Inspiration
A plant extract inhibits a parasite. It contains hundreds of compounds. Which molecule did it?
That question inspired DECOUPLE. When candidate compounds appear together across active extracts, the existing screen may not distinguish their contributions. Isolating a candidate too early risks spending limited laboratory resources on the wrong explanation. Potency alone can also prioritise compounds that harm host cells.
We wanted to help researchers answer a more useful question first: Which experiment should we do next?
Built for Track 2: Natural Products & Drug Discovery, DECOUPLE turns uncertainty into a concrete recommendation for follow-up testing.
What it does
DECOUPLE is an interactive research workbench for choosing the next extract to assay under a limited budget.
Researchers select a compound pair, inspect the recommended extract, and reveal archived assay outcomes one at a time. After each reveal, the system updates four competing explanations: compound A, compound B, both, or neither.
The workbench brings together:
- Candidate abundance profiles and extract provenance.
- Parasite inhibition and L6 host-cell toxicity measurements.
- Model probabilities and recommendations for the next assay.
- A final nomination, shortlist, joint explanation, abstention or cytotoxicity flag.
- A comparison against strongest-first selection and an exportable decision trace.
These outputs are hypotheses for experimental follow-up. A joint explanation does not establish synergy, and a nomination does not confirm a compound’s identity, safety or therapeutic value.
How we built it
We built the analysis and decision engine in Python, using NumPy for numerical computation and RDKit for chemical information. The interface uses HTML, CSS and JavaScript, with Geist typography and a restrained visual design.
Our data pipeline uses the ENPKG archive, which provides chemistry and assay records from a 1,600-extract plant library. We prepared 12 measured-data replay episodes that judges can explore locally without an account, API key or full dataset download.
The model uses continuous log-abundance information with a detection floor. The interactive workbench chooses assays by their expected contribution to a confident nomination. We separately implemented and evaluated adaptive contrast, which selects extracts with informative differences in candidate abundance while penalising redundant choices.
We built reproducibility checks, provenance records and automated tests alongside the prototype. AI coding assistance supported development; the scientific decision engine uses explicit statistical rules rather than an LLM.
Our pitch uses Manim, FFmpeg and ElevenLabs narration. Its demonstration reconstructs captured engine states from archived measurements; it is not a live browser recording or a new laboratory experiment.
Challenges we faced
Separating co-occurrence from causation. Compounds that travel together can support the same explanation. We had to identify useful contrasts without treating an undetected signal as proof that a compound was absent.
Accounting for toxicity. Strong parasite inhibition is insufficient when the same candidate may also explain host-cell toxicity. We incorporated both readouts into nomination.
Evaluating without known compound-level truth. Archived extract measurements do not establish which compound caused an effect. We therefore kept measured-data replay separate from a semi-synthetic benchmark that combines real co-occurrence structure with simulated pharmacology.
Communicating uncertainty clearly. A probability can look more authoritative than it deserves. We worked to distinguish measurements, model outputs and simulated results throughout the interface, documentation and presentation.
Results and what we learned
Our semi-synthetic benchmark contains 952 episodes from 238 compound pairs.
On the hard stratum—58 episodes from 16 pairs—adaptive contrast and strongest-first each nominated the true driver in 19.0% of episodes, compared with 8.6% without additional assays. The 95% pair-clustered interval around each 19.0% estimate was 6.9–34.4%. Adaptive contrast did not outperform strongest-first on these hard cases.
Under an assumed 50% prevalence of cytotoxic drivers, selectivity-aware nomination reduced simulated wasted isolations from 32.35 to 0.21 per 100 episodes, while maintaining the same overall rate of correctly nominating a developable driver: 32.7%.
Those are simulation results, not measured laboratory savings. Separately, in the measured screen, 7 of 15 extracts above 80% parasite inhibition failed the host-toxicity criterion, illustrating why selectivity matters.
Our main lesson was that the decision to nominate deserves as much attention as the decision about what to measure next. We also learned that abstention can be a useful scientific outcome, and that a familiar baseline must be tested fairly.
What we’re proud of
We built a working prototype with inspectable recommendations, measured-data replay, baseline comparison and an exportable evidence trail.
We also made the project reproducible: public source code, installation instructions, data-processing scripts, dataset references and automated checks are included. All 52 local tests passed, and the published repository passes GitHub CI.
Most importantly, we report where our approach ties the baseline and where its evidence remains limited.
Impact and next steps
DECOUPLE aims to help researchers allocate scarce follow-up assays and avoid premature isolation decisions. Demonstrating that benefit requires prospective testing.
Our next steps are to:
- Compare recommendations against strongest-first in a blinded laboratory study.
- Expand evaluation beyond the small hard stratum.
- Test sensitivity to model assumptions and cytotoxic-driver prevalence.
- Confirm candidate identities and effects using purified compounds and mixture experiments.
Responsible research and attribution
ENPKG and the published screening findings belong to Gaudry and colleagues. Our contribution is the decision-support implementation, evaluation and prototype workflow.
Compound identities remain mass-spectrometry annotations. Benchmark pharmacology is simulated, and simulation and inference share modelling assumptions. DECOUPLE is not clinically validated and does not establish medical efficacy or safety.
Dataset and font licences, provenance and media credits are documented in the repository. We retain attribution to published traditional-knowledge sources, claim no ownership over that knowledge, and do not assume that public availability establishes permission for downstream development.
Team: Panos Georgakopoulos and Apostolos Dedeloudis.
