Michael Lazear

Dude is an monster... and only 16s off of Remco on the TT today.

Can confirm that it really increases (subjective) MS1 XIC quality for us.

I'm not an IM expert, but isn't IM extremely strongly correlated with m/z of the precursor though (at least for TIMS)? High resolution IM is probably just as good as knowing the quad isolation window. And multiple charge states are already fragmented at once in DIA?

For further AI-readiness stuff, I am the wrong person to ask. Nearly 100% of my code is artisanal, organic, small-batch, and handcrafted. Exceptions are occasionally made for one-off scripts that I know how to write, but LLM can write faster.

As Tine mentioned, data cleaning/prep is usually the most important part. Some kind of metadata tracking system for all of your samples (could be a CSV file, a SQL database, etc) is critical. I know Claude can interact with SQL databases directly using some of the MCP plugins

🧪 Transforming biomedical discovery needs proteomics that is specific and accurate, but also faster and cheaper

@parallelsq.bsky.social is proud to introduce 9-plex PSMtags: a new mass tag that improves sequencing and increases throughput in sensitive proteomics 👇

www.biorxiv.org/content/10.1...

@michaellazear.bsky.social & @jspaezp.bsky.social I'm Running Sage on a dual Epyc 128 thread box w/ 2TB memory (I think) . Searching 19 ddaPASF files directly with built in timsRUST!! Dang this is screaming fast !!

You can definitely burn parchment paper in an oven.

For gaming, 4060. I'm still rocking an old 1080 Ti

What's the use case? More GPU RAM is generally best.

Chiang and Collins making a fun Casanovo + Sage (w/Mokapot et al.) metaproteomics pipeline (Bittremieux, @michaellazear.bsky.social @willfondrie.com) … and an exceptional TOC graphic.

I have just a gravel bike right now that I've slapped some fast tires on - but frequently find myself spinning out on the 1x10 gearset. So looking for something faster, and I've enjoyed test riding the racier bikes

That is what I'm leaning towards!

Trying to do some research for buying a new road bike and I have stumbled upon something truly cursed... what appears to a forum of just LLMs talking to one another: www.cyclingforums.com/threads/gian...

Oh, for sure. Not all tools need to be hyper-optimized. The point of the parent post was why bother writing a GPU search engine if you haven't squeezed out every CPU cycle first 😉

This is a completely meaningless way to measure "pushing CPUs to their limit". You can make an infinite spin loop that does nothing and has 100% utilization. How many spectra are you processing per ms? What is the throughput in GB/s?

Yes, it's for personal use. But I also write proteomics software, so it kinda blurs the line 😉. Being able to search fast is critical for keeping a tight iteration loop.

AFAIK DIA-NN still isn't GPU for it's deep learning-based spectral predictions despite using (I believe) torch under the hood. So don't think it would do any good anyways

Re: single thread, I think most of the tools are pretty good about being highly parallel, to varying extents. But faster is usually still better. I just built a new PC with AMD 7950x (32 cores at ~4.5 GHz) and I'm pretty happy with it lol

That's gonna be highly specific to your software. Disk speed will always be important for reading data - for a closed search using Sage, reading data takes longer than searching data. FP/DiaNN/MQ all do pretty heavy disk reads and writes throughout their execution, so that will impact them more.

Prolucid GPU and Andy Kong's unpublished GPU search engine come to mind. GPUs are pretty hard to program for though... And to be honest, most extant tools aren't even pushing CPUs to their limits yet.

Horizontal scaling (e.g. spinning up 100s or 1000s of VMs to each search a single file) can be faster, but is much less user friendly if you aren't a programmer.

RAM speed and memory bandwidth can be pretty important. Unfortunately many tools don't run on Apple Silicon or other ARM CPU architectures, but they tend to have incredibly high memory bandwidth. You will always hit a wall with how much you can vertically scale though.

It's one of the really cool and unique applications for MS. Hard or impossible to do many kinds of chemoproteomics experiments with non-MS methods!

Yep, you can run locally and script it. Just need to do "sage params.json -f fasta1.fasta", etc

I would be happy to help run that - already have a nice distributed/fan out system in place. You can find some examples in the docs too: sage-docs.vercel.app/docs/configu...

👍nobody asked for your response either! And yet, that is the point of using a social media app like this.

Lisp (1960; R has some strong Lisp influence) famously has first class AST/homoiconicity. If you havent played around with a Lisp or read some of SICP, definitely worth a shot and a nice introduction into programming language theory (lambda calculus, etc)

R is a terrible programming language. Seems great for stats though!