Zero ms2 predictions with transfer learning function

[yangkl96]

Describe the bug
Make sure your bug is not addressed in the troubleshooting section or in previous issues. If not, provide a clear and concise description of what the bug is.

To Reproduce
Steps to reproduce the behavior:
I installed the developer version of peptdeep with pip install -e ".[stable,development-stable]". I tried peptdeep.pipeline_api.transfer_learn() and got prediction matrices of all zeros. I confirmed this behavior with multiple speclib.tsv and pepxml+mzml files I loaded into global_settings.

Expected behavior
Non-zero predictions

Logs

2025-07-25 17:48:11> [PeptDeep] Running train task ...
2025-07-25 17:48:14> Loading PSMs and extracting fragments ...
100%|█████████████████████████████████████████████████████████████████████████████████| 27848/27848 [00:11<00:00, 2421.25it/s]
Empty string is detected in the `mobility` column, fill with 0.0
2025-07-25 17:48:33> Loaded 27848 PSMs for training and testing
2025-07-25 17:48:33> Training RT model ...
2025-07-25 17:48:33> 1224 PSMs for RT model training/transfer learning
2025-07-25 17:48:33> Testing pretrained RT model:
   R_square         R     slope  intercept  test_num
0  0.906195  0.951943  0.962368   0.231975       200
2025-07-25 17:48:33> Training with fixed sequence length: 0
2025-07-25 17:48:33> Testing refined RT model:
   R_square         R     slope  intercept  test_num
0  0.906195  0.951943  0.962368   0.231975       200
2025-07-25 17:48:33> Finished training RT model
2025-07-25 17:48:33> Training MS2 model ...
2025-07-25 17:48:37> Testing pretrained MS2 model on testing df:
         PCC    COS     SA         SPC
count  200.0  200.0  200.0  200.000000
mean     0.0    0.0    0.0    0.051298
std      0.0    0.0    0.0    0.215521
min      0.0    0.0    0.0   -0.454129
25%      0.0    0.0    0.0   -0.085351
50%      0.0    0.0    0.0    0.046855
75%      0.0    0.0    0.0    0.214629
max      0.0    0.0    0.0    0.652426
>0.90    0.0    0.0    0.0    0.000000
>0.75    0.0    0.0    0.0    0.000000
2025-07-25 17:48:37> 1692 PSMs for MS2 model training/transfer learning
2025-07-25 17:48:37> Training with fixed sequence length: 0
[Training] Epoch=1, lr=4e-05, loss=0.03749007191509009
[Training] Epoch=2, lr=6e-05, loss=0.03095195721834898
[Training] Epoch=3, lr=8e-05, loss=0.023990619213630757
[Training] Epoch=4, lr=0.0001, loss=0.02098238399873177
[Training] Epoch=5, lr=0.0001, loss=0.01814938588067889
[Training] Epoch=6, lr=9.045084971874738e-05, loss=0.017319106155385573
[Training] Epoch=7, lr=6.545084971874738e-05, loss=0.015489165981610617
[Training] Epoch=8, lr=3.4549150281252636e-05, loss=0.015003920874247949
[Training] Epoch=9, lr=9.549150281252633e-06, loss=0.014223656989634037
[Training] Epoch=10, lr=1.0000000000000002e-14, loss=0.013656104945888122
2025-07-25 17:48:42> Testing refined MS2 model on training df:
          PCC     COS      SA          SPC
count  1692.0  1692.0  1692.0  1692.000000
mean      0.0     0.0     0.0     0.096597
std       0.0     0.0     0.0     0.211188
min       0.0     0.0     0.0    -0.483608
25%       0.0     0.0     0.0    -0.042768
50%       0.0     0.0     0.0     0.088416
75%       0.0     0.0     0.0     0.227828
max       0.0     0.0     0.0     0.735943
>0.90     0.0     0.0     0.0     0.000000
>0.75     0.0     0.0     0.0     0.000000
2025-07-25 17:48:42> Testing refined MS2 model on testing df:
         PCC    COS     SA         SPC
count  200.0  200.0  200.0  200.000000
mean     0.0    0.0    0.0    0.051298
std      0.0    0.0    0.0    0.215521
min      0.0    0.0    0.0   -0.454129
25%      0.0    0.0    0.0   -0.085351
50%      0.0    0.0    0.0    0.046855
75%      0.0    0.0    0.0    0.214629
max      0.0    0.0    0.0    0.652426
>0.90    0.0    0.0    0.0    0.000000
>0.75    0.0    0.0    0.0    0.000000
2025-07-25 17:48:42> Finished training MS2 model

Screenshots
If applicable, add screenshots to help explain your problem.

Version (please complete the following information):

• Installation Type [e.g. One-Click Installer / Pip / Developer]
• If no log is available, provide the following:
  2025-07-25 17:42:49> Platform information:
  2025-07-25 17:42:49> system - Windows
  2025-07-25 17:42:49> release - 10
  2025-07-25 17:42:49> version - 10.0.22621
  2025-07-25 17:42:49> machine - AMD64
  2025-07-25 17:42:49> processor - Intel64 Family 6 Model 85 Stepping 7, GenuineIntel
  2025-07-25 17:42:49> cpu count - 16
  2025-07-25 17:42:49> ram - 53.4/95.7 Gb (available/total)
  2025-07-25 17:42:49>
  2025-07-25 17:42:49> Python information:
  2025-07-25 17:42:49> alphabase> -
  2025-07-25 17:42:49> alpharaw> -
  2025-07-25 17:42:49> autodocsumm - 0.2.14
  2025-07-25 17:42:49> click - 8.1.7
  2025-07-25 17:42:49> contextfilter - 0.3.0
  2025-07-25 17:42:49> furo - 2024.8.6
  2025-07-25 17:42:49> jinja2 - 3.1.6
  2025-07-25 17:42:49> lxml - 5.3.0
  2025-07-25 17:42:49> myst_parser - 3.0.1
  2025-07-25 17:42:49> nbmake -
  2025-07-25 17:42:49> nbsphinx - 0.9.7
  2025-07-25 17:42:49> numba - 0.60.0
  2025-07-25 17:42:49> numpy<2 -
  2025-07-25 17:42:49> pandas - 2.2.3
  2025-07-25 17:42:49> peptdeep - 1.4.1-dev0
  2025-07-25 17:42:49> pre-commit -
  2025-07-25 17:42:49> psutil - 5.9.0
  2025-07-25 17:42:49> pydivsufsort -
  2025-07-25 17:42:49> pyteomics - 4.7.5
  2025-07-25 17:42:49> pytest -
  2025-07-25 17:42:49> python - 3.9.21
  2025-07-25 17:42:49> scikit-learn - 1.6.0
  2025-07-25 17:42:49> seaborn - 0.13.2
  2025-07-25 17:42:49> sphinx - 7.4.7
  2025-07-25 17:42:49> streamlit - 1.43.1
  2025-07-25 17:42:49> streamlit-aggrid - 1.0.5
  2025-07-25 17:42:49> streamlit> -
  2025-07-25 17:42:49> torch - 2.5.1+cu121
  2025-07-25 17:42:49> tqdm - 4.67.1
  2025-07-25 17:42:49> transformers - 4.47.0

Additional context
This issue may be related to #224, MannLabs/alphadia#409, or MannLabs/alphadia#406.
I modified the function update_sliced_fragment_dataframe in alphabase.peptide.fragment.py at line 605 (https://github.com/MannLabs/alphabase/blob/2a758adcad512a60957960ffdf5ae1f07216c334/alphabase/peptide/fragment.py#L604C9-L604C25) to add a line that directly updates fragment_df instead of only fragment_df_vals:

if charged_frag_types is None or len(charged_frag_types) == 0:
        fragment_df_vals[frag_slices, :] = values.astype(fragment_df_vals.dtype)
        fragment_df.iloc[frag_slices, :] = values.astype(fragment_df_vals.dtype)

After adding this line, transfer learning runs smoothly

2025-07-25 17:42:49> [PeptDeep] Running train task ...
2025-07-25 17:42:52> Loading PSMs and extracting fragments ...
100%|█████████████████████████████████████████████████████████████████████████████████| 27848/27848 [00:11<00:00, 2390.36it/s]
Empty string is detected in the `mobility` column, fill with 0.0
2025-07-25 17:43:11> Loaded 27848 PSMs for training and testing
2025-07-25 17:43:11> Training RT model ...
2025-07-25 17:43:11> 1224 PSMs for RT model training/transfer learning
2025-07-25 17:43:11> Testing pretrained RT model:
   R_square        R     slope  intercept  test_num
0  0.930724  0.96474  0.985213   0.224196       200
2025-07-25 17:43:11> Training with fixed sequence length: 0
2025-07-25 17:43:11> Testing refined RT model:
   R_square        R     slope  intercept  test_num
0  0.930724  0.96474  0.985213   0.224196       200
2025-07-25 17:43:11> Finished training RT model
2025-07-25 17:43:11> Training MS2 model ...
2025-07-25 17:43:15> Testing pretrained MS2 model on testing df:
              PCC         COS          SA         SPC
count  200.000000  200.000000  200.000000  200.000000
mean     0.918067    0.925407    0.777811    0.782237
std      0.083174    0.075037    0.113016    0.160973
min      0.476020    0.511491    0.341813    0.099542
25%      0.894274    0.905665    0.721255    0.693150
50%      0.942028    0.945901    0.789638    0.809570
75%      0.972704    0.974979    0.857291    0.898580
max      0.998703    0.998764    0.968340    0.999954
>0.90    0.725000    0.770000    0.120000    0.245000
>0.75    0.955000    0.970000    0.645000    0.650000
2025-07-25 17:43:15> 1692 PSMs for MS2 model training/transfer learning
2025-07-25 17:43:15> Training with fixed sequence length: 0
[Training] Epoch=1, lr=4e-05, loss=0.024538329740365347
[Training] Epoch=2, lr=6e-05, loss=0.022115163567165533
[Training] Epoch=3, lr=8e-05, loss=0.019727862812578677
[Training] Epoch=4, lr=0.0001, loss=0.01797805226718386
[Training] Epoch=5, lr=0.0001, loss=0.016415516970058282
[Training] Epoch=6, lr=9.045084971874738e-05, loss=0.016211210160205762
[Training] Epoch=7, lr=6.545084971874738e-05, loss=0.014452268121143181
[Training] Epoch=8, lr=3.4549150281252636e-05, loss=0.014053952569762865
[Training] Epoch=9, lr=9.549150281252633e-06, loss=0.013430366354684035
[Training] Epoch=10, lr=1.0000000000000002e-14, loss=0.012770313334961732
2025-07-25 17:43:20> Testing refined MS2 model on training df:
               PCC          COS           SA          SPC
count  1692.000000  1692.000000  1692.000000  1692.000000
mean      0.936367     0.941652     0.822450     0.762332
std       0.102090     0.094025     0.132790     0.150148
min       0.146507     0.190090     0.121756    -0.042815
25%       0.932747     0.938469     0.775512     0.676017
50%       0.975781     0.977678     0.865236     0.784977
75%       0.990360     0.990995     0.914503     0.870501
max       0.999855     0.999859     0.989299     1.000000
>0.90     0.825059     0.835106     0.338652     0.169031
>0.75     0.942080     0.952128     0.790189     0.579196
2025-07-25 17:43:21> Testing refined MS2 model on testing df:
              PCC         COS          SA         SPC
count  200.000000  200.000000  200.000000  200.000000
mean     0.959992    0.963744    0.856733    0.794086
std      0.067847    0.060791    0.098201    0.128814
min      0.483538    0.516474    0.345511    0.295009
25%      0.960632    0.963947    0.828534    0.722612
50%      0.981664    0.983253    0.883327    0.806845
75%      0.991650    0.992322    0.921061    0.889037
max      0.999463    0.999478    0.979430    1.000000
>0.90    0.915000    0.925000    0.385000    0.225000
>0.75    0.975000    0.980000    0.895000    0.670000
2025-07-25 17:43:21> Finished training MS2 model

[jcharkow]

This seemed to fix the issue I was having as well.