RECON2.2 gives far more flat-zero reaction penalties than RECON2_mat (75% vs. 33%) on identical human scRNA-seq data — expected, or a bug?

[VittoriaOssanna]

Thank you for this awesome tool! I have trouble understanding some behavior I'm seeing with the RECON2.2 model and wanted to check before I trust (or discard) my results.

I used RECON2.2 for human scRNA-seq data since it's the most recent curated update to the Recon human metabolic reconstruction (Swainston et al. 2016) that I'm aware of. I found it via compass --help — it's listed as a first-class --model choice (RECON1_mat, RECON2_mat, RECON2.2, Human1, Mouse1) — but I couldn't find it covered anywhere in the tutorial or docs site, and (unlike RECON2_mat) no reaction metadata file ships with it (same gap reported in #52). So I'm not sure if this is an actively supported/validated path or something present in the CLI but not really vetted for the standard human pipeline.

Running it on the same 20-cell test set with identical parameters aside from --model, I get a much higher fraction of reactions that are a hard 0.0 across every single cell:

Model	Reactions	All-zero across all cells
RECON2_mat (default)	10,211	33.4% (3,414)
RECON2.2	11,567	75.0% (8,670)

RECON2.2 run:

compass --data smoke_CPM.tsv.gz \
        --model RECON2.2 --species homo_sapiens \
        --output-dir output \
        --temp-dir temp \
        --num-processes 20 \
        --num-threads 1 \
        --lpmethod 2

RECON2_mat run (only --model changed):

compass --data smoke_CPM.tsv.gz \
        --model RECON2_mat --species homo_sapiens \
        --output-dir output \
        --temp-dir temp \
        --num-processes 20 \
        --num-threads 1 \
        --lpmethod 2

Some zero-penalty reactions are legitimately gene-less (no GPR), but a meaningful chunk of the RECON2.2 zeros are reactions tied to genes that clearly have expression in my input (e.g. a reaction requiring DBI, at CPM 400–700 in most sampled cells, still comes out 0.0).

Is this difference expected given the two models' different structure/gene curation, or does it point to a mapping issue with RECON2.2 specifically? Which output should I trust for downstream analysis — and is RECON2_mat the intended default for a reason (e.g. is RECON2.2 known to still need work)?

Thanks in advance for any pointers!