Reference Biosciences builds physics-based tools that predict how drugs are metabolized — with the accuracy of quantum chemistry and the speed of machine learning. Know your metabolic liabilities before you synthesize.
Every year, drug candidates fail because of unexpected metabolism. Soft spots missed in early screening become toxic metabolites in the clinic. The current tools — rule-based predictions and empirical models — top out at ~75–80% accuracy and offer no insight into why a prediction was made or how much to trust it.
The result: expensive experimental cycles, late-stage surprises, and clinical holds that cost $50K–$500K per day.
We took a different approach. We started from the physics.
Reference Biosciences built the world's largest dataset of quantum-mechanically computed activation barriers for drug metabolism — over 32,000 barrier calculations across cytochrome P450, aldehyde oxidase, and FMO enzymes. We use this dataset to train delta-learning models that correct fast semi-empirical estimates to DFT-level accuracy.
Every prediction traces back to quantum mechanical transition states — not pattern matching on historical data. When chemistry changes, our predictions change with it.
Every prediction comes with a calibrated confidence interval and a transparency dashboard. You know when to trust the model and when to run the experiment.
Screen 5,000 compounds across multiple CYP enzymes in 24 hours. Get per-molecule predictions in under 30 seconds. Fast enough to fit inside your design-make-test cycle.
One platform, three tiers — from quick predictions to regulatory-grade analysis.
Paste a SMILES string. Select your enzymes. In 30 seconds, get a full regioselectivity map, ranked metabolite predictions, and a transparency dashboard showing transition state geometries, spin state analysis, and uncertainty decomposition.
Upload a compound library. Get back a ranked report: metabolic liability scores, site-specific soft spot analysis, predicted metabolites, and a prioritized list of compounds to validate experimentally.
Comprehensive QM-level characterization of all metabolic sites for a single drug candidate. Full DFT and CASSCF transition states for the highest-risk sites, with uncertainty quantification suitable for IND/NDA filings.
We believe predictions without transparency are just guesses. Every MetaboQM result traces back to real quantum mechanical calculations — and we show our work.
The largest dataset of its kind — activation barriers computed across three enzyme classes using validated DFT methods with DLPNO-CCSD(T) anchoring.
Every prediction includes a confidence interval with uncertainty decomposition. You always know when to trust the model and when to run the experiment.
We validate against established benchmarks including the Sheridan drug-like set. Publications and benchmark results will be shared as they become available.
For the computational chemist who wants to know.
SMILES → 3D conformer generation → identification of all C–H sites and heteroatom lone pairs.
GFN2-xTB single points at each site — barrier estimates, Mulliken charges, bond orders, Wiberg indices, HOMO–LUMO gaps.
ECFP4 fingerprints, pharmacophore features, electronic descriptors from xTB.
A gradient-boosted ensemble predicts the correction from xTB to DFT-quality barriers, trained on 32,000+ data points with active learning.
A classifier flags sites with likely multi-reference character (DFT spread > 3.0 kcal/mol) and applies CASSCF corrections where needed.
Template-based transition state rendering and metabolite prediction using established P450 reaction rules.
"We have 500 compounds in our SAR campaign. Which ones will have metabolic problems?"
MetaboQM Screen ranks your library and identifies exact soft spots so your medicinal chemists can fix them before synthesis.
"This scaffold looks promising but it gets cleared too fast."
MetaboQM Predict shows which sites are vulnerable and how blocking them — fluorine, deuterium, steric shielding — would change the metabolism profile.
"FDA is asking us to characterize all metabolites for our Phase II candidate."
MetaboQM Validate delivers a comprehensive QM-based metabolism report formatted for regulatory submission.
Reference Biosciences is a Boston-based computational chemistry company focused on one goal: making drug metabolism predictable.
We started with a simple observation — the tools available to DMPK teams rely on empirical pattern matching and offer no insight into prediction confidence. We believed that starting from quantum mechanical first principles and building rigorous uncertainty quantification into every prediction would produce a fundamentally better tool.
So we built one. Over 12 months, we computed more than 32,000 quantum mechanical activation barriers across three major enzyme classes, developed an automated triage system for multi-reference electronic structure, and trained delta-learning models that deliver DFT-quality accuracy in seconds.
The result is MetaboQM — a platform built by scientists, for scientists, designed to fit inside real drug discovery workflows.
We offer risk-free pilots: send us 50 compounds where you know the experimental metabolites. We predict blind. If we beat 85% accuracy, let's talk about a license. If not, you owe us nothing.