting for AspN and especially for GluC. The attached files are for the seemingly successful ArgC transfer learning (frd_17.pdf) and the failed ones for AspN (fdr_20.pdf) and GluC (fdr_21.pdf). Given the success for ArgC, I would like to try the other proteases again if there is a specific bug responsible for the over-fitting behavior. I am happy to provide the complete results folder if helpful.
fdr_21.pdf
fdr_20.pdf
fdr_17.pdf