From molecules
to better decisions.
Three working research prototypes. Explore chemical representations, challenge a model's predictions, and choose what to measure next.
Loading reproducible research assets…
Molecular Similarity Explorer
RDKit · 24 curated structuresHow much does a small structural change matter? Rank a curated molecular library by Morgan fingerprint similarity, inspect physicochemical descriptors, and export your comparison.
Method, contribution and limits
RDKit calculates eight descriptors and 2,048-bit Morgan fingerprints with radius 2. Similarity is the Tanimoto intersection-over-union of fingerprint bits. This project adds a ranked, filterable comparison workflow and a reproducible asset generator on top of RDKit. Structures and descriptors are precomputed, so the demo works without sending molecular inputs to a server.
The curated library is deliberately small. Some structures omit stereochemistry. Similarity does not establish biological activity, safety, efficacy or synthetic accessibility. cLogP is a calculated descriptor. No drug-discovery model is trained in this explorer.
Solubility Validation Lab
Measured ESOL data · held-out evaluationA model should earn trust on unfamiliar chemistry. Compare a random split with a scaffold-group split, check a simple baseline, and inspect whether prediction intervals actually cover the held-out measurements.
Descriptor importance
Random-forest impurity importance, not causal attribution.
What the benchmark tests
A 250-tree random forest uses eight RDKit descriptors. A training-mean predictor supplies the baseline. Canonical SMILES duplicates are removed before splitting, keeping the first measurement. Calibration data determine 90% split-conformal intervals; the test set is used only for final evaluation.
Scaffold groups use Bemis–Murcko structures, with all acyclic compounds in one group. Group splits target 60/20/20% of groups, not exact molecule counts. Scaffold separation deliberately changes the chemical distribution: nominal conformal coverage is not guaranteed under this shift. This is a fixed-seed baseline study, not evidence of state-of-the-art performance or prospective validation.
Data provenance and reproducibility
Source: Delaney's ESOL dataset as distributed through DeepChem. Original publication: Delaney, 2004. The Python script accepts a locally supplied CSV, records its SHA-256 checksum and package versions, and rebuilds the displayed evaluation. Raw benchmark data are not bundled in the source package.
Active Learning Sandbox
Synthetic simulation · runs in your browserSpend the next experiment where it can help. A Gaussian process updates after each observation. Change the exploration weight to balance predicted performance against uncertainty.
Blue: posterior mean and pointwise 95% model bands. Dark dots: observed experiments. Maximum 25 observations per run.
Experiment log
| Step | Candidate x | Observed score |
|---|
Acquisition rule and simulation assumptions
Choose the unmeasured candidate with the highest upper confidence bound: μ(x) + βσ(x). The Gaussian process has a zero mean, unit-amplitude RBF kernel with length scale 1.1, and diagonal numerical jitter 10⁻⁶. It starts with observations at x = 0, 5 and 10 and chooses from 81 equally spaced candidates.
The hidden objective is sin(0.8x) + 0.45cos(1.9x) − 0.08x, without measurement noise. This is an illustrative one-dimensional simulation, not molecular property prediction. The kernel settings are fixed; the bands describe this model's uncertainty and are not experimentally calibrated. The original browser implementation includes numerical tests.