Noeris

AI for drug-induced liver injury

Predict Toxicity Early

Noeris bridges animal studies to human safety decisions. We read the preclinical data pharma already generates and tell drug developers whether a molecule will injure the human liver, while it is still cheap to change course.

Get in touch How it works

01 · The problem

Liver injury is a major blind spot.

Every drug passes a rodent study before a human sees it. Those studies are a good negative filter and a poor positive one, so liver risk is found late: in the clinic, or years after approval.

51%

Rat findings predict human liver effects barely better than a coin flip.1 Liver is among the organs animal studies predict worst.2

#1

reason approved drugs are withdrawn is liver injury: 81 of 462 withdrawals over six decades.3

Liver injury is a leading reason drugs fail in clinical trials.

U.S. Food and Drug Administration, CDER, 20244

DILI is drug-induced liver injury. IND is the Investigational New Drug application, the gate to human trials.

02 · Why now

The pipeline widened. The gold standard didn't change.

More molecules reach the safety gate every year. Every one of them is still judged by the same rat study, read by eye.

1

More drugs, same gate

+24%. Phase I drugs grew from 3,263 to 4,061 in three years.5 Safety is still judged by the same rat study, read by eye.6 It predicted human liver injury poorly in 2000, and it still does in 2025.1, 2

2

Pathology AI is already in use

Medicare pays for Imagene's AI lung cancer test, and Paige and Ibex AI tools help pathologists diagnose cancer in routine practice.7

3

The FDA is pushing for change

In 2025 the FDA's roadmap set out to cut animal testing, and ISTAND, its route for qualifying new methods, became permanent.8

03 · Approach

How Noeris is different.

Rat tissue AI trained on human outcomes. Three independent witnesses, one verdict.

Chemistry

The molecule's structure and predicted bioactivity.

Human liver cells

How human hepatocytes respond when exposed to the drug.

Rat liver tissue, read by AI

Whole-slide images from the standard rat study, read for human outcomes.

Noeris verdict

One human liver-risk score. Each witness sees different biology, so their mistakes rarely overlap.

molecule · cell · tissue · patient

A whole living organ

Chemistry sees the molecule. Cell tests see cells in a dish. Liver injury happens in a whole organ, and only a living liver shows how it responds to the drug.

Trained on human outcomes

Other pathology AI learns the pathologist's read. Ours learns what happened to patients.9

Sees what the eye can't

A slide holds billions of pixels. A pathologist turns it into a grade. The AI reads every one of them.

No new experiment

Uses the rat study every drug already runs. No added cost, time, or change to the lab workflow.

04 · Results

Same slides. Far better prediction.

Tested on 86 drugs with known human liver outcomes. Every one passed every preclinical check and reached the market, which makes them a harder subset.10, 16

Accuracy (AUC), same 86 drugs
Rat pathology read
0.55
Noeris chemistry model
0.67
Noeris
0.75
coin flip 0.50

Bars start at 0.4. An AUC of 0.5 is a coin flip, 1.0 is perfect. Today's rat pathology read scores about a coin flip.10

37 of 48

liver-toxic drugs caught at the same false-alarm rate. Pathology grades catch 17. The Noeris chemistry model, on par with the published state of the art, catches 21.10, 11

Tissue

is the witness that contributes most. Remove it and accuracy falls from 0.74 to 0.62. Remove any other witness and accuracy barely moves.12

Case study

Same rat study. Different answer.

Perhexiline, an angina drug, later withdrawn for liver injury.

Today
Rat study pathologist
No liver findings. Every treated slide graded normal.13
Noeris
Chemistry model
Looks safe. Ranked 79th of 86 for risk. On par with the published state of the art.10, 11
Human liver cells
Risk. Ranked 12th of 86.10
Rat tissue, read by Noeris AI
Risk. Ranked 15th of 86, from the same slides the pathologist graded normal.10
Noeris verdict
High human liver risk. Two of three witnesses agree.
What happened
Withdrawn for liver injury in the UK in 1985, and in most countries by 1988.14 A failure at that stage costs about $1.3B in development spend, on the industry average.15

Retrospective analysis on held-out predictions. Illustrative, not a prospective claim.

05 · Leadership

Biology, AI, and Clinical Research.

Omer Shenhar

CEO

Builder-operator: turns big problems into running organizations, fast.

  • Co-founded a nonprofit that mobilized $25M in weeks
  • Led 90+ people; advisor to Israel's Minister of Energy
  • Commanded 18 Special Forces medics (Maglan)
  • B.Sc. Biology, Hebrew University; partner at Coeus Ventures

Hebrew University · Ministry of Energy · Let's Do Something

LinkedIn

Yoav Beck

CTO

Builds medical AI that passes clinical validation.

  • 13+ years in healthcare AI
  • Clinical model with AUC above 0.90 (Donisi Health)
  • Multiple-instance learning expert
  • MRes, University College London

Donisi Health · Kohler · Sourasky · UCL

LinkedIn

Noa Davis, PhD

CSO

Has taken drugs and diagnostics from lab to clinic.

  • PhD Human Genetics, Tel Aviv University
  • Head of Drug Development: preclinical, PK/PD
  • FDA and IRB submissions
  • 15 publications

Tel Aviv University · Micromedic · CAPS Medical · Biomica

LinkedIn

06 · Contact

Working on a liver safety question?

We run blinded retrospective pilots on partners' own compounds, using the rat studies they have already completed. We would like to hear from drug developers, CROs and investors.

info@noeris.bio