Request: Add documentation for training

[a-ill]

Could you add documentation on how to train the models? Specifically:

• What format should the input files have?
• What command is used to run training?

I see training code in the repository, but there are no explanations on how it works.

[chenghui03]

I attempted to train the model, although the performance did not reach the results reported in the paper (in fact, it is substantially worse).
Below is the code I used:

from lib.PTModels import train_model

train_model(
    input_data='data/raw/phos_st_training.tsv',
    test_file='data/raw/phos_st_testing.tsv',
    db='data/raw/DeepMVP/swiss_prot_human_20190214_processed.fasta',
    out_dir='./mytrain',
    peptide_length=28*2+1,
    p_model=None,
    model=model,
)

The input_data and test_file come from:
https://deepmvp.ptmax.org/database/train_test/phosphorylation_sty/phosphorylation_st_train.tsv
https://deepmvp.ptmax.org/database/train_test/phosphorylation_sty/phosphorylation_st_test.tsv

The db file must be preprocessed; only the UniProt ID portion in the FASTA headers should be retained:
https://deepmvp.ptmax.org/download/swiss_prot_human_20190214.fasta

The model is taken from the GitHub repository:
https://deepmvp.ptmax.org/download/models.tar.gz

Below is my training output. It is puzzling because the performance is far below what the paper reports, even though the architecture and dataset are the same. If you have successfully reproduced the results, I would appreciate it if you could share your training logs.

5155/5155 [==============================] - ETA: 0s - loss: 0.2907 - accuracy: 0.8761
Epoch 48: val_accuracy did not improve from 0.83340

If any part of the file formats is unclear, feel free to ask for clarification.

[Devin-jun]

After checking the code, it seems that DeepMVP’s default inference strategy is:
• Using all models (0–9) for prediction
• Combining outputs via combine_rts() using mean averaging, with remove_outlier=True enabled

This means inference runs as an ensemble by default rather than using a single model.

[chenghui03]

Yes, I noticed that ensemble-based prediction strategy as well, and you’re absolutely right. However, the performance of the pretrained models provided in the paper is still much higher than what I am getting from my own training. So perhaps you could share your training results or process, just to help confirm that I haven’t made any mistake or missed something.