How is it possible to predict RT when DeepLC doesnt know the LC gradient

[candidaorelmex]

Hi,

DeepLC enjoys a good reputation which is why I'd like to use it to predict retention times of peptides with with a specific gradient. Unfortunately I cannot provide calibration peptides with that gradient.

When I feed the DeepLC streamlit GUI my peptides to predict it returns a table with valuesin the "predicted_tr" column between 1 and 10.

1. how do I interpret this value?
2. how can DeepLC predict the retention time without knowing anything about the LC gradient I am using? (with he simple scenario that my peptides elute somewhere between 0% and 100% ACN)

Thank you for your help!

[RobbinBouwmeester]

Hi candidaorelmex,

Unfortunately without calibration peptides DeepLC is not able to predict exact retention times. As you point out without knowing the gradient (and a multitude of other parameters of your LC setup) DeepLC cannot predict actual retention times. You can still predict "uncalibrated retention times".

These retention times are based on the dataset that the model was originally trained on, these predictions were normalized in a range between [0,10]. So that means that these normalized values correspond to a retention time in a specific experiment and the %B in that experiment. With calibration, these values are mapped to the retention times of your specific LC setup and gradient. The major assumption there is that the elution order is conserved (which for proteomics RP is a fair assumption) and we can directly map between the [0,10] values to the calibration retention times.

If you let me know the application I will likely be able to help you to implement DeepLC effectively.

Hope that helps!

Kind regards,

Robbin

[RobbinBouwmeester]

FYI, most of the times calibration peptides are chosen from the most confident identifications. So you can use the peptides you are already very certain of their identification, to evaluate less confident identifications.

[RobbinBouwmeester]

I will close this issue for now, but feel free to re-open it if you need any further help.

[candidaorelmex]

Dear Robbin, thanks a lot for your elaobarate answer.

We want to compare weather the it makes a big difference to us if we use calibration peptides (defined from some other run that ran on the column) vs just using the uncalibrated DeepLC output. We then saw that the uncalibrated output is hard to interpret without more details.

What was the %B used for the runs used for the development of the prediction model=

Do you have a couple of examples how to interpret the values? E.g., peptide X has tr = 5.5 in the unpredicted model, corresponding to Y %ACN in the chormatography gradient?

Thanks for your help!

Best, Ilja