Elman · Drug development AI you can audit, end to end

The problem

AI for drug development has four
compounding failure modes

Each failure compounds the next. Stitched together, they produce predictions that look credible and aren't.

Priors bias predictions

16% of top-20 oncology targets flip when names are revealed.

Cuccarese (Recursion VP), 2026 · public disclosure

Elman→grounded in extracted features, not priors

Memorisation hides as reasoning

Up to 87% of medical pre-training content persists through fine-tuning.

Memorization in LLMs in Medicine, 2025.

Elman→claims trace to source arm + figure / table / sentence

Brute-force averages bury the signal

Phenotype-driving cells: <1% of a sample. Bulk -omics averages them out.

Van de Sande et al., Nature Reviews Drug Discovery, 2023.

Elman→resolves signal at cell, tissue, sub-population level

Depth doesn't scale

Reading thousands of papers at depth breaks on cost, time, and orchestration debt.

Elman→hundreds-to-thousands of papers and in-silico runs, at depth

Approach

Three properties.
One assembled intersection.

Each component is replicable. The intersection is what holds.

01 · Full provenance

Every edge resolves to a source

Every claim points to a paper figure, in-silico run, or experiment ID. The model never paraphrases from memory; it reads what it cites.

02 · Rational pruning

Edge-by-edge, not bulk-average

Reasoning agents chase specific questions through the graph, surfacing the low-magnitude signals at cell and sub-population scale where disease-driving biology sits.

03 · Compounding

One schema. Every domain.

A fact extracted in cancer is reusable in autoimmune, neuro, ageing. Graph grows ~88 arm-level facts per discovery cycle.

McKinsey · 2024 ~32% of pharma have scaled gen AI; ~5% report consistent differentiation.

Verifiable output at scale is what unblocks the 95%.

Live demo

See the engine run.

A step-by-step replay of one real discovery. Thousands of siloed datasets narrowed to one validated target. Press play or step through it yourself.

Open full screen ↗

The product suite

Six products. One platform.

In production today

FTO

Live

Claim-level freedom-to-operate.

Per-component decomposition with a claim-level evidence trail on every verdict, not a similarity score on the whole filing.
Defensible under IP-committee review: one isolated agent per patent, no cross-contamination.

See it running ↓

Verification

Live

Auditable claim validation.

Source-span resolution on every claim: paper sentence, trial arm, or figure.
Audit-grade depth across hundreds-to-thousands of papers and in-silico runs per claim, with pharma-specific provenance.

See it running ↓

Elman FTO Studio: a completed freedom-to-operate analysis showing 589 patents searched and ranked by risk, with an executive summary and a per-patent risk table. — **FTO Studio.** A completed freedom-to-operate run: 589 patents ranked by risk, every high-risk family traceable to its source. *Hover a card to switch; click to enlarge.* **Elman Verifier.** A hypothesis disproved, with the supporting and refuted edges and the full knowledge graph. *Hover a card to switch; click to enlarge.*

Elman Verifier: a hypothesis validation report that disproved a hypothesis, showing the supporting and refuted edges and the full knowledge graph. — **FTO Studio.** A completed freedom-to-operate run: 589 patents ranked by risk, every high-risk family traceable to its source. *Hover a card to switch; click to enlarge.* **Elman Verifier.** A hypothesis disproved, with the supporting and refuted edges and the full knowledge graph. *Hover a card to switch; click to enlarge.*

FTO Studio is open to new users. See how it works and request a pilot →

The rest of the suite

Hypothesis Generation

What testable hypotheses fall out of this dataset?

Patentability

What in here is novel and defendable?

Bioinformatics-at-scale

What does the raw cohort data reveal at single-cell resolution?

Effect Simulation

If I modify a target (add, inhibit, or combine), what happens downstream?

In active development

Applied use-cases

Nine use-cases.
Composed from the suite.

Available today

FTO assessment

↳ FTO

Patentability assessment

↳ Patentability

Verification audit

↳ Verification + Bioinformatics-at-scale

Target identification

↳ Hypothesis Generation + Verification + Bioinformatics-at-scale

Drug-drug interactions

↳ Hypothesis Generation + Verification + Bioinformatics-at-scale

Via scoped pilots

Biomarker discovery

↳ Hypothesis Generation + Verification + Effect Simulation + Bioinformatics-at-scale

Drug repurposing

↳ Hypothesis Generation + Verification + Effect Simulation + Bioinformatics-at-scale

Trial cohort selection

↳ Hypothesis Generation + Verification + Effect Simulation + Bioinformatics-at-scale

Combination design

↳ Effect Simulation + Bioinformatics-at-scale

The full picture

Drug development, end to end.

Five iterative steps from data to therapy. We push verifiable AI into every step of the way.

Step 1

Collect datasets

Step 2

Produce hypotheses

Step 3

Find targets

Step 4

Model effects

Step 5

Engineer mitigation

Wet-lab data

Traction

Partner. Platform. Product.

Three proof points: a research collaboration, a graph in use, a product customers pay for.

Partner

Allen
Institute

A joint research collaboration on Alzheimer's single-cell data, now concluded. The Allen Institute's preprint names Elman and links to our output. Read the preprint.

Platform · today

~170k

Arm-level experimental facts in the graph, reusable across cancer, autoimmune, neuro, ageing.

Product · live

Live

FTO in live IP-committee decisions: claim-level evidence per verdict.

Founders

Two senior technical co-founders.

Working at the intersection of biology, AI, and architecture.

Dr. Moustafa Khedr

Co-founder · biology + architecture

linkedin.com/in/khedrm

PhD, Crick / UCL: 3D iPSC skeletal-muscle constructs for Duchenne muscular dystrophy modelling and ASO drug testing.
Buy-side biological-DD on gene-therapy assets: mechanism-risk and animal-model interpretation.

Dr. Francesco Moramarco

Co-founder · AI / engineering + architecture

linkedin.com/in/fm2g11

PhD in medical-text generation with large language models, University of Aberdeen.
Research Scientist (NLP), UK clinical-AI scale-up; live medical-note generation from inception to production deployment.

Together at Deep Science Ventures

Co-architects of Elman, the platform behind the Allen Institute collaboration and the FTO products in production today.

Drug development AI
you can fully audit.

AI for drug development has four
compounding failure modes

Priors bias predictions

Memorisation hides as reasoning

Brute-force averages bury the signal

Depth doesn't scale

Three properties.
One assembled intersection.

Every edge resolves to a source

Edge-by-edge, not bulk-average

One schema. Every domain.

See the engine run.

Six products. One platform.

In production today

FTO

Verification

The rest of the suite

Hypothesis Generation

Patentability

Bioinformatics-at-scale

Effect Simulation

Nine use-cases.
Composed from the suite.

Available today

Via scoped pilots

Drug development, end to end.

Collect datasets

Produce hypotheses

Find targets

Model effects

Engineer mitigation

Partner. Platform. Product.

Two senior technical co-founders.

Dr. Moustafa Khedr

Dr. Francesco Moramarco

Let's build verifiable AI
for drug development, together.

Drug development AIyou can fully audit.

AI for drug development has fourcompounding failure modes

Priors bias predictions

Memorisation hides as reasoning

Brute-force averages bury the signal

Depth doesn't scale

Three properties.One assembled intersection.

Every edge resolves to a source

Edge-by-edge, not bulk-average

One schema. Every domain.

See the engine run.

Six products. One platform.

In production today

FTO

Verification

The rest of the suite

Hypothesis Generation

Patentability

Bioinformatics-at-scale

Effect Simulation

Nine use-cases.Composed from the suite.

Available today

Via scoped pilots

Drug development, end to end.

Collect datasets

Produce hypotheses

Find targets

Model effects

Engineer mitigation

Partner. Platform. Product.

Two senior technical co-founders.

Dr. Moustafa Khedr

Dr. Francesco Moramarco

Let's build verifiable AIfor drug development, together.

Drug development AI
you can fully audit.

AI for drug development has four
compounding failure modes

Three properties.
One assembled intersection.

Nine use-cases.
Composed from the suite.

Let's build verifiable AI
for drug development, together.