SARVault

SARVault is an exploration of cheminformatics at warehouse scale. The guiding questions are the ones a medicinal chemistry team working on cytotoxic ADC payloads keeps asking: how does potency rank across the series, where are the activity cliffs, which analogues exist, what does the chemical neighbourhood look like?

The raw material is public - ChEMBL, UniChem, PDBe - but fragmented across sources with different identifiers, formats, and update cycles. Integrating heterogeneous sources into one consistent, queryable model is the same problem that dominates any enterprise informatics deployment. SARVault works this problem end to end. Every step of the pipeline is reproducible.

Scale is not what makes this hard. Under two thousand compounds fit in memory on any laptop. What makes it hard is that the same molecule wears a different name in every source, while a potency value means nothing without the assay it came from. Both problems stay invisible until an answer is already wrong. A join that silently misses is worse than a join that fails.

27 dbt models carry the material from raw API responses through staging into marts and an analytics layer, orchestrated as a Dagster asset graph with the dbt tests surfaced as asset checks. UniChem is read from the bulk mapping dumps rather than the live API, which proved the more reliable route. PDBe ligand codes are resolved to concrete co-crystal entries, so the viewer above shows a real structure rather than a promise.

A ranked list of the pairs worth looking at. 296 activity cliffs across the target set, ordered by steepness rather than by similarity alone, each rendered side by side with both structures.

A map of the chemical space. 1,868 compounds projected to two dimensions from their fingerprints, coloured by potency, clickable through to the underlying series.

A bridge from a molecule to its crystal structure. 370 links from a compound to a PDB entry, resolved from ligand codes, which is what feeds the 3D viewer at the top of this page.

Nine views over one warehouse. Overview, SAR, selectivity, series, space, library, cliffs, payload comparison and data quality, all reading the same marts.

The scope is deliberately narrow. Four targets, chosen around cytotoxic and tubulin-targeting chemistry, the space behind ADC payloads and classical chemotherapeutics. This is not a general-purpose bioactivity warehouse and does not pretend to be.

There is no predictive model. SARVault describes and organizes, it does not predict potency. That is a decision rather than a gap: the question here was whether public data can be made trustworthy enough to reason over, not whether a regressor can be fitted to it.

Snowflake is prepared, not proven. The same dbt models build against a second profile, which has not been run against a live account.

The data is public and the code is MIT. Everything on this page can be reproduced from the repository, against ChEMBL release 36.