#llms-for-bioc

2025-02-02

Vince Carey (11:52:36): > @Vince Carey has joined the channel

Bastien CHASSAGNOL (11:53:19): > @Bastien CHASSAGNOL has joined the channel

Sebastian Lobentanzer (11:53:19): > @Sebastian Lobentanzer has joined the channel

Tyrone Lee (11:53:19): > @Tyrone Lee has joined the channel

Ludwig Geistlinger (11:53:19): > @Ludwig Geistlinger has joined the channel

Stephanie Hicks (11:53:19): > @Stephanie Hicks has joined the channel

Alex Mahmoud (11:53:19): > @Alex Mahmoud has joined the channel

Sean Davis (11:53:20): > @Sean Davis has joined the channel

Vince Carey (11:53:37): > If this is redundant with another channel please indicate here.

joe mcconnell (19:55:50): > @joe mcconnell has joined the channel

2025-02-05

Ludwig Geistlinger (08:42:06): > Thanks for creating the channel@Vince Careyand congrats to an outstanding piece of work in that space@Sebastian Lobentanzer, just out in Nature Biotechhttps://www.nature.com/articles/s41587-024-02534-3 - Attachment (Nature): A platform for the biomedical application of large language models > Nature Biotechnology - A platform for the biomedical application of large language models

Sebastian Lobentanzer (10:38:23) (in thread): > Thanks a lot for sharing@Ludwig Geistlingerand indeed many thanks@Vince Careyfor creating the channel

2025-02-10

Vince Carey (07:10:58): > Can we use this channel to discuss issues with genAI approaches in bioinformatics? I would like to hear from people who have well-maintained deployments of geneformer, scGPT, GEARS … as I find the configuration/use of these systems very challenging to say the least. scGPT has open issues regarding obsolescence of configuration spec, as an example. Or would it be wise to start another channel?

Sebastian Lobentanzer (08:50:49) (in thread): > Good question.I am interested in both, we have ongoing projects with genAI models, but technically they are not LLMs, and their purpose is also different.While LLMs for me are mainly a tool for accessibility (whetherit’sbioconductor generally or some specific software), the models you mentioned are specialist tools.Happy to chat about both; if you create a new channel, please add me.:)

Sebastian Lobentanzer (08:52:12) (in thread): > I’m also not sure what the overlap would be of people interested inone or the other here in the Slack.

Tyrone Lee (10:08:31) (in thread): > CCB has experimented with scGPT and geneformer. From that experience, there is quite a bit of fine-tuning needed to perform downstream tasks (see thistutorialfrom czi, they did the fine-tuning for embeddings on their data) . Their virtual cell models platform seems to be designed for this use case, fine tuning biology models for transcriptomics and imaging.

Tyrone Lee (10:11:23) (in thread): > Fine-tuning has quite a (still) substantial training compute cost which why not a lot of labs do this. I think it will be quite some time before something like this becomes commonplace for bioinformatics models, or Bioconductor for that matter.

Vince Carey (10:11:37) (in thread): > i will make a genAI-bioc channel.

2025-03-06

Tyrone Lee (11:55:13): > ProCyonhttps://github.com/mims-harvard/ProCyon - Attachment (zitniklab.hms.harvard.edu): ProCyon > A foundation model for protein phenotypes

2025-03-08

Alex Mahmoud (18:45:06): > https://medarena.ai/came across this today

2025-03-14

Sean Davis (16:00:36) (in thread): > Only for clinicians. For any such effort that comes across in the context of clinical use cases, be VERY careful to read the terms of service and to abide by them. My guess is that most university employees would be told by their information security teams that this is not an application that they can use. I’m guilty of using services that aren’t fully vetted, but when they are in a clinical/medical context, I pay much closer attention.

Sean Davis (16:01:20) (in thread): > Looks like an interesting application, though!

2025-04-13

Dario Righelli (16:33:04): > @Dario Righelli has joined the channel

2025-04-17

Juan Henao (17:30:49): > @Juan Henao has joined the channel