Improvements multiwgcna

[ivokwee]

Note: Needs PR bigomics/playbase#420 on playbase.

1. Added AI report and diagram for WGCNA and multiomics WGCNA.
2. Various fixes.

[zitoa]

Multi-omics WGCNA should not appear in example-data (as well as any other single-omics data).

[zitoa]

2 new R packages needed: tidyprompt + DiagrammeRsvg

[zitoa]

Error: you can replicate with example-data. Select grok-llama-instant as model.

[ESCRI11]

not sure if related to grok, but once it worked fine, now I just get this error

[ESCRI11]

also, if I compute multiWGCNA AI report and then go to standard WGCNA and try to get AI report I get errors

[ESCRI11]

with mox-brca I just get big crash

[zitoa]

mox-brca dataset. Please load from your library. Then try multi-omics WGCNA (enabling beta features). It crashes...

[zitoa]

@ESCRI11 Exactly same crash i get...

[ivokwee]

Error: you can replicate with example-data. Select grok-llama-instant as model.

This is the DOT diagram code sometimes has errors. Try regenerating in the hamburger menu.

[ivokwee]

ok. mox-brca still has problems. Please try mox-geiger.pgx (proteomics + metabolomics). But the diagram still has often problems in the generated code.

[ESCRI11]

Is the diagram problem related to our code or what we retrieve from llm?

[ivokwee]

OK mox-brca should work with the latest playbase "wgcna-upd2" but the diagram in fact still often fails. It seems to depend on the model. It seems better with grok. or please add groq:openai/gpt-oss-20b in your options. this is like gpt5 but faster.

[ESCRI11]

Maybe we can catch failed diagram and re-send to model until something useful comes up?

[ivokwee]

I added few manual corrections. It should be much less now. Also Llama is pretty poor for the diagram. Yeah, we could perhaps do a 2nd round like "check and correct this code" (could be cool). Or maybe igraph is easier for them?

[ivokwee]

also, if I compute multiWGCNA AI report and then go to standard WGCNA and try to get AI report I get errors

OK. same error. Something in getTopModules. I will look at it.

[ESCRI11]

Ahh so llm is building the diagram code itself? That sounds flaky and unreliable. Maybe get bulletpoint list of nodes + direction of connection and aux function that builds the graph is more reliable, way easier to handle bad llm output for sure

[ivokwee]

True. But sometimes the LLM gives nice annotation on edges, groups, etc. I think it is now much more reliable. GPT and Grok definitely do better job.

I also fixed the getTopModules error in playbase. please check again.

[ivokwee]

Error: you can replicate with example-data. Select grok-llama-instant as model.

The diagram is coded up by LLM but llama seems bad at coding. Grok or GPT seems better.

[zitoa]

Am testing again the PR. Where the corresponding playbase PR? @ivokwee

[zitoa]

The AI diagram errors as per above are still appearing often. I suggest adding a shiny::validate/need message saying (please regenerate diagram) instead of displaying the user with the red error line.

[ESCRI11]

We should detect diagram is not fine and regenerate ourselves (max N times), at that point just notify user diagram could not be generated.

[zitoa]

Diagram seems to be generated by LLM directly....that is what happens

[zitoa]

Multi-WGCNA still appears for single-omics data (example data). I assume this is not the intended behavior.

[ESCRI11]

https://github.com/bigomics/omicsplayground/blob/dev-multiwgcna/components/board.wgcna/R/wgcna_html_report.R#L212 typo here that will crash app

[ESCRI11]

Multi-omics WGCNA should not appear in example-data (as well as any other single-omics data).

added this commit to prevent this from happening 410969e

[zitoa]

playbase PR used for testing bigomics/playbase#418; dataset is mox-brca: crash in multiWGCNA:

[zitoa]

just committed a push to fix a typo most likely causing an edge-case crash in the future.

[zitoa]

mox-brca: crash

[zitoa]

I pushed some fixes into R/pgx-wgcna.R in bigomics/playbase#418. There are other errors and needs testing. If we do not take action in refactoring the code the pgx-wgcna.R will soon be not maintainable.

[zitoa]

I believe there are changes not committed or something went off. There are multiple missing parameters in functions resulting into "objects" or "parameters" not found from corresponding playbase PR. I pushed some fixes...more needed.

[zitoa]

Loading fresh example data. Error:

[ESCRI11]

Loading fresh example data. Error:

i cant reproduce

[zitoa]

mox-brca dataset: "Preservation WGCNA" is broken. Reproducible by clicking at the compute button.

[ESCRI11]

mox-brca dataset: "Preservation WGCNA" is broken. Reproducible by clicking at the compute button.

Reproduced the error on my end as well + same dataset on consensus two boards are not right

update: mox-geiger also breaks on preservation

[ESCRI11]

ok regarding mox-brca and consensus, it seems changing selected trait fixes it, so maybe a more graceful error is required or not allow useless trait selection

on a different note, if on same analysis session i change dataset, the trait list is not properly update, check now im using mox-geiger and trait is the same (actual list of selector is the same not just selected)

i need to succesfully recompute consensus with some condition that works to actually update selectors, they should be cleaned on dataset change

[ESCRI11]

on a separate note, ai diagram still fails sometimes, we need a mechanism to gracefully fail instead of displaying a random error message to the user (which everytime is different)