Additional Filters

[mrtoronto]

TL;DR:

We may be able to derive some more filters to remove samples that aren't relevant for the classifications but I'm not sure what they should be.

Details:

I did a 500 sample test run recently with the cells and multichannel filters on and it was reduced down to ~200 samples. Looking through the results, I think we still have some samples that don't fit a category.

I put all the samples into a google sheet here and added a manual category column where I started adding what categories certain samples should go in. Rows marked flagged were picked up by the rough regex flags I already put together. I also added data from a 2k sample run (filtered to ~800 samples) but didn't get to adding any manual category labels there.

While I haven't gone through the whole thing yet, I found a couple rows that mentioned something like gene expression in a tissue but didn't mention mice or a treatment to a mouse. Examples I found already were on lines 3 and 5 of the 500 sample sheet.

• Do you think we should use a filter to capture these samples and remove them?
  • If so, what's the higher-level group these fall under?
  • If not, do they fit into a category that I'm missing or why do we want to save them?
• Also, feel free to add ideas about any other filters below.

[roburrhoades]

I see some regexes have been added for wild-types, genetic interventions, and chemicals. I think that is a start, though my question would be how confident we could be before the ability to actually match gene and chemical names? Or do you think that detail of matching is overkill for the purpose of categorization? (keep in mind there are a lot of nuance even in mouse studies: environmental exposure,s maternal diets, fasting/CR, etc)

again, maybe we just need some regexes to cover more of these situations without a very clear and accurate protocol. At least at that point we can start scaling classifiers on experimental data fed to the metabolic modeling algos. Agree?

[mrtoronto]

I figure if we used a similar approach to the ages, we'd just have samples get flagged one way or another and the flags could be used as signals to pass samples through to a specific function made to extract info from whatever that flag corresponds to.

Detecting the type of intervention feels like a full piece of work that we can try to iterate on and refine, increasing the accuracy and the coverage as we go. Feeding the flagged samples into another function which will return the active part of the intervention seems like a second piece to me. We may be able to use some overlapping regex for the detection and extraction but I think doing both at the same time would get too complex.

I think iterating on the detection regexes will be a good start, just until we have a decent number of samples in the major groups then we can look at flagged samples and try to extract and verify info from them.

To answer you question about confidence in the flags, I think we could eventually be reasonably confident once a sample passed the flag-level test that it was legit but there would definitely be some false positives. Maybe we have a multi-tier rule system to check out the positives for common signs of a false positives but that's close to becoming a rule rabbit-hole. We may also be able to send samples in for verification using one of the services we've talked about. We could do this separate from the scraper too so we wouldn't have to worry as much about speed.

That all being said, I think getting decent coverage across flags would be a good starting point.