Full-stack AI drug discovery

Our engine finds the target. We take it to clinic. Pharma licenses the drug.

28M PubMed abstracts — the totality of human biomedical knowledge, indexed.

Folded into a substrate that ranks the root drivers no single paper reveals. The engine is not the product. The assets are. SAPU003, in-licensed from Sapu Biosciences for pediatric neurology and longevity.

See the programs Explore the engine

Built on Qdrant Core method patent-pending — favorable WIPO ISR, nothing granted Lead program from an approved active Presented at Vector Space Day 2026

The Approach

We compress both. Timeline and uncertainty.

Speed alone doesn't fix a 1-in-10,000 funnel. Reaching the wrong answer faster is still the wrong answer.

Axis 1 — Timeline

The wrong target is what costs the years

Concept to development candidate — target, validation, formulation, preclinical package — runs the industry 4–7 years and ~$300M. The expensive part isn't the bench work. It's years of bench work spent on a target chosen from a slice of the literature. Get that call right and everything downstream gets shorter. The engine ran that half twice — SAPU003 and SAPU006, concept to clinic in 24 months for $1–3M each. Both were paid engagements for Sapu Nano, not internal projects. PDAOAI holds SAPU003 in pediatric neurology and longevity.

Axis 2 — Uncertainty

Known molecule. New delivery.

SAPU003 starts from an approved active — everolimus — reformulated. Decades of human safety already exist. The novelty is delivery and indication, not an untested molecule. Molecule risk is retired before the expensive years start.

Novel target. Known molecule. That's the leapfrog.

Repurposing is the strategy, not the fallback. When the biology is already in the literature and the molecule has already cleared humans, you leapfrog — straight past the years everyone else spends proving a new chemical entity is safe. De novo chemistry is what you do when nothing exists. It's also why most AI discovery programs still carry full molecule risk into Phase II.

Industry standard — discovery half

Time & cost4–7 years · ~$300M

TargetChosen from the literature a team can read

MoleculeNew entity — safety unknown until humans

You find outPhase II. After the money is spent

PDAOAI-assisted — discovery half

Time & cost24 months · $1–3M

Target28M abstracts folded and ranked, traceable to source

MoleculeApproved active, reformulated — safety established

You find outDay one. Checked against patient outcomes

The compression is in discovery — not in clinical trials or FDA review, and it does not replace wet-lab work. A reformulated active still requires its own clinical program. Industry ranges from Tufts CSDD, Wouters et al. (JAMA 2020) and BIO; they reflect typical experience, not a controlled comparison.

The Pipeline

One flow. Target to candidate.

Not a bag of tools. A sequence — the engine finds the root driver, patient data validates it, quantum optimization designs against it.

Step 1

Find the root driver

28M abstracts folded and ranked by structural convergence. Surfaces the gene everything else depends on — not the most-cited one.

Knowledge engine · Qdrant vector layer

→

Step 2

Validate it

Checked against public patient-outcome cohorts before anything advances. Candidates that don't survive the data don't move.

Kaplan-Meier · Public genomic cohorts

→

Step 3 · In development

Design the molecule

Generative chemistry hands you a novel entity and its safety risk back. We optimize against the root driver starting from chemistry that already cleared humans.

Quantum computing · In development

→

Output

Drug Candidate

Repurposing hit or de novo lead, ready for IND-enabling. SAPU003 and SAPU006 both came through this flow.

Fold the space → clear the ghosts → the root driver glows. Search finds the needle. We find the magnet — the thing every needle points to.

See the full method →

Proof

We asked what drives taxane resistance. The engine returned YAP1.

Run as a positive control — a question with a knowable answer, to test whether folding the corpus recovers what a domain expert knows.

YAP1

Recovered, rank 1

2.5–4× more literature sprawl than any other candidate.

Sprawl score 0.92. Reached from corpus structure alone — no paper in the set makes the full connection. Then checked against held-out patient-outcome data.

Four genes, univariate, HR 1.17–1.45, unadjusted — an association within a treated population, not a treatment-interaction test. A method demonstration in oncology, not a PDAOAI program. Full case study, with the keynote slides →

The Programs

SAPU003. Built from an approved active.

In-licensed from Sapu Biosciences. Pediatric neurology and longevity only — Sapu retains worldwide oncology and the compound IP.

Cornerstone asset

SAPU003 — IV nano-everolimus

Rare pediatric neurology (mTORopathies) and healthy aging — the two fields PDAOAI holds. Human safety and pharmacology from prior clinical use of the approved active. Orphan Drug, Rare Pediatric Disease and a Priority Review Voucher are the planned regulatory path — none filed or granted.

See the asset →

SAPU003 came out of the engine, as did SAPU006 — each concept to clinic in 24 months, both run as paid work for Sapu Nano. Two runs at that pace, so the cycle is a reproduced result, not a projection. PDAOAI's licensed field is SAPU003 in pediatric neurology and longevity.

The Business

Who buys.

PDAOAI is not a software company selling seats. It's a drug developer running its own engine.

Primary

Pharma licenses the asset

The engine finds and validates the target. PDAOAI takes the asset through IND-enabling and early clinical, then partners at the value inflection. Pharma is the buyer.

Secondary

Point the engine at their problem

Discovery groups sitting on a disease with 100+ implicated genes and no clear driver. A ranked list, traceable to source. Research collaboration, not a SaaS seat. Sapu Nano did exactly this and paid for it — twice. Both programs reached clinic.

Why it compounds

Every program sharpens the engine

Each run widens the corpus and sharpens the method — outcome data stays held out of the ranking. Assets fund the platform. The platform makes the next asset cheaper to find.

The Frame

Two companies just proved the category. Neither is doing what we do.

This isn't a bet on whether AI belongs in drug discovery. That argument closed in 2026.

The model works

Insilico took an AI-originated drug to Phase 3

Rentosertib entered Phase 3 in IPF in July 2026, and the company guided to its first profitable half. Full-stack AI drug discovery is fundable — that question is settled. Their lead asset is a novel chemical entity, which means twelve years in, molecule risk is still riding along.

The layer matters

Anthropic paid ~$400M for a nine-person team

Coefficient Bio was pre-revenue and eight months old. Weeks later Anthropic shipped Claude Science, a workbench wired into 60+ scientific databases — and said plainly it is not a better biology model. The most valuable AI lab in the world is betting the bottleneck is knowledge tooling.

One validates the business model. The other validates the layer.

Tools don't own drugs, and novel molecules don't retire safety risk. PDAOAI runs Insilico's structure — engine finds the target, we take it to clinic, pharma licenses it — on a knowledge layer of the kind Anthropic just paid $400M to go build, and against a molecule that already cleared humans. The engine is not the product. The asset is.

Insilico H1 2026 figures are per the company's preliminary, unaudited profit alert. Public-company references are for competitive context only and are not affiliations, endorsements, or partnerships.

Partnership

Built on Qdrant.

Qdrant is the vector database underneath the engine. PDAOAI held a keynote slot at Qdrant's Vector Space Day 2026 in San Francisco, co-presented with Bastian Hofmann, Qdrant's Head of Product.

Built on

"What Oncotelic has built on top of our primitives — manifold folding for biomedical literature — is one of the most ambitious applications of vector search in healthcare we have seen."
— André Zayarni, CEO, Qdrant · published statement

Watch the VSD 2026 keynote — "Building the DNA of Search" — on YouTube →