Terranova: Plant Compounds for Paracetamol Analgesia

Inspiration

Paracetamol is widely used for pain relief, yet its biology involves distinct pathways. One produces AM404, which is associated with TRPV1 and cannabinoid signalling involving CB1. Another can produce NAPQI, a reactive metabolite linked to liver injury in overdose.

That contrast inspired our question: could we use computational screening to find plant-derived compounds worth investigating against targets in the analgesic pathway, while flagging chemical features that may raise safety concerns?

What it does

Terranova screens plant-derived molecules for predicted interactions with TRPV1, CB1 and FAAH. It combines those predictions with chemical similarity estimates, structural alerts and hepatic-liability information to help prioritise candidates for future experiments.

The pipeline screened 145,247 plant-derived structures and produced a shortlist of 50 hypotheses: 35 prioritised around TRPV1, 11 around CB1 and 4 around FAAH. A dashboard lets users explore compounds, scores, safety flags, plant sources and literature references. A knowledge graph connects plants, compounds, targets, pathways and evidence, distinguishing measured, literature-supported and predicted relationships.

These are computational predictions, not experimental findings. Our team has not tested the shortlisted compounds in a laboratory.

How we built it

We curated experimental bioactivity records from ChEMBL for TRPV1, CB1 and FAAH, and used LOTUS as the plant natural-product library. We standardised molecular structures and represented them with 2,048-bit Morgan fingerprints and 14 physicochemical descriptors. We trained and compared Random Forest and LightGBM models using scaffold-based evaluation.

The screening and analysis pipeline is written in Python. The local dashboard uses FastAPI with a plain HTML, CSS and JavaScript frontend. It presents precomputed results and includes an interactive compound explorer and knowledge graph.

Challenges we ran into

One challenge was data imbalance: many curated measurements were labelled active, which can make a model appear more capable than it is. Presumed-inactive decoys helped address the imbalance, but they are assumptions rather than confirmed negatives.

Another challenge was that plant chemistry differs from the medicinal chemistry represented in many training datasets. We therefore report applicability-domain estimates to show when a molecule is unlike the chemistry the model has seen.

Our target models also performed unevenly in external evaluation. ROC-AUC was 0.918 for TRPV1, 0.917 for CB1 and 0.625 for FAAH. These values measure ranking performance on the evaluated data; they do not show that a candidate will work.

Accomplishments that we're proud of

We built an end-to-end pipeline that connects public bioactivity data, natural-product screening, model evaluation, safety flags and source references. The result is a shortlist that can be traced back to plant and literature information, alongside a dashboard and a knowledge graph for exploring the evidence.

We are also proud that the project makes uncertainty visible. Measured, literature-supported and predicted links are labelled separately, and the weaker FAAH performance and inconclusive literature check are included in our reporting.

What we learned

We learned that model performance depends on data quality and coverage as much as on the algorithm. We also learned that natural-product models can be outside their area of experience, and that a high prediction score should be treated as a lead for investigation, not proof.

What's next for Terranova: Plant Compounds for Paracetamol Analgesia

Next, we want to improve FAAH data coverage, incorporate stronger confirmed-inactive examples and refine how the models handle compounds outside their applicability domain. The shortlisted candidates need target-specific laboratory assays, followed by selectivity and cytotoxicity studies.

Our goal is to make natural-product discovery more traceable, transparent and useful for deciding what to test next.