[GSoC 2025] Generate MS1 and MS2 Peptide Feature Masks for Targeted Deconvolution

[singjc]

Background

Our diffusion model for DIA-MS data deconvolution can benefit from conditioning signals to guide the deconvolution process. Similar to how segmentation masks guide image super-resolution in computer vision, peptide feature masks can provide valuable prior information about where signals are expected in the MS data. This task involves generating these masks from established DIA-MS analysis tools (OpenSwath, DIA-NN, or Spectronaut, DIA-UMPIRE) and integrating them as conditioning signals in our diffusion model.

Task Objectives

• Extract peptide feature masks from one or more DIA analysis tools:
  • OpenSwath
  • DIA-NN
  • Spectronaut
  • DIA-UMPIRE
• Process and format these masks to be compatible with our model's conditioning input
• Implement the conditioning mechanism in our diffusion model architecture
• Evaluate the impact of different mask sources on deconvolution performance

Technical Details

• Feature masks should capture the expected m/z and RT positions of target peptides
• Masks may need to be converted into compatible format for model input
• Consider both binary masks and probabilistic/scored masks based on tool confidence scores

Deliverables

• Scripts to extract and process peptide feature masks from at least one of the target tools
• Integration of the masks as conditioning signals in the model architecture
• Documentation explaining the mask generation process and format specifications
• Evaluation of deconvolution performance with and without mask conditioning
• Comparative analysis if multiple mask sources are implemented

Resources

• OpenSwath documentation
• DIA-NN documentation
• DIA-UMPIRE
• Spectronaut documentation

Difficulty

Intermediate - Requires domain knowledge of proteomics tools and understanding of conditioning mechanisms in deep learning models.

[dev-coder2]

Hi @singjc , I'd love to contribute to Issue #20 ('Generate MS1 and MS2 Peptide Feature Masks for Targeted Deconvolution'). Since this task is crucial for improving our diffusion model’s conditioning, I’d like to start by exploring how peptide feature masks are extracted from DIA-NN, OpenSwath, and Spectronaut. I'll first focus on understanding the data format and compatibility with our model. Let me know if there are specific preferences for the tool I should start with. Looking forward to your guidance! Thanks!

[singjc]

Hi @singjc , I'd love to contribute to Issue #20 ('Generate MS1 and MS2 Peptide Feature Masks for Targeted Deconvolution'). Since this task is crucial for improving our diffusion model’s conditioning, I’d like to start by exploring how peptide feature masks are extracted from DIA-NN, OpenSwath, and Spectronaut. I'll first focus on understanding the data format and compatibility with our model. Let me know if there are specific preferences for the tool I should start with. Looking forward to your guidance! Thanks!

Great! See the Data section under key resources in the preamble

You will probably want to download the raw data under the data availability section, as opposed to the one uploaded on dropbox, which is in sqMass format.

A good place to start would be to use DIA-NN since it's a simple easy to use GUI tool. This will provide you with identified peptides, which you can use to then re-extract the identified regions in the raw data to generate peptide feature masks

[dev-coder2]

Hi @singjc,

I wanted to provide an update on my progress with the 'Generate MS1 and MS2 Peptide Feature Masks for Targeted Deconvolution' issue.

Initially, I faced some issues with MS Reader failing to read the raw data. To resolve this, I used MSConvert to convert the raw files into mzML format, which allowed me to proceed with my work.

I've successfully loaded the FASTA file into DIA-NN and initiated peptide and retention time predictions. I also ran into a few warnings (e.g., MBR requiring at least two files and Prosit forcing C(cam) as a fixed modification), but I believe these are expected given my current settings.

I'm currently waiting for the prediction step to complete and will verify if the output files are generated correctly. If all goes well, I plan to move on to processing the raw DIA-MS data with the generated library.

Since this task is crucial for improving the diffusion model's conditioning, I was wondering if there’s any prior work or existing research I should be aware of on extracting peptide feature masks from DIA-NN, OpenSwath, or Spectronaut? Any additional insights or specific suggestions on improving this approach would be greatly appreciated!

Looking forward to your feedback!
Thanks!

[singjc]

Great, thanks for the update! Are you running the library prediction step on the 6 narrow window DIA files?

Since this task is crucial for improving the diffusion model's conditioning, I was wondering if there’s any prior work or existing research I should be aware of on extracting peptide feature masks from DIA-NN, OpenSwath, or Spectronaut? Any additional insights or specific suggestions on improving this approach would be greatly appreciated!

You can refer to the Dinosaur paper, it should give you a good idea of what we want to do here. https://github.com/Roestlab/diffusion-deconvolution-dia-msms-data/wiki/Google-Summer-of-Code-2025#literature--reading

[dev-coder2]

report.pdf

peptide_feature_mask.csv

Hi @singjc ,

I generated a peptide feature mask heatmap from my 6 narrow .mzML files using DIA-NN. Attached is the heatmap visualization. Could you confirm if this looks correct?

Additionally, here are the DIA-NN parameter settings that were changed from default:

FDR Threshold: 1% (--qvalue 0.01)
Threads Used: 4 (--threads 4)
Verbose Logging Level: 5 (--verbose 5)
Protein Inference Disabled: (--no-prot-inf)
Reanalysis Mode Enabled: (--reanalyse)
Retention Time Profiling Enabled: (--rt-profiling)
MS1 Mass Accuracy: 8 ppm
MS2 Mass Accuracy: 12 ppm
Please let me know if these parameters look good or if any adjustments are needed. Looking forward to your feedback!

[singjc]

I generated a peptide feature mask heatmap from my 6 narrow .mzML files using DIA-NN. Attached is the heatmap visualization. Could you confirm if this looks correct?

@dev-coder2, great, this is a good start. So what you have now is a list of precursor/peptide coordinates (m/z and RT) identified by DIA-NN. Now we want to generate the actual precursor masks from the raw data, so you probably want to do the following:

1. filter the IDs for maybe 5% or 1% FDR (precursor level should be fine) for confident IDs.
2. Using those confident IDs, go back to the raw data to extract the signal for each precursor using the m/z and RT coordinates identified by DIA-NN. To perform this step, you can either write your own code to perform the re-extraction (using pyopenms, pyteomics, pymzml) or you can use MassDash

We essentially want to extract a feature mask like the following:

The signal in blue would be the peptide that DIA-NN has identified, and the red signal would be signal belonging to other analytes. Because want a feature mask, we would only want to have the blue signal, and the red signal would all be zeroed out. This is how we can create precursor feature masks as conditioning signal to train the model to separate the blue target signal of interest from the rest of the red signal.

[dev-coder2]

Hi @singjc,

Thanks for your previous guidance! I want to proceed efficiently and avoid redundancy. Since the README and issue mention that this method has already been worked on and enhanced, could you clarify the existing progress on precursor feature masks?

Specifically:

Are there existing scripts or pipelines for extracting precursor feature masks from raw DIA-MS data?
How can I integrate my DIA-NN identifications with the existing approach?
What areas still need improvement, and where should I focus my efforts?

I’d like to align my work with what’s already been developed so I can contribute meaningfully rather than starting from scratch. Also, could you officially assign this issue to me? I’m gaining confidence in resolving it and would love to take ownership.

Looking forward to your insights!

Thanks!

[dev-coder2]

Hi @singjc,

I’ve attempted to generate a feature mask similar to the one you showed using pyOpenMS for my first narrow .mzML file. Could you confirm if this approach looks correct so that I can proceed with creating binary masks?

Additionally, I extracted precursor features using the combined main output (report.parquet) from all six narrow files. Should I extract feature masks for individual narrow files from this combined result, or should I process each narrow file separately in DIA-NN and then extract features from their respective outputs?

For the red and blue intensity lines, I used 50% of the maximum intensity as the threshold for classification.

Looking forward to your insights!

Thanks!

[singjc]

Are there existing scripts or pipelines for extracting precursor feature masks from raw DIA-MS data?

I'm not aware of any, so you most likely will need to write a script to do this. But you can use any of the packages I mentioned above to make it easier to extract from the raw data. But you can first do a literature/google search to see if there exists any existing scripts/pipelines to do this.

How can I integrate my DIA-NN identifications with the existing approach?

Once we have precursor feature masks, we would want to update the data loader and the model to utilize these feature masks as additional conditioning signals. @LLYX may have more input here.

What areas still need improvement, and where should I focus my efforts?

There currently is no methods implemented for peptide feature mask extraction and use of them as conditioning signals for this project. So it's still very exploratory and very much up to your discretion in the direction you take it and how to implement/integrate it into the project.

Could you confirm if this approach looks correct so that I can proceed with creating binary masks?

For the examples you showed, is this just showing the MS1 precursor profile? What does the associated MS2 fragment ions look like?

Additionally, I extracted precursor features using the combined main output (report.parquet) from all six narrow files. Should I extract feature masks for individual narrow files from this combined result, or should I process each narrow file separately in DIA-NN and then extract features from their respective outputs?

It's fine to process all six narrow files with DIA-NN, since each of the files have a different region of mz that was acquired. To create the feature masks, you can use the combined main output from DIA-NN, just make sure the precursor m/z is within the narrow window acquisition window (if it's not it might be some false target or technical artifact). For example, for the first narrow window file the target acquisition m/z range is 400 - 500 m/z. So you would only expect to see precursors of m/z within that range.

[dev-coder2]

Hi @singjc @LLYX , I have an update on Precursor & MS2 Feature Extraction.

I have successfully extracted MS1 and MS2 features using pymzML from both narrow and DDA files. The extraction process was efficient, and I have also visualized the extracted features.

Key points:

Used pymzML for parsing and extracting spectral data.
Extracted MS1 precursor and MS2 fragment ion features.
Successfully visualized the extracted features in a 3D intensity map.
Below is an example from 23aug2017_hela_serum_timecourse_4mz_narrow_1.raw (400-500 m/z):

Could you review these visualizations to confirm if they look correct? If everything is fine, I will proceed with creating the feature mask and modifying our diffusion model.