Proposal: Fine-tuning DreaMS on lipid data for lipid annotation

[Sajjadkhani]

Hi, I’m planning to work on fine-tuning DreaMS using labeled lipid MS/MS data to improve lipid annotation.

The idea would be to add a small fine-tuning objective + a minimal example/tutorial.

Before starting, I wanted to ask whether this direction fits the project’s scope and if you’d be open to a PR along these lines.

Thanks!
Sajjad

[roman-bushuiev]

Hi Sajjad,

It definitely sounds great! Let me know if you face any issues with fine-tuning.

Are you planning to focus on a specific type of lipids? If your problem can be framed as a binary classification, I think you could reuse our approach for fluorine detection (i.e., linear head and focal loss). We have tested it on several binary classification problems beyond fluorine detection and it worked well.

Best,
Roman

[Sajjadkhani]

Hi Roman,

We’ve already tested multiclass lipid classification using the base DreaMS model, and we’re seeing around ~0.7 accuracy. At the moment, our main interest is improving fingerprint (FP) prediction rather than classification alone.

Our plan is to first clean and curate labeled lipid MS/MS data from MoNA plus some in-house data, and then fine-tune the base model using this dataset. The goal is to see whether fingerprint prediction performance can be improved, at least locally, for data generated at our facility.

I’ll keep you posted as things progress, and thanks again for being open to this direction.

Best,
Sajjad

[roman-bushuiev]

Hi Sajjad,

That sounds great 👍

We are currently working on DreaMS-Mol, a model for retrieving/generating molecules from MS/MS spectra, without relying on an intermediate fingerprint prediction step. So, we didn't further optimize fingerprint prediction beyond what is reported in the paper. Currently, we have only the first version of the model but later we could also test it or fine-tune it on lipids.

Best,
Roman

[Sajjadkhani]

Hi Roman,
Thanks for the update.
FP prediction is also a middle step for us. When are you planning to release the DreaMS-Mol?
Best
Sajjad

[roman-bushuiev]

Hi Sajjad,

We are planning to release it this year but currently we have only the first version of the model.

Best,
Roman

[Sajjadkhani]

Hi Roman,
I’m looking forward to the new release. Dreams is already very useful, and hopefully in the meantime I’ll be able to achieve some improvements through fine-tuning.

Best,
Sajjad