AI Drug Discovery
AI that finds better drug candidates from far less data.
Built for pharma, biotech and drug-discovery teams who cannot afford to screen everything before they know what is worth screening.
The Problem
Before AI is useful in drug discovery, someone has to pay for the data.
More than nine in ten drug candidates still fail in the clinic. The industry's answer was more data and more compute. That has not moved the number. A large screening campaign can run to more than 100,000 compounds before a shortlist even exists, against a high-throughput screening market that is growing fast.
High-throughput screening market global
Grand View Research, High Throughput Screening Market (2024–2030), published August 2024.
Most AI drug-discovery tools still need that kind of scale to work: large labelled datasets, expensive physical screening, months of effort before anyone has something worth pursuing.
The Data Tax
It falls hardest on the organisations with the least room to pay it: mid-sized pharma, biotech companies, the regional pharma ecosystem, neglected-disease programmes, and venture studios evaluating early assets.
The core pain is not just accuracy. It's the upfront cost of producing enough data for conventional AI to work.
The Approach
TERNIA finds the winners before the expensive screening starts.
An engine that works from a single small assay. No pre-training. No million-dollar screen first.
-
Works from small data
A ranked shortlist from one small assay, not a million-compound screen.
-
Knows when it doesn't know
A confidence flag on every call, so teams act on the strong ones and set the rest aside.
-
Days, not months
Early triage that moves a programme forward while conventional screening would still be running.
The outcome: lower early-stage screening costs, and months of lead discovery cut to days.
What is proven, and what is not
The results so far, stated plainly.
TERNIA was tested on a public benchmark built to defeat models: the real hits are rare, and the decoys look like them.
Single-assay benchmark Tox21 SR-ARE
We tested TERNIA on an extremely difficult public dataset: the SR-ARE assay from the Tox21 Challenge. SR-ARE (Antioxidant Response Element) measures activation of the Nrf2 pathway, which cells use to fight oxidative stress.
Latest run
PRC-AUC
0.001.00
Base rate: 0.1599 (random classifier)
Deep Learning models require massive pre-training across millions of historical, multi-task data points just to learn feature representations. TERNIA uses information geometry to extract its features directly and exclusively from the target assay, eliminating the need for massive pre-training databases.
ROC-AUC
0.001.00
Base rate: 0.1599 (random classifier)
TERNIA beats every model in the Tox21 Challenge without being pre-trained on multi-assay data. Tox21 Challenge winner on SR-ARE scored 0.840 (source: Measuring AI Progress in Drug Discovery: A Reproducible Leaderboard for the Tox21 Challenge)
Accuracy
0.001.00
No data enrichment has been performed, whether using information from other assays in the Tox21 database or external pre-trained data. All runs have exclusively used the information in the individual assays.
Delivering actionable enrichment
Latest run
Top 0.5%
Top 1%
Top 2%
06.3
Top 5%
Top 10%
06.3
Enrichment factors (higher = better, 1.0 = random)
On blind Tox21 SR-ARE splits, TERNIA achieved in the reported runs the maximum theoretical Enrichment Factor in the Top-1 and near-perfect in the Top-2 without external pre-training. †
† Depending on the split and stochastic elements.
Physics inspired performance
Latest run
Enrichment factor, by depth
06.3
Precision
0.001.00
By deploying a multi-agent swarm to discover natural activity cliffs, TERNIA returned 15 real hits in its Top-20, an enrichment factor of 4.7, proving that our physics-inspired platform can outperform Deep Learning models. Across the compounds it scored as reliable, 72.7% of its positive calls were real hits. ††
TERNIA has a higher PRC and strong Enrichment Factor compared to traditional methods, making it a powerful driver to improve ROI.
While Tanimoto’s global PRC is buoyed by its performance in the middle of the distribution, wet-lab realities dictate that usually the very top of the list is actionable (EF@1% or 2%).
†† Published Tox21 benchmarks report ROC-AUC primarily. Per-task PRC-AUC is rarely reported due to the extreme class imbalance of many assays in this database. The base rate of about 12% to 16% for SR-ARE set the random-classifier floor.
What the difference buys you
Three things standard AI can't do.
-
A clear signal, not noise
A typical screen with a 10 to 20% hit rate leaves chemists guessing why some molecules worked and similar ones didn't.
Practical output TERNIA separates molecules that look alike but behave differently, so a chemist stops guessing which one to carry forward.
-
Genuinely new chemistry
Most tools surface near-copies of known drugs, landing you in crowded, heavily-patented space.
Practical output TERNIA finds molecules that look different but behave the same in the assay, so what it hands you is not a near-copy of a drug that already exists.
-
Tune it to your goal
Run it for precision when you need a few sure leads, or for reach when you want candidates further from what's already known.
Practical output One engine tuned to a pipeline's actual priorities, not a fixed setting.
How it works
The mathematics is public. Anyone can test it.
Standard AI learns from pairs of examples, which is why it needs volume before a pattern holds. TERNIA's agents reason about molecules three at a time. Past a threshold, they self-organise: the structure comes from the interactions between agents, not from the size of the dataset behind them.
We do the physics. Most others curve-fit the data.
Publicly available in two preprints from a team with a track record of peer-reviewed work elsewhere.
Eduardo Salazar (2025, rev. January 2026), Introducing COGENT3: An AI Architecture for Emergent Cognition (arXiv:2504.04139). Eduardo Salazar (2026), Composite-Operator Scaling on Triadic Hypergraphs (arXiv:2604.27038).
The Field
Everyone else needs the data first.
| Player | Approach | What it needs to work |
|---|---|---|
| Atomwise | Structure-based deep learning | Large training data |
| Optibrium | ADME, toxicity and property prediction | Curated property datasets |
| Insilico Medicine | Generative chemistry | Large training data |
| Schrödinger | Physics-based modelling and ML | Heavy compute |
| TERNIA | Physics-native, single assay | None of the above |
Approaches summarised from public company materials. A fuller landscape is available on request.
How we get paid
Paid to prove it. Then paid to license it.
-
Year 1
Paid pilots
Clients pay TERNIA to run the engine on their own targets, at $100,000 to $300,000 a pilot.
-
Year 2 and beyond
Licences and CRO deals
Pilots that work convert into ongoing revenue.
-
In parallel
Grant co-funding
Neglected-disease partnerships fund open validation and provide compound access, on a non-dilutive basis.
We reach clients three ways: direct pilots with pharma and biotech, introductions through our advisors’ trusted networks, and neglected-disease and public-health partnerships that build open proof points and give access to compound libraries. Discovery conversations are under way with pharma and biotech R&D teams, weighted towards biotech and regional partners on shorter contracting cycles. We’re targeting three pilots in 2026
Who is building it
Technical depth, commercial execution, and scientific credibility.
Two founders, twenty years working together. Advisors with deep AI and mathematics pedigree.
-
Dr Eduardo Salazar
Co-Founder & Lead Technology Architect
Mathematician and economist, forty years in complex systems. Built and coded the TERNIA engine.
-
Kofi Osei-Ntem
Co-Founder & Strategy Lead
Medicinal chemistry, strategy and business development background. Leads investor engagement and go-to-market.
-
Dr Sergio Alvarez-Teleña
Advisor, AI
Computer scientist, founder of SciTheWorld, a UCL spin-out. Algorithmic-trading pioneer; OECD GPAI.
-
Prof. Hugo Scolnik
Advisor, Computational Mathematics
Optimisation and cryptography pioneer. Founded Latin America's first computer science department.
-
Dr Haresh Mirchandani
Advisor, Pharma Strategy & BD
Thirty-year pharma veteran, formerly VP of BD at J&J. Led licensing and acquisitions across therapy areas.
Get in touch
Nobody should take our word for it.
Hand us a proprietary assay and we will run it blind.