Drug-discovery analyses: connectivity-map / signature reversal, synergy & dose-response, genetic interactions

[Zethson]

Context

pertpy already ships the drug/MoA/compound metadata scaffolding (pt.md.Drug, pt.md.Moa, pt.md.Compound) and drug2cell Enrichment, but not the analyses that consume it.
These are core pharma workflows sitting right next to metadata we already provide.

What's missing

• Connectivity-map / signature reversal (LINCS/L1000-style): given a disease/query signature, rank perturbations by how strongly they reverse it. Complements drug2cell enrichment.
• Synergy models for combinations: Bliss / Loewe combination-index models. Cinemaot covers single-cell synergy via optimal transport, but not the classic combination-index framework.
• Dose-response curve fitting: Hill / EC50 / IC50 as first-class output, building on the dose-response quantification already in the perturbation-space base class.
• Genetic-interaction / epistasis scoring (Norman 2019-style GI scores) for double perturbations.

Proposal / API

Add these under pertpy.tools, reusing the perturbation-space representations and Distance metrics where possible.
Signature reversal can operate on perturbation-space embeddings; GI scoring on paired-vs-single perturbation representations.

[daveringelberg]

Hello,

I'd be interested in helping with a small first slice of this issue.

Would a PR for CMap-style signature reversal be welcome? I'm thinking of starting with only:

• input: perturbation-level AnnData with one row per perturbation and genes as variables
• query: up/down gene sets
• output: a ranked DataFrame of perturbations by reversal score

I would leave Bliss/Loewe synergy and dose-response for separate follow-up PRs, since those are distinct pieces of functionality.

Does that sound like a useful first scope?

[Zethson]

Dear @daveringelberg,

thank you for your interest! This issue is unfortunately AI generated just so that I remember eventually looking at this.

However, yes, I think this would be a useful first scope. We should just think about the API & UX first.
Could you please draft here how the user would call it and what kind of insight & plot they'd get out out of it? Note that pertpy builds on scverse data structures so it should feel natural and we often avoid just returning DataFrames. It's bad for the way we're doing plots etc.

[daveringelberg]

Thanks for the response.

The main insight would be: which perturbations most strongly reverse this disease/query signature, and are therefore candidates for follow-up validation.

I would center this on AnnData rather than returning a DataFrame. A possible way to call it would be:

ps = pt.tl.PseudobulkSpace()
ps_adata = ps.compute(adata, target_col="perturbation", mode="mean")

sr = pt.tl.SignatureReversal()
sr.score(
    ps_adata,
    up_genes=["GENE1", "GENE2"],
    down_genes=["GENE3", "GENE4"],
    key_added="signature_reversal",
)

sr.plot(ps_adata, key="signature_reversal", n_top=20)

I'd use a ranked bar plot first.

Does this match the direction you have in mind?

[Zethson]

I wonder whether this should be its own tool or whether we can come up with a more general tool where this would then be an option. Kind of like pertpy has the Distance API. plot is also too generic. Can we not reuse a scanpy plot or something?

I also wonder whether https://pertpy.readthedocs.io/en/latest/api/tools_index.html#enrichment can be reused here? Maybe this should rather be included into that already?

[daveringelberg]

Maybe we could start with Enrichment and move to a more general tool if it outgrows it? Possibly:

ps_adata = pt.tl.PseudobulkSpace().compute(
    adata,
    target_col="perturbation",
    mode="mean",
)

enr = pt.tl.Enrichment()
enr.signature_reversal(
    ps_adata,
    up_genes=["GENE1", "GENE2"],
    down_genes=["GENE3", "GENE4"],
    key_added="signature_reversal",
)

For plotting perhaps:

top = ps_adata.obs.nsmallest(20, "signature_reversal_rank").index
sc.pl.matrixplot(
    ps_adata[top],
    var_names=["GENE1", "GENE2", "GENE3", "GENE4"],
    groupby="perturbation",
)

Would this be closer to what you have in mind?

[Zethson]

Yes, this is much closer! Let's start with this

[Zethson]

Great! Now we have a first step. I'm open to more ideas and PRs here if you'd be up for it.
Thank you

[daveringelberg]

Absolutely, I'll take a closer look at the remaining parts, thanks again!

[daveringelberg]

The synergy package already implements Hill, Bliss and Loewe. Given its GPL-3.0+ license, would using it through an adapter be acceptable, or would you prefer a native Pertpy implementation?

I'd be inclined to start with Hill fitting, since Loewe requires fitted single-agent curves, and leave Bliss/Loewe for later. The fitter could use distances returned by dose_response() to estimate an EC50 for the transcriptomic-distance response, or a normalized inhibitory response to estimate an IC50. Does that scope sound acceptable?

[Zethson]

Unfortunately, GPL 3 is a big no go so we need a native implementation. Your plan sounds great!