Extension for microbiome analyses

[Gian77]

@CMits and team,
Great work! thanks for implementing this!
Before trying it out, I’m curious: how feasible would it be to extend this to microbiomes? Or, if that’s too broad, could it target specific groups like bacteria or Pezizomycotina (fungi)?
Thanks!
G.

[CMits]

Hi G,

Thanks for your question.

For a single microbe, it is very feasible. There is actually more perturbation data available than for plants, so it would work very well.

For a microbiome community, the current architecture would not work, as it is not designed to model interactions between taxa. It would need a different approach, along with new ways to capture and validate the changes. That said, I imagine you could probably use agentic AI to do it, combining several agents.

Which one were you referring to?

Cheers,
Christos

[Gian77]

Hi Christos,
Thanks for your comment. I was thinking more about a community but honestly, both could be interesting. If I need to name a specific bacterium I would say Bradyrhizobium.
G.

[CMits]

Hi,

I run a network in the plant version for "nitrogen fixation in Bradyrhizobium diazoefficiens USDA110" and it gave me a network but I do not know if it is any good. The accuracy was 92.9 % across 42 perturbations.

Let me know if you want to send you the whole outcome of the run, including the graphml, so you can explore it better, as I cannot attach it here. Or you can try the agent and see how it goes. The good thing is that you can then fix a few things that are better suited for microbes in a forked version of yours.

Let me know if you have any other questions.

We are also thinking of a way to make the agent evolve and have new skills for the future. So later on can be like a new skill for microbes.

Cheers,
Christos

[Gian77]

Hi,

This is actually great. Thank you. I am running it right now asking a different question. It is a little bit harsh on tokens, any suggestions? Have you compared different models? Maybe there is no need to run it with the largest model.

G.

[CMits]

Hi,

Did you run the Flash-P_Plant? We have made it so it can run with Claude pro subscription. Me and others have tried it many times and runs fully end to end if the session's usage limit is fresh with claude code Pro subscription. It usually burns 3 M tokens in cache and less than 200k for read and write if I remember correctly. The Flash-P_Paper version is very expensive, you will run out in usage tokens in few minutes with claude code pro subscription. They have tried it for plant traits and species though maybe what you tried was more rich in information and was big network so it run out of tokens faster.

Yes, I agree it does not use Opus in all of the steps, that is why, in Flash-P_Plant, we use Sonnet and Haiku for most of the steps. We use Opus only for the Builder, to do the best possible job in wiring the network. You can change this as you wish in the settings. You can use only Sonnet and Haiku.

Maybe you had already used a lot of the session's usage limits that 4-hour window? Do you use a Claude Pro subscription, or do you try it with Codex? Or do you use just API? We have not used the cheaper models in Codex, probably used the most expensive model as default. I will change the settings in Codex either way as I did not do that in Codex.

Let me know what the case was, as it is supposed to run in a session with the minimum subscription, if it is not, then I will need to prune more in token usage.

Thanks for trying it out, as it is important for us to make sure it runs with the lowest subscription.

By the way I am setting out a discord channel to create a community that uses flash-p. Feel free to join if you want https://discord.gg/bXtTM6MDJ.
I just made it. No worries if it is easier communicatiing from here.

Cheers,
Christos

[Gian77]

Hi Christos,

I used Flah-p_Plant with my Claude Pro subscription, but I think it was not on a clean 4 hours session. I think I had used the model already on other projects, maybe was at 30% usage, and Flash-p ended up running the last agent at about 98% usage, something like that. I can try that again on a fresh session.

As a suggestion, I did this on my local install of flash-p, people would like to visualize the network without running Cytoscape which require additional skills and it's a bit finicky to use. So, I added an agent that plot the network in R for a quick visual :)

Thanks for sharing the channel!

G.

[CMits]

Hi,

Yes, let me know how it goes. I will work in making it cheaper in tokens usage later on.

Ohh yes, great suggestion, thank you! We have a nice visualisation and layout built into the website. Check it out: https://flash-p.com/traits/Arabidopsis_thaliana/Stomatal_Conductance.

I will transfer the scripts of visualsation and layout here in the export step, and a command.

Let me know if you have any other suggestions or questions while using it.

Cheers,
Christos

[CMits]

I pushed the changes for the visualiser in the main you can try it as a command. It creates an svg, png and html that is interactive and you can click to nodes and see the literature used. It will be done in the new networks in the export stage from now on.

[Gian77]

Amazing. Thank you!

[Gian77]

Was just playing around with the website link you sent and I think an import option for people to upload their flash-p-generated network would really great. In that way they can run the simulation they want after the perturbation they applied. I think should no be hard to implement this and, unless I missed it, I think is not possible right now.
Best,
Gian

[CMits]

Hi,

Thanks for the suggestion, we totally forgot it. We made a command {/run-flashp-perturb} that gives you an HTML that you can use to perturb the networks you have exactly as you did in the website. Every new one you do is automatically added to the HTML in the export stage of the agent. So, only run the command once to give you the HTML for the ones you have. We are still testing it for bugs. It is in the Visual-Perturb branch at the moment. I will merge it soon to the main.

Give it a try and let me know how it goes. We added also (/run-flashp-help} for describing the commands.

We will add it to the website soon as well.

Cheers,
Christos