About the DeepRT+ calibration

[Chaohk]

Hi, could you please tell me that how the DeepRT+ do the calibration using a certain ratio of the test-group peptides after pretraining with the big data?

Thanks.

[horsepurve]

Hi! For the calibration from a pretrained model (run1) to another LC run (run2), please:

1. Select a certain ratio of peptides from run2 as the calibration peptides (better with high identification confidence, and span across the whole LC run). Save as a txt file and assign its file path to "train_path" in "config.py".
2. Set "pretrain_path" to the file path of your pretrained model, set "dict_path" and "max_length" to be the same with those of the pretrained model (see tutorial for details).
3. Run by "python capsule_network_emb.py" and see if the prediction is improved and better than not using a model pretrained from bigger data.

[Chaohk]

Thank you for telling me the usage of the calibration function. Could I know what algorithm and methods are using in the RT calibration. Like piecewise linear regression?

[horsepurve]

The method is transfer learning. The model is first pre-trained on data1 from LC run1, then fine-tuned on data2 from LC run2.

The only difference between using/not using this pretraining & fine-tuning scheme is that: without it, we train a deep neural network (DNN) from scratch whose weights are randomly initialized; whereas pretraining provides a well-educated guess for the weight of DNN so that DNN can converge to a better minimum faster by stochastic gradient descent.

This pretraining & fine-tuning routine is extremely widely adopted in Computer Vision, Natural Language Processing, etc. In LCMS, our work is perhaps the first that uses this to integrate/calibrate multiple runs across different LCs/experiments/labs.

[Chaohk]

Fully understand. Thanks a lot!