Skip to content

Science

Off-target risk,predicted before trials.

Faster, cheaper, and more precise: predict why a candidate would fail before the money is spent — traditional programs find out after.

4–6 weeks

to a de-risked computational lead · Epitorix estimate

Versus 3–5 years to a candidate by the traditional route
Our in-house models

Two paths to a candidate

Same target. The same question, asked earlier.

The difference is when you learn a molecule will fail: after 3–5 years and $5–10M, or after 4–6 weeks and up to $100K of in-silico design and de-risking.

From target to a preclinical candidate

up to65×

faster from target to candidate

4–6 weeks versus 3–5 years, to a candidate ready for preclinical testing

up to100×

lower cost from target to candidate

up to $100K versus $5–10M, to a candidate ready for preclinical testing

  • Traditional approach

    3–5 years

    from target to a development candidate

    $5–10M

    cost to a development candidate

    • Preclinical development
    • Clinical trials begin
    • Toxicity discovered late

  • Epitorix

    4–6 weeks

    from target to a de-risked computational lead

    up to$100K

    cost to a de-risked computational lead

    • Predict biodistribution & PK
    • Identify organ & cell-type risk
    • Optimize candidate before clinic

Epitorix estimates. ‘Preclinical’ means a designed, ranked and de-risked candidate ready to enter preclinical testing; in vitro and in vivo studies, GLP toxicology, GMP manufacturing and regulatory review follow before any first-in-human study.

HER2 extracellular domain with trastuzumab Fab · PDB 1N8Z · rendered from deposited coordinates

How the models work together

Design, de-risk, advance.

The same models design the candidate, predict where it goes and what it hits, and rank what to test next. Every product runs on them.

DESIGNEPiQ™ · XenocleaDE-RISKPharmPhax™ · PharmacoLawADVANCETherapeutic Pipeline01Target02Binding-siteanalysis03Generation~10,000site- & target-specific candidates04Affinityranking~1,000/smolecules rankedby predicted Kd05Optimization06Off-target~1 hourproteome-wide07Biodistribution& organ riskup to 93%liabilities detected,retrospective analyses08In vitro10+ labsEpi-labs,United States09In vivovalidation10Translational& regulatory11Clinic
  1. Target — Design (EPiQ · Xenoclea).
  2. Binding-site analysis — Design (EPiQ · Xenoclea).
  3. Generation — Design (EPiQ · Xenoclea). ~10,000: about 10,000 site- and target-specific candidates.
  4. Affinity ranking — Design (EPiQ · Xenoclea). ~1,000/s: about 1,000 molecules per second, ranked by predicted Kd; Pearson R 0.84 for peptides in internal benchmarks.
  5. Optimization — Design (EPiQ · Xenoclea).
  6. Off-target — De-risk (PharmPhax · PharmacoLaw). ~1 hour: the whole proteome in about one hour, over 90% recall on known off-target interactions.
  7. Biodistribution & organ risk — De-risk (PharmPhax · PharmacoLaw). up to 93%: organ, cell-type and biodistribution risk; up to 93% accuracy detecting reported development and safety liabilities in retrospective analyses of radiopharmaceuticals with documented clinical outcomes.
  8. In vitro — Advance (Therapeutic Pipeline). 10+ labs: Epi-labs — over 10 labs across the United States.
  9. In vivo validation — Advance (Therapeutic Pipeline).
  10. Translational & regulatory — Advance (Therapeutic Pipeline).
  11. Clinic — Advance (Therapeutic Pipeline).
Figures from retrospective analyses and internal benchmarks; a preprint describing the evaluation is forthcoming.

Head to head

No display library. No million compounds. A shortlist worth testing.

Site-specific ligand discovery has meant mRNA or phage display and libraries of millions of compounds. Epitorix's re-ranking model returns a shortlist ordered by confidence — the candidates most worth taking to the bench first.

  • Display librarymillions of compounds · months
  • Epitorix re-rankinga shortlist · confidence-ordered

  • 92.4%

    Small molecules

  • 78–100%

    Peptides

    canonical 81% · macrocyclic 78% · non-canonical 100%

  • 78%

    Mini-proteins

  • 86%

    Antibodies

  • 73%

    RNA ligands

  • 56%

    DNA ligands

Hit rate = shortlisted candidates confirmed to bind at < 10 nM · Epitorix re-ranking model · internal benchmark

The re-ranking model runs inside EPiQ and Xenoclea.

Inside EPiQInside Xenoclea

PSMA (GCPII, PDB 4NGM) turning among ten real ligands of ten modalities — trastuzumab Fab 1N8Z · RNA aptamer 1F1T · octreotide 7T11 · nanobody 1MEL · PSMA-1023 (5O5R) · exendin-4(9–39) 3C59 · DNA aptamer 148D · de novo miniprotein LCB1 7JZU · somatostatin-14 7T10 · JB7 (4NGM) — rendered from deposited coordinates
An intact Y-shaped IgG as a pale bone ribbon, hinge in slate blue, a small violet chelator on a hinge lysine, and a ringed inset at lower left magnifying it.
An intact IgG at atomic resolution with a single Lu-DOTA chelator modelled onto its most exposed hinge lysine, drawn at true relative scale — the chelate is about one nanometre across, the antibody fifteen times that; the inset shows the same chelate magnified two and a half times · PDB 1IGT, chelator built in RDKit

Our in-house models

We are not just combining published models. We have our own.

Epitorix's in-house models are its own architectures, trained on data we generate ourselves and validated at the bench — for peptides, antibodies, nanobodies and mini-proteins. They run beside published models in EPiQ — 30+ models in all, selected per target rather than defaulted to one — and every product runs on them.

The same models in EPiQ

100%

of designs taken to the bench bound their target

In internal benchmarks; the number of designs, targets and the assay threshold are not yet published.

30+

models in EPiQ, published and in-house

0.84

Pearson R, predicted vs measured pKd — peptides

in internal benchmarks

Targeted generation

Ten thousand candidates, one binding site.

Targeted generation is de novo design against one chosen binding site: new backbones and sequences sampled by generative models rather than selected from a library. Phage and mRNA display take months and return noise. Epitorix generates ~10,000 site- and target-specific candidates — about six months saved, success rates up about 60% versus display libraries, in internal benchmarks.

AI in radiopharmaceutical design

~10,000

site- and target-specific candidates

~6 months

saved generating high-quality candidates

versus display libraries

~60%

success-rate increase

versus display libraries, in internal benchmarks

Octreotide as a slate-blue backbone tube with ring side chains, seated in the cut-open pocket of somatostatin receptor 2, a bone surface, lining in violet.
Octreotide in the orthosteric pocket of somatostatin receptor 2, its D-Trp and Phe side chains making most of the contact and its disulfide closing the ring; the pocket is shown cut open — the darker faces are the cut — and the two hairlines mark the lipid bilayer · PDB 7T11

Off-target prediction

Off-target binding, proteome-wide, in about an hour.

Off-target prediction is the screening of a candidate against proteins other than its intended target, to find the binding that causes toxicity in healthy tissue. A microarray is limited by what is on it. Epitorix screens binders, peptides and small molecules against the whole proteome computationally, in about an hour — over 90% recall on known off-target interactions.

Off-target binding, explained

>90%

recall on known off-target interactions

in retrospective analyses

  • NCAM1
  • SSTR1
  • CXCR4
  • ATRNL1
  • PRKG1
  • BCL2
  • FXYD2
  • PLPP1
Proteome-wide, schematic · cell-surface and intracellular proteins · ~1 hour · >90% recall on known off-target interactions
56789105678910Measured pKdPredicted pKdleast-squares fity = xPEARSON R · PEPTIDES0.84n = 140 · illustrative
Illustrative distribution for R = 0.84; not experimental data. pKd = −log10 Kd. Internal benchmark; preprint forthcoming.

Affinity ranking

Binding affinity, Kd at scale.

Binding affinity is how tightly a ligand holds its target, measured as the dissociation constant Kd: lower is tighter. SPR and BLI measure it in months; molecular dynamics and FEP+ take hours per ligand pair. Epitorix predicts Kd at about 1,000 molecules per second — Pearson R 0.84 for peptides, 0.70 for small molecules, in internal benchmarks.

Kd, pKd and binding affinity, explainedpKd calculator

0.70

Pearson R, predicted vs measured pKd — small molecules

in internal benchmarks

~1,000

molecules per second

>80%

of true top ligands recovered

from 100 generated, in internal benchmarks

Biodistribution and organ risk

Biodistribution and organ risk, before trials.

Biodistribution is where a radiopharmaceutical goes in the body and how much reaches each organ over time — with PK and PD, which healthy tissues pay for the dose. Toxicity often surfaces after trials begin. Epitorix predicts biodistribution, organ and cell-type risk, PK and PD first — detecting previously reported liabilities with up to 93% accuracy in retrospective analyses.

PharmPhax, the report built on thisBiodistribution and dosimetry, explained

up to93%

accuracy detecting reported development and safety liabilities

in retrospective analyses of radiopharmaceuticals with documented clinical outcomes

Predicted organ and cell-type risk for one candidate
OrganCell typePredicted risk
Salivary glandsSalivary acinar cellsLow
LiverHepatocytesModerate
KidneysProximal tubular cellsHigh
MarrowHematopoietic stem & progenitor cellsModerate

The same models run our own programs.

The four programs now in vivo were designed with the models on this page.

See the pipelineBring us a target

Figures from retrospective analyses and internal benchmarks; a preprint describing the evaluation is forthcoming.

Contact

Bring us a target.

A target to design against, a candidate to de-risk, or the pipeline data.

What are you working on?

Privacy policy

Laboratory bench · liquid handler and assay plates