Signature bug

[allonw]

The following session throws an error on v2.1.0 (code for producing this vis object is also available). Creating the same object with a small set of standard signatures runs successfully.

` library(VISION)
vis = readRDS("/data/yosef/users/allonwag/Tim3/processed_sources/First10x/outputs/Vision/vis_bug.rds")

vis = analyze(vis)
`

Beginning Analysis

Clustering cells...completed

Projecting data into 2 dimensions...
Running method 1/2: UMAP ...
Running method 2/2: tSNE30 ...

Evaluating signature scores on cells...

|================================================================| 100%, Elapsed 00:02
Evaluating signature-gene importance...

|================================================================| 100%, Elapsed 00:05
Creating 5 background signature groups with the following parameters:
sigSize sigBalance
1 6 0.9987715
2 11 0.9948271
3 28 0.9887866
4 53 0.6131914
5 96 0.9385173
signatures per group: 3000
Computing KNN Cell Graph in the Latent Space...

Evaluating local consistency of signatures in latent space...

|================================================================| 100%, Elapsed 00:00
|================================================================| 100%, Elapsed 00:02
|================================================================| 100%, Elapsed 00:13
|================================================================| 100%, Elapsed 00:00
Clustering signatures...

Error in [.data.frame(sigMatrix, , keep_significant, drop = FALSE) :
undefined columns selected

[mattjones315]

Hi Allon,

does this vision object have any meta data associated with it? If so, NA's in the meta data will cause an error like this.

I'm happy to take a look at your signature data as well if you'd like to send that over.

-Matt

[allonw]

Yes, some of the metadata have NAs. I'll try again with replacing these with explicit values - but I imagine this isn't an uncommon scenario - so maybe check that before running and give a friendly error message in that case?

[mattjones315]

Sure, it'd be good idea to add this. Thanks for the suggestion.

I'm going to close this issue now, but please let me know if you encounter any other issues.

[allonw]

Actually, it doesn't seem to be related to NAs in the metadata. I'm attaching a minimal example, taking only the first 1000 signatures, for example runs successfully.

Input: vis_bug.rda.zip

library(VISION)
load("vis_bug.rda")
vis = Vision(minimal_mat,
signatures = minimal_sigs, #or: minimal_sigs[1:1000] to run successfully
meta = minimal_meta,
projection_genes = projection_genes,
projection_methods = c(),
pool = FALSE,
latentSpace = minimal_latent,
latentSpaceName = "scvi_latent_space")

vis = analyze(vis)

---

output:
Beginning Analysis

Clustering cells...completed

Projecting data into 2 dimensions...

Evaluating signature scores on cells...

|================================================================| 100%, Elapsed 00:01
Evaluating signature-gene importance...

|================================================================| 100%, Elapsed 00:01
Creating 5 background signature groups with the following parameters:
sigSize sigBalance
1 6 1
2 10 1
3 17 1
4 37 1
5 121 1
signatures per group: 3000
Computing KNN Cell Graph in the Latent Space...

Evaluating local consistency of signatures in latent space...

|================================================================| 100%, Elapsed 00:00
|================================================================| 100%, Elapsed 00:11
|================================================================| 100%, Elapsed 00:00
Clustering signatures...

Error in [.data.frame(sigMatrix, , keep_significant, drop = FALSE) :
undefined columns selected

[mattjones315]

Hi Allon,

Thanks for the reproducible example & re-opening the issue. I believe I've found the issue - it appears that signatures are duplicated in the signature list you're providing VISION (hence why running on the first 1000 signatures works, but not the entire set). In particular, I've found the following signatures duplicated:

[1] "REACTOME_PEXOPHAGY"
[2] "REACTOME_SIGNALING_BY_ERBB2_ECD_MUTANTS"
[3] "REACTOME_SEALING_OF_THE_NUCLEAR_ENVELOPE_NE_BY_ESCRT_III"
[4] "REACTOME_SIGNALING_BY_PDGFR_IN_DISEASE"
[5] "REACTOME_SIGNALING_BY_PDGFRA_TRANSMEMBRANE_JUXTAMEMBRANE_AND_KINASE_DOMAIN_MUTANTS"
[6] "REACTOME_SIGNALING_BY_MEMBRANE_TETHERED_FUSIONS_OF_PDGFRA_OR_PDGFRB"
[7] "REACTOME_NERVOUS_SYSTEM_DEVELOPMENT"
[8] "REACTOME_DISEASES_OF_DNA_REPAIR"

I was able to successfully run VISION after filtering out the duplicated signatures, arbitrarily with this code:

new_sigs = list()
for (sig in names(minimal_sigs)) {
    if (! (sig %in% names(new_sigs)) ) {
        new_sigs[sig] = minimal_sigs[[sig]]
    }
}

Then, with re-parameterizing the Vision object with new_sigs, I was able to successfully analyze the data. I suggest you do something similar.

I will add a statement from within the Vision object constructor to make sure cases like this are flagged.

Thanks again, let me know if this fixes everything on your end!

-Matt