Arvee BioPharmaTechnologies

Computational Discovery

Computation to narrow the experiment, not replace it.

A lean, computation-first approach to early drug discovery. The aim is to make the experimental loop narrower and cheaper, and to check every prediction against real measurements.

In development

The workbench

One loop, end to end

Biology, evidence, structure, molecules, simulation, ranking, experiment and feedback, run as a single loop. The experiment is the step that turns software into discovery.

  1. Biology

    Disease mechanism, genetic and pathway evidence, known drugs, failures and resistance.

  2. Evidence

    Public structures and activity data (ChEMBL, BindingDB, PDB), literature and patents.

  3. Structure

    Experimental structures where available, predicted models where needed; pocket identification and preparation.

  4. Molecules

    Known chemical libraries first, then similarity search and QSAR / machine-learning models.

  5. Simulation

    Docking, molecular dynamics and binding-energy estimates.

  6. Ranking

    Combining model outputs with explicit uncertainty estimates.

  7. Experiment

    A small set of compounds tested in a real biochemical or cell assay.

  8. Feedback

    Predicted against observed: recalibrate, retrain and repeat.

Current work

A blinded methods benchmark on public data

The first project is deliberately unglamorous. We take a well-characterised target with extensive published ligand data, hold back part of the measurements, and test whether the pipeline recovers the known structure–activity relationships.

The questions are simple and measurable:

  • Can it separate potent compounds from weak ones?
  • Can it tell actives from inactives?
  • Can it flag selectivity against a related protein, or poor ADMET properties?

Tools

Open-source stack

Established, inspectable tools and public data sources, so that methods can be reproduced and checked.

  • RDKitCheminformatics
  • OpenMMMolecular simulation
  • AutoDock VinaDocking
  • Boltz-2Structure and affinity prediction

Known limits

Docking can produce plausible binding poses while ranking real biological activity poorly, and target selection and biological validation remain hard problems even with modern AI models. We treat computation as a filter on what to test. The assay decides.

Longer term, the focus is chronic disease.

Interested in methods, data or assay collaboration?

We are open to conversations with researchers and laboratories working on similar questions.

Get in touch