A full-day team challenge: predict nine ADMET endpoints on a blinded test set, using everything from Week 1 — descriptors, ensembles, external data, multitask learning, and pretrained models.
Start here
Everything for today is in the shared Drive folder OpenADMET Hackathon!
Open the shared folder link you were given.
Open Notebooks/00_start_here.ipynb.
File → Save a copy in Drive. The shared copy is read-only, so if you skip this you will lose your work when the tab closes.
Work in your copy. Do the same for every notebook you open.
Change CHANGE-ME in the first cell to your pair name and run it. Use the same pair name in every notebook — that is what links your work together, so your split and your saved models follow you from one notebook to the next.
Read README.md for the plan for the day and REPORT_TEMPLATE.md for what you are presenting at the end. Read the report template early; it tells you what to write down as you go.
"Could not find common.py" → you skipped a setup step above.
Anything else → Runtime → Restart session, then re-run from the top. After a restart you have to run the cells above the one you were on.
Notebooks
Start with 00 to get on the leaderboard, then move through the rest as time allows. Each one opens directly in Colab.
Get a first submission on the leaderboard with the simplest possible model. Then look at the data itself: what was measured, how it was transformed, and how much the compounds drift over time.
Build a trustworthy way to estimate your test score, since you only get a handful of leaderboard submissions today. The trick: you're not looking for the split that gives the best score, you're looking for the split whose score best predicts the test score — and those pull in opposite directions.
Compare classical molecular descriptors and fingerprints paired with different regressors to see how much the choice of representation moves predictive performance. It's the first lever most teams pull, and often the one with the biggest early payoff.
Combine models that fail in different ways to see when an ensemble actually beats its best individual member. The question isn't which model wins alone, but how correlated your models' errors are — before you spend a submission finding out.
Pull in public ADMET data from sources like TDC, Polaris, and ChEMBL to expand your training set beyond what's provided. The catch: an assay value measured by someone else isn't the same measurement as yours, and this notebook walks through making them comparable.
Train one Chemprop model with a shared molecular encoder and nine task-specific output heads, instead of nine separate models. Data-rich endpoints end up improving the shared representation that data-poor endpoints then get to use for free.
Fine-tune CheMeleon, a Chemprop encoder pretrained on a large molecular corpus, and try TabPFN, a tabular transformer that predicts on your data without training on it at all. Everything here is single-task, so you can see the effect of pretraining on its own before combining it with multitask learning.