Demetrius DiMucci

Machine learning reveals missing edges and putative interaction mechanisms in microbial ecosystem networks. I made this. Neat. biorxiv.org/content/early/…

Part 2: So why DO Random Forests work?! On this, I’ll have to disagree with Elements of Statistical Learning (my first time ever 💔) EoSL says the success of forests should be understood as a consequence of variance reduction *alone*, but I think that’s not a good intuition 1/n

@SamHaney12 @ABindoff @hardsci It wasn't obvious to me. This has happened to friends of mine.

@ABindoff @hardsci Would this be an obvious joke to someone on the science/grad school community? To someone who doesn’t do this for a living, there’s absolutely no way to tell this is a joke.

Published this morning. A major leap forward in my own understanding of just what it is a Random Forest is doing: frontiersin.org/articles/10.33…

@baym @SamPinzkaBerry Can you point out some of these? Interpretable models are relevant to my own projects.

@SamPinzkaBerry Agreed. And there's some super interesting work happening on interpretable models (and elsewhere in AI)

Will I lose my computational biologist card if I admit that I think the growing focus on all AI/DL all the time is really boring scientifically?

Goodnight Moon Knight

@PhDemetri That guest spot on a podcast.

10,000 people follow me and honestly I could not tell you why.

@Astropartigirl @SquidForADay That's where I'm going. To see what happens.

@SquidForADay Well I should really say, even we *think* protons will decay, because we have not actually seen a proton decay yet!

You're given the opportunity to time travel. Your spaceship and suit protect you from ANY danger, and you have all tools you need for research (telescope, laptop, etc). What do you choose? A) 10^-36 s after Big Bang B) 200 Myr into the past A) 2 Gyr from now C) 10³⁴ yr from now

@MeredithAlberta The time a TA failed me on a problem set despite my answers matching up with the answer key because "In his country not every student is given a passing grade" and the dean essentially responding to my complaint with "Git gud scrub."

@upulie He's losing GA

PSA: Uninterrupted exponential growth is NEVER a realistic assumption for an outbreak model and if you think it’s ok to use for a “first pass at the numbers” then you really really shouldn’t be posting about your model on the internet.

@Brave_Recovery @ReaLauren89 @mlipsitch @DrEricDing Those aligned sequences are so short, they'll get perfect matches to all kinds of things. Looks to me like a classic case of cherry picking in order to tell a story.

@ReaLauren89 @mlipsitch @DrEricDing They are working to get to the bottom of it and I trust them fully

Just off the phone w reporter for major global newspaper whom I had to talk down from a story on "I've read that use of antibiotics in agriculture can select for superbugs, you know, really bad viruses that can infect people." Glad I took the call as I assumed it was some...[1/n]

@victorveitch Ah gotcha. I've seen Frank Harrell talk about calibration curves - just plotting the predicted class probability against the observed class probabilities, then the r^2 could be monitored. That's the extent of what I've come across to date in this vein.

@DiMucciDM To be clear, the data is simply features and binary labels. The question is what are simple metrics to monitor to see if the predicted probabilities are behaving sensibly

Is there a standard(ish) metric for calibration useful in the following setting? Unknown to you: Objects with features in set 'A' are 1 wp 30% Objects with features in set 'B' are 1 wp 10% Question: do I have a good model of P(Y=1 | features) ?

@victorveitch Wouldn't we be better served with a predicted probability of class membership given a set of features rather than a binary 0/1 call in this situation?

@DiMucciDM Probably not? Anything that's a measure of classification performance seems inappropriate because the optimal classifier always predicts 0.

@victorveitch Would the precision and recall for Y | features be useful?

I'd like something I can monitor in a similar fashion to accuracy. Accuracy, ofc, doesn't work because the ground truth classifier would just predict 0 for every feature.

My article has been recommended on @F1000Prime f1000.com/prime/73436720…

Really excited to have my first paper from grad school published in @elife! The nicely formatted version is now out! A computational analysis of human oral microbiome organism’s biosynthetic capabilities doi.org/10.7554/eLife.…

@MelisaLKO @isayhionthe75 @segrelab @BostonBacteria Cool! Is the poster available to see online?

Excited to share a pre-print of our Penicillium experimental evolution project. We discovered rapid loss of pigment, secondary metabolite, and spore production in Penicillium when passaged on cheese. Evolved strains also smelled wildly different (see awesome VOC data!).

Beet salad at @nelcentro Portland food tour.

Recently discovered my paper has been cited for the first time. Happy to know I've contributed to the conversation and not just shouting into the void.