Live with first pharma partner

Pick the right molecule, before you make it.

Albion TechBio builds dynamics-trained neural networks for medicinal chemistry. We rank congeneric series in minutes, so chemists know which compounds to make next, before the synthesis cycle starts.

FEP+ Benchmark
0.497
Weighted-mean Pearson on 199 compounds, 8 targets. Second only to a fully supervised baseline trained on ~98k labelled complexes.
Inference
<1.5s /cmpd
CPU-only. No GPU cluster, no infra rewrite. Drops into existing pipelines.
Affinity Labels
0
Stage 1 trains on physics, not on binding measurements. Generalises to novel targets.
00Mission

We build neural networks that capture how molecules move.

Drug binding is a dynamic process. Most computational scoring still treats it as a still image, learning from frozen crystal structures. Albion trains on molecular dynamics, so its rankings reflect the physics of binding rather than a memorised snapshot. That is what makes the model hold up on targets it has never seen.

01The Problem

Static structure is the wrong training signal for binding.

What AI sees today

One frozen conformation per crystal.

X-ray crystallography gives one snapshot, one conformation.
  • ~200,000 crystal structures, one frame each
  • Memorises shapes, not the physics of interaction
  • Fails out-of-distribution on novel targets
What actually happens

Thousands of conformations, real motion.

Molecular dynamics gives thousands of frames of real physical behaviour.
  • Trained on motion, not snapshots
  • Learns the physics of binding
  • Generalises to novel targets
02Technology

A graph neural network trained on physics simulations, not labelled data.

The Albion engine learns from molecular dynamics trajectories, the same simulations medicinal chemists already trust to ground-truth their FEP+ runs. The result is an architecture that scores in seconds on commodity CPU and stays accurate when a target has never been seen before.

Trained on motion, not measurements.

Most scoring models are trained on labelled binding affinities. Labelled data is sparse, biased toward well-studied targets, and gives the model no information about why a ligand binds, only how strongly it bound in one experiment.

Albion takes the opposite path. We run molecular dynamics on the protein and ligand together, then train the network to recover physically meaningful targets from the trajectory. Stage 1 sees zero affinity labels. The model learns what physically stable binding looks like, then ranks unseen ligands by how close they get to it.

Built for the medicinal chemist's workflow.

Inference runs in under 1.5 seconds per compound on a laptop CPU. A 500-compound congeneric series is ranked in minutes. The output is a per-compound priority with a confidence interval, so the team can hand the top 30 to FEP+ or directly to synthesis.

Boltz-2, IsoDDE, and Schrödinger's FEP+ all need a GPU cluster. Albion drops into the existing chemistry pipeline as a CPU API. That is what makes it the integration partner, not the procurement battle.

01 · Input

Candidate molecules from chemistry

200 to 500 compound congeneric series. SMILES with docked or cofolded poses. Standard medicinal chemistry inputs, nothing exotic.

02 · Albion engine

Physics-trained GNN ranks all of them

16-checkpoint ensemble. ~1.5s per compound on commodity CPU. No GPU. Zero affinity labels in stage-1 training.

03 · Output

Ranked list. Which to make next.

Per-compound priority with confidence intervals. Hand the top 30 to FEP+ or to synthesis. The chemist stays in the loop.

Physics
Learns from what is physically possible
No labels
Zero affinity labels in stage-1 training
Speed
Under 1.5s per compound, CPU only
03Results

Every competitor was given the answers. Albion was not.

On the FEP+ congeneric benchmark and on three held-out novel targets, Albion is competitive with or beats methods trained on 15 to 469 times more labelled data. On the targets where static methods sit at chance, Albion ranks first.

FEP+ congeneric benchmark
0.497

Weighted-mean Pearson PCC across 199 compounds and 8 targets, 95% CI [0.393, 0.594]. Second only to AEV-PLIG augmented (0.59), which trains on roughly 98,000 labelled crystal complexes. Albion uses none.

Out-of-distribution
#1on novel targets

On three held-out novel targets, Albion ranks first against every supervised baseline tested. Where static methods sit at chance because they have never seen the target before, dynamics-trained scoring continues to discriminate.

FEP+ benchmark · Schrödinger congeneric set · 199 compounds · 8 targets
Novel-target evaluation · three held-out targets · zero training exposure
04Platform

Two production configurations. One shared encoder.

Same GNN backbone, two heads. One ranks compounds for synthesis priority. The other catches the cases where cofolding tools confidently produce physically implausible poses. Both run on CPU.

Albion-A · Affinity

Congeneric ranking and pre-FEP triage.

Rank a 200 to 500 compound series in minutes. Hand the top 30 to 50 to FEP+. Cuts alchemical compute by an order of magnitude while preserving the compounds most likely to advance.
Ensemble
5 checkpoints
Signal
MD-stability, mean-rank
FEP wmPCC
0.497
Inference
<1.5s / compound, CPU
Albion-S · Stability

Cofold-QC and adversarial pose detection.

Every AlphaFold-3, Boltz-2, Chai or IsoDDE cofold gets a reliability score in milliseconds. Catches the cases where cofolding confidence inverts on physically implausible ligands, before they reach chemistry.
Ensemble
16 checkpoints
Signal
MD-derived stability
Adversarial
Correct on 3 / 3 classes
Inference
Milliseconds / pose, CPU
05Team

Built by people who have lived this problem.

A founder who has run drug-discovery operations and built the model from first principles, advised by a senior medicinal chemist with 25-plus years of pharma experience.

MF
Mitchell Frizelle
Founder & CEO
  • Chemist with operational drug-discovery experience inside a venture-backed biotech.
  • Built the Albion model end to end, from first principles.
06Get in touch

If you run a small-molecule programme, we should talk.

We are working with chemistry teams running one to four lead programmes. Pilots take a single congeneric series, run it through Albion, and produce a ranked list inside a week. If the answer is good, we go from there. If it is not, you have lost a week.