Siegfried Gessulat

Hello #ASMS2025! Get ready for two exciting presentations from our team in Baltimore — happening tomorrow!

🔗 msaid.de/conferences/...
📍 Stop by Booth #203

#TeamMassSpec #TeamProteomics #Proteomics #MassSpectrometry

#TeamMassSpec, don't miss my talk about double dipping in DIA data analysis on Tuesday afternoon at #ASMS2025! It's not as savoury as the one below, but will contain a couple of spicy takes. 🙃

ASMS 2025 | MSAID ASMS 2025 – Find more information about our participation in conferences and events.

5/5 All of @msaid-de.bsky.social at #ASMS2025 can be found at www.msaid.de/conferences/.... Two more highlights:
- @martinfrejno.bsky.social's oral TOA pm 04:10 shows a confirmation bias in DIA that inflates data completeness
- Daniel Zolg's oral ThOA pm 03:30 showing our platform

Enjoy!

4/5 Our newly introduced TimsTOF ✈️ support deserves its own spotlight! Visit @tkschmidt.me on Tuesday at TP 507. We show how Chimerys 5 and Inferys 5 perform on dda-PASEF and dia-PASEF and outperform or match alternatives out-of-the-box.

3/5 What about speed? 🏃‍♀️ Tokenization can compress the input by more than 100x. Weight quantization doubles our speed, and knowledge distillation gives another 30%. Combined, Inferys 5 is faster than Inferys 4 at the same high accuracy.

2/5 Instead of amino acid tokens, we show Inferys 5 the full molecular structure 🧬. Tokenization and data augmentation enable us to vary the structure's detail level. We can trade speed for out-of-distribution generalization similar to chain-of-thought reasoning models.

🧵1/5 Although I'll miss #ASMS2025 I'm hugely excited about @lizimamisashvili.bsky.social's oral 🕵️‍♀️ TOE am 09:50. Inferys 5 is our biggest ML update in years. We support any PTM input in principle, dramatically increase speed, and add support TimsTOF. All with a single model.

We are working hard on generalizing our peptide property prediction models to more diverse peptide classes! Watch out for my colleague @lizimamisashvili.bsky.social's ASMS talk (TOE am) to input any molecular structure to our models. Next, we need to generalize our model output too.

Thanks for bringing this to our attention - this is caused by Met-containing peptides in the LFQ Benchmark DIA files. We state in the Methods that we filter these out, but in E10-C they slipped through. We fixed this and quantification now looks as expected (github.com/msaid-de/chi...)

We are working hard on generalizing our peptide property prediction models to more diverse peptide classes! Watch out for my colleague @lizimamisashvili.bsky.social's ASMS talk (TOE am) to input any molecular structure to our models. Next, we need to generalize our model output too.

A really solid paper from the MSAID team.
I’d guess there’s 30 minutes worth of information in the supplementary alone (check out the discussion).
Covers many aspects of spectrum-centric proteomics and why it’s the best single approach for DDA, DIA, and PRM data.

(i don't have a car)

Surf's up! Visit the famous Eisbach wave See surfing action in the middle of Munich - all year long

People Munich's city center take their boards to the Eisbach wave by bike or bus. 🏄‍♂️
www.muenchen.de/en/sights/ei...

Here in Berlin, I could reach the sailing class I took last year by S-Bahn in 30 min after work. It's located at the former Socialist State Yacht. en.wikipedia.org/wiki/A._K%C3...

I am happy that our first publication showcasing the MSAID Platform is out! Read all about its unique features or try it for free at platform.msaid.io!