How to train a new model?

[biocc]

I retrained a Casanovo model (v5.1.0) using the following command:

CUDA_VISIBLE_DEVICES=1 nohup casanovo train \
  --validation_peak_path /mnt/data_nvme1/data/MassiveIVE/3mod/Human_HCD_Spectral_Library_only3mod_casanovo5_valid5.mgf \
  /mnt/data_nvme1/data/MassiveIVE/3mod/Human_HCD_Spectral_Library_only3mod_casanovo5_train95.mgf \
  --config /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/casanovo.yaml \
  --output_dir /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/train \
  > /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/run.log 2>&1 &

After training finished, I ran de novo sequencing with the trained checkpoint:

CUDA_VISIBLE_DEVICES=1 nohup casanovo sequence \
  /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/denovo/sample_preprocessed_spectra.mgf \
  -c /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/denovo/casanovo.yaml \
  -m /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/train/best.ckpt \
  -d /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/denovo/result \
  > /mnt/data_nvme1/models/casanovo5.1.0/2m/3mod/denovo/run.log 2>&1 &

However, the de novo results look very strange (see attached screenshot). Could you please help us understand what might be going wrong?

casanovo.yaml

[wsnoble]

Can you provide us with your training data (or a small subset thereof) and the log file from the training run?

[biocc]

Can you provide us with your training data (or a small subset thereof) and the log file from the training run?

https://drive.google.com/drive/folders/1uH9MOCpmFPx7rHrwxXr13054M5HJaJTK?usp=sharing

[wsnoble]

The log file does not show any obvious problem. Did you check whether all of the results in your de novo sequencing output looked strange? Sometimes the very first spectra in a run are junk. As a sanity check, try doing sequencing on your training set and make sure it works there. If you still can't figure it out, maybe send us your test set. And also try the released model on your test set.

[biocc]

The log file does not show any obvious problem. Did you check whether all of the results in your de novo sequencing output looked strange? Sometimes the very first spectra in a run are junk. As a sanity check, try doing sequencing on your training set and make sure it works there. If you still can't figure it out, maybe send us your test set. And also try the released model on your test set.

I reviewed the de novo sequencing output, and all test results exhibit abnormalities consistent with those shown in the figure I previously shared. However, when I apply the released model to the same test set, the results appear normal. Consequently, I suspect that the issue lies within my training process rather than in the test data or the spectra themselves.

[Lilferrit]

I'm not sure how large your train set is, but looking at the log file quite a bit of the spectra are being skipped:

2025-11-28 10:56:40,154 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 140204 spectra with invalid information.Last error was: 
 Insufficient number of peaks
2025-11-28 10:56:40,154 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 140204 spectra with invalid information.Last error was: 
 Insufficient number of peaks
2025-11-28 10:56:59,603 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 7406 spectra with invalid information.Last error was: 
 Invalid precursor charge.
2025-11-28 10:56:59,603 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 7406 spectra with invalid information.Last error was: 
 Invalid precursor charge.

If this is the (vast) majority of your training set it's totally possible that Casanovo isn't seeing enough data to learn anything useful during training.

[biocc]

I'm not sure how large your train set is, but looking at the log file quite a bit of the spectra are being skipped:

2025-11-28 10:56:40,154 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 140204 spectra with invalid information.Last error was: 
 Insufficient number of peaks
2025-11-28 10:56:40,154 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 140204 spectra with invalid information.Last error was: 
 Insufficient number of peaks
2025-11-28 10:56:59,603 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 7406 spectra with invalid information.Last error was: 
 Invalid precursor charge.
2025-11-28 10:56:59,603 WARNING [py.warnings/MainProcess] warnings._showwarnmsg : UserWarning: Skipped 7406 spectra with invalid information.Last error was: 
 Invalid precursor charge.

If this is the (vast) majority of your training set it's totally possible that Casanovo isn't seeing enough data to learn anything useful during training.

1818143 spectras in the training set, 95722 spectras in the valid set. (data link: https://drive.google.com/drive/folders/1uH9MOCpmFPx7rHrwxXr13054M5HJaJTK?usp=sharing)

[bittremieux]

Your peptide precision also remains 0 throughout the entire training run, so something is clearly going wrong here. I can't immediately see the issue though, the data and settings look fine.

[biocc]

Your peptide precision also remains 0 throughout the entire training run, so something is clearly going wrong here. I can't immediately see the issue though, the data and settings look fine.

Could you please provide me with a demonstration of the training data?

[bittremieux]

The peptide precision being 0 was actually a bug that will be fixed in an upcoming release (#546). So that comment is not applicable.

I was looking at your data and trying to figure out what could be the problem, but I didn't see anything obviously wrong and haven't been able to find the issue yet. I'll try to do some training with your data to see if I can reproduce the issue.

[biocc]

The peptide precision being 0 was actually a bug that will be fixed in an upcoming release (#546). So that comment is not applicable.

I was looking at your data and trying to figure out what could be the problem, but I didn't see anything obviously wrong and haven't been able to find the issue yet. I'll try to do some training with your data to see if I can reproduce the issue.

Have you reproduced this issue?

[wsnoble]

I don't know if Wout reproduced it, but I just started my own version of your training run. However, in doing that I noticed that you are reporting that you get weird sequencing results on the file sample_preprocessed_spectra.mgf. Is it possible that there is something awry with that file? What happens if, as a sanity check, you run the casanovo sequence command on your validation set (Human_HCD_Spectral_Library_only3mod_casanovo5_valid5.mgf)?

[wsnoble]

OK, this is weird. I tried training on your dataset using this command line:

casanovo train \
  --config casanovo.yaml \
  --output_dir . \
  --output_root casa_balanced \
  --validation_peak_path /tmp/6289856.1.noble-long.q/Human_HCD_Spectral_Library_only3mod_casanovo5_valid5.mgf \
/tmp/6289856.1.noble-long.q/Human_HCD_Spectral_Library_only3mod_casanovo5_train95.mgf

And I got this traceback error:

File "/net/noble/vol1/home/noble/miniconda3/envs/casanovo/lib/python3.13/site-packages/lance/dataset.py", line 2644, in write_dataset
    inner_ds = _write_dataset(reader, uri, params)
OSError: LanceError(IO): Generic LocalFileSystem error: Unable to create dir /mnt/data_nvme1/wenbin/models/casanovo5.1.0/2m/3mod/model/train.lance/data: Permission denied (os error 13), /home/runner/work/lance/lance/rust/lance-io/src/encodings/plain.rs:123:17

It seems like Casanovo is trying to write to mnt/data_nvme1/wenbin/models/

Full stderr output is attached.

job.stderr.txt

[wsnoble]

FWIW, the same error appears on the dev branch.

[bittremieux]

I trained a model on the provided data and similarly got incorrect predictions. I haven't been able to dig into the model functioning yet to understand where this bug comes from.

I noticed that you are reporting that you get weird sequencing results on the file sample_preprocessed_spectra.mgf. Is it possible that there is something awry with that file?

No, that's the standard example data in this repository. With our own weights, predictions on these spectra are correct, so it's not a problem with the data. The issue is with the training / the resulting weights, which don't give good predictions.

It seems like Casanovo is trying to write to mnt/data_nvme1/wenbin/models/

This doesn't make sense, and is something completely unrelated. For some reason Casanovo is trying to create data in a directory that you don't have access to. Can it have something to do with how you copied data over (e.g. symlinks)?

In any case, this has nothing to do with the original issue, so if you can't track it down, I suggest opening a new issue to not have multiple unrelated discussions in this thread.

[wsnoble]

Haha, never mind. I didn't notice that the provided yml file had the lance directory encoded in it.