You have 8,528 commercially available carboxylic acids and a budget of 600 DFT labels — about 70,000 CPU-hours. Train a Chemprop D-MPNN that predicts 156 conformer-ensemble descriptors, and know how uncertain it is.
The premise
Every label in this tutorial is a published DFT number. The descriptors come from Haas, Hardy, Sowndarya, Adams, Coley, Paton & Sigman, Digital Discovery 2025, 4, 222 — 8,528 carboxylic acids, 71,324 conformers, computed at M06-2X/def2-TZVP // B3LYP-D3(BJ)/6-31G(d,p) over Maestro conformer ensembles with GoodVibes quasi-harmonic Gibbs energies, Boltzmann-weighted at 298.15 K.
Their paper reports over 1,000,000 CPU hours for the acid library. That is roughly 117 CPU-hours per molecule, so your 600-label budget is about 70,000 CPU-hours — eight CPU-years. The notebook prints that number before you spend anything.
Nothing in this tutorial recomputes chemistry. The labels already exist and are served over an open API. What is scarce is not compute — it is your permission to look at them. That is exactly the situation active learning is for.
156 targets: 39 base descriptors × 4 conformer aggregations (_min, _max, _low_e = lowest-energy conformer, _boltz = Boltzmann average). Atom labels follow the paper: C1 carboxyl carbon, O2 carbonyl oxygen, O3 hydroxyl oxygen, H5 acidic hydrogen, C4 α-carbon.
| Family | n | What |
|---|---|---|
| buried volume | 12 | %Vbur at C1 and C4 (3, 4, 5 Å) plus C1 max/min hemispheres |
| NBO charges | 5 | natural charges on C1, C4, H5, O2, O3 |
| frontier orbitals | 5 | HOMO, LUMO, hardness η, chemical potential μ, electrophilicity ω |
| spectroscopic | 4 | NMR shifts (C1, C4, H5) and the harmonic C=O stretch |
| Sterimol | 3 | L, B1, B5 along C1→C4 — the sterics of R seen from the acid carbon |
| SASA | 3 | surface area, enclosed volume, sphericity |
| geometry | 3 | O–H length and the two carboxyl dihedrals (acid proton syn/anti) |
| electrostatic | 4 | dipole, isotropic and anisotropic polarisability, molecular volume |
Look at the units in that table before you write an acquisition function. volume runs to thousands of Bohr³, IR_freq sits near 1,800 cm⁻¹, and NBO_charge_H5 is about 0.5 e. Five orders of magnitude — and Section (d) is where that bites.
What you will learn
Instructions
Open the Colab notebook above and run the setup cell — it installs Chemprop and downloads the ~10 MB data bundle from the course repo. Everything then lives in one notebook with five sections, so state carries and nothing needs re-running.
Sections (a)–(c) spend no budget. They run on a free, fully-labelled 1,000-molecule dev set. That is deliberate: rehearse on free data before you queue DFT.
(a) Look at the data. Pool composition, acid subclasses, and the descriptor table with its units. Then the plot that matters: how conformational each descriptor is, using the authors' own _boltz_stdev. The carboxyl dihedrals come out on top — their _min and _max are barely the same physical quantity, and a 2D graph never sees a conformer. Commit three opinions before moving on.
(b) Choose a GNN. A quick architecture sweep, then the choice that actually matters — the predictor head. regression with one model has no sigma at all; mve alone reports only aleatoric uncertainty, which is the wrong thing to acquire on. Only an ensemble gives you a real epistemic signal. Commit a ModelSpec.
(c) Choose an initialisation. Random, MaxMin, k-means medoids, and scaffold-balanced, measured by mean pairwise Tanimoto, distinct scaffolds, and coverage distance. Then re-run random with five seeds: if the method gap is inside the seed noise, your comparison measured nothing. Report that honestly if so.
(d) Choose an acquisition function. Run random first — it is the thing to beat. Then max_variance, which should fail, and the one-line fix. Diagnose where the budget went with the per-family breakdown, look at within-batch Tanimoto to see redundancy, try BALD and core-set, then write your own.
(e) Package and submit. validate_submission.py checks the files, the budget, the log, the target order and the predict.py checksum, then actually runs your predictor on 50 public molecules. Post your dev score to the live board, upload the zip, and compare yourself against the paper's own GNN.
Do not sort the target list. Use bundle.targets verbatim. The buried-volume column names contain the substrings min and max — %Vbur_C1_min_hemisphere_3Å_min is a real column — so a stray sorted() silently scrambles twenty of them. The validator catches it; the scorer disqualifies for it.
Questions
Answer these as you work through the notebook. Discuss with your neighbour — most have no single right answer, and one of them is a trap.
After the 100-molecule seed, what is your dev scaled MAE? A model that just predicts each target's mean scores about 0.8. How much of the way from 0.8 to 0 did 100 molecules — roughly 12,000 CPU-hours — buy you?
Easymax_variance sums raw predictive standard deviations over all 156 targets and should do no better than random. Which two descriptors dominate that sum, and why does the leaderboard metric hide the damage? Check the per-family breakdown to confirm which families it starved.
These DFT labels are deterministic — run the calculation again and you get the same number. So what is the “aleatoric” uncertainty that an MVE head reports actually measuring here? Use it to explain why bald_ensemble declines to buy a molecule that max_variance_scaled happily buys.
Plot mean pairwise Tanimoto within each acquired batch for top-k uncertainty sampling and for BADGE. At ~117 CPU-hours per molecule, put a number on what the redundancy in the top-k batches costs you across the whole campaign.
MediumWrite an acquisition function that beats max_variance_scaled at round 4 rather than round 10 — early efficiency is what active learning is actually for. Then show the gap exceeds the seed-noise scale you measured in section (c). If it does not, say so: a negative result honestly reported is worth more here than a ranking you cannot defend.
Live leaderboard
Section (e) POSTs your score on the public dev set straight from Colab. It has to: you cannot compute your score on the hidden test set, because you have never seen those molecules. Resubmit as often as you like — each team's best dev_combined is shown.
The real ranking comes later. After the session every submitted checkpoint is run against a scaffold-disjoint hidden test set — 1,000 acids whose molecular skeletons appear nowhere in the pool. Teams move between the two boards, sometimes a long way. Whoever tuned hardest against dev falls furthest, and that gap is what “does your model generalise?” means in practice.
Ranked on combined = mean_sMAE + mean_ENCE, lower is better. The first term is the mean over 156 targets of MAE divided by that target's spread, so a partial charge and a buried volume count equally. The second is calibration — a constant error bar scores badly by construction.
Resources
Descriptor data from Haas, B. C.; Hardy, M. A.; Sowndarya S. V., S.; Adams, K.; Coley, C. W.; Paton, R. S.; Sigman, M. S. Digital Discovery 2025, 4, 222–233, DOI 10.1039/D4DD00284A, used under CC BY 4.0. Served via the MolSSI Descriptor Libraries API. Model: Chemprop 2.3.1 (Heid et al., J. Chem. Inf. Model. 2024, 64, 9).