Semsiman calculation taking huge amout of time and resources

[souzadevinicius]

I'm trying to calculate semantic similarity profiles using Phenio ontology comparing different term sets

1. HPxHP
2. HPxMP
3. HPxZP

• HP term set: 17097 entries
• MP term set: 13809 entries
• ZP term set: 39373 entries

Ontology used: Phenio Library versions

semsimian                  0.2.11
oaklib                     0.5.24

---

command line execution example:

runoak --stacktrace -vvv  -i semsimian:sqlite:phenio-monarch.db similarity -p i \
--set1-file hp_terms.txt \
--set2-file hp_terms.txt \
--min-ancestor-information-content 4.0 \
--min-jaccard-similarity 0 \
--autolabel \
-O csv \
-o phenio-monarch-hp-hp.0.semsimian.tsv

I tried to run these experiments locally (32 and 64 GB RAM machines) and in a HPC (writing output process took more than one week and then was killed)

[caufieldjh]

One week, oof! I have previously been able to complete a very similar PHENIO HP vs MP on a GCloud instance with < 64 GB memory, though it did consume a lot of that resource. What kind of resource usage do you see with higher thresholds, like a min of 10 for AIC and 0.4 for Jaccard? The labeling can also consume a surprisingly large amount of resources and is very redundant for this sort of comparison, so I'd suggest dropping that parameter and mapping CURIEs to labels after the comparison is complete.

[matentzn]

@souzadevinicius lets try a 0.4 Jaccard threshold and removing the labelling options and see if that makes it at least possible to run HP-ZP

[souzadevinicius]

Ok

@souzadevinicius lets try a 0.4 Jaccard threshold and removing the labelling options and see if that makes it at least possible to run HP-ZP

[cmungall]

what's the status of this?

[justaddcoffee]

Discussing in the MWF hackathon now

We were thinking we would deploy semsimian/oak on our build server and run on a regular cadence. This way we have an objective measure of how much memory/time we are talking about here, and we can also emit a new artifact with a PURL so people can use this downstream.

@caufieldjh perhaps we already have a repo to do this?

[justaddcoffee]

Ah okay, Harry has already made a repo for this here

[hrshdhgd]

Sorry I'm a little late to the party but @souzadevinicius , did you run this without --autolabel or specify --no-autolabel? Just to get an idea how fast it'll be.

[justaddcoffee]

The last build in Aug '23 took 1h and 18m.

Sorry I'm a little late to the party but @souzadevinicius , did you run this without --autolabel or specify --no-autolabel? Just to get an idea how fast it'll be.

Yep good question @hrshdhgd

Harry says a previous build with auto-label turned on took 15h so this might be at least one thing that is slowing down Vinicius's run

[caufieldjh]

Note that the Jenkins build performed by that repo takes a bit over 1 hr without autolabel and ~15 hrs w/ autolabel.

[caufieldjh]

For reasons not entirely clear to me, this build took 3 hours. Here's the command:

runoak -i semsimian:sqlite:obo:phenio similarity --no-autolabel -p i --set1-file HPO_terms.txt --set2-file MP_terms.txt -O csv -o HP_vs_MP_semsimian.tsv --min-ancestor-information-content 4.0

That's with:

semsimian-0.2.11
oaklib-0.5.25

The product: http://kg-hub-public-data.s3.amazonaws.com/monarch/HP_vs_MP_semsimian.tsv.tar.gz