Evidence maps
Diseases, genes, targets, molecules and studies in one sourced map, with the contradictions left visible and every inference marked as such.
In preparation - access by invitation
DNOVA by NEO AI
Biotech intelligence, delivered with its proof.
DNOVA is the NEO AI platform being built to open computational biology to biotech teams, investors and research partners. Evidence maps, target dossiers, model benchmarks and asset reviews, prepared with LILI and each delivered with its sources, its limits and a proof anyone can verify.
Artistic illustration, not a structure. Drag to turn it.
Most AI in biotech hands you an answer. DNOVA will hand you the answer, the data it came from, the runs that produced it, and what it cannot prove.
8capabilities in preparation on one platform
4independent proof statements on every deliverable
50+public data sources in the connector scope, each qualified before use
1AI brain, LILI, shared with every NEO AI vertical
The platform
What DNOVA is being built to deliver
Eight capabilities on one platform, sharing the same sources, the same proof model and the same AI. Each one is in preparation: nothing here is sold or available today.
Target dossiers
Supporting and contradicting evidence for a target, a prior-art search that states its sources, and a validation plan. Never a clinical prediction.
Molecules and targets
Structures, targets and activity measures with units, species and protocol. IC50, EC50, Ki and Kd are never merged; predicted and experimental structures are always told apart.
Image model benchmarks
How a segmentation or counting model really behaves on biological images, compared on identical protocols with data split by plate, laboratory or donor.
Cell model evaluation
Whether a single-cell or perturbation model generalises to contexts it has never seen, against simple baselines, with its uncertainty.
Clinical and competitive radar
Trials and programmes by disease, target, molecule, sponsor and stage, with every change of a source dated.
Asset reviews
A scientific, competitive and operational review of a biotech asset, with the missing documents and the questions to ask. Facts, company claims and our assessment stay separate.
Partner workspaces
Private project spaces for partner laboratories, reviewers and auditors, with roles and permissions per action, project and data category.
Who it is for
Built for the people who must decide on evidence
01
Biotech and pharma teams
Targets, molecules and models examined on public data, with the limits stated before the decision, not after.
02
Investors and funds
Asset reviews that separate what is documented, what is claimed and what is assessed, with the open questions listed.
03
Research laboratories
Benchmarks on held-out data, partner workspaces and external validations recorded with their protocol and their limits.
04
Institutions and funders
Reproducible dossiers that a reviewer or an auditor can check without trusting us.
The engines
Three research engines from NEO Bio n Tech
DNOVA runs on the computational research programmes of NEO Bio n Tech, the research vertical of NEO AI. Each asks narrow, measurable questions of existing public data and states what was computed, what would need a laboratory, and what remains unknown.
Illustration - synthetic data
- Sourced link
- Inferred by a model
- Contradicting evidence
Evidence maps and testable hypotheses
Discovery Atlas
Diseases, genes, proteins, molecules, pathways, trials and studies, linked by relations that each carry a source, a version, an evidence type and their limits.
Questions it answers
- Which lines of human genetic, expression and intervention evidence support or contradict a target for a defined indication?
- Which hypotheses are under-explored once a documented prior-art search has been run?
- How much of an apparent consensus comes from independent studies rather than the same dataset reused?
What it produces
- Sourced evidence maps with dates and versions
- Testable hypotheses with counter-arguments and a validation plan
- Evidence dossiers with a sealed proof bundle
Limit we state up front. The absence of results in the sources consulted does not prove that no prior art exists.
Raw imageModel masksDrag the line
Illustration - synthetic data
How image models really behave
BioVision
Segmentation, counting and morphology on biological images, compared on exactly the same protocol, with data split by plate, laboratory or donor rather than by random image.
Questions it answers
- Does a segmentation model keep its accuracy on images from a laboratory it has never seen?
- Do popular benchmarks overstate performance because training and test images share plates or sources?
- Which failure modes dominate, and at what cost?
What it produces
- Benchmarks with confidence intervals, compute time and cost
- Failure cases kept and shown, not filtered out
- Reproducible reports from real results
Limit we state up front. No improvement is announced before it is measured on a locked test partition.
Illustration - synthetic data
Predictions that must generalise
Cell Intelligence
Single-cell and perturbation data analysed with documented statistics, and models compared against simple baselines on perturbations or cell contexts they have never seen.
Questions it answers
- Does a candidate model beat no-change, mean and linear baselines on held-out perturbations, within uncertainty?
- Which signatures stay consistent across studies, donors and batches?
- Where do models fail: by cell type, by perturbation class, by dataset?
What it produces
- Baseline comparisons with confidence intervals
- Benchmark studies, negative results included
- Model cards recorded in the Proof Ledger
Limit we state up front. Proximity in a projection is a visual aid, not a mechanism.
How an engagement works
From your question to a result you can check
Every engagement follows the same path. Each step leaves a dated, signed trace, and the steps that need a human decision cannot be skipped by software.
- 01
Frame the question
A named scientist and your team agree on a precise, testable question. Prior work is searched and the sources consulted are listed.
- 02
Fix the protocol and the rights
Comparators, metrics, data splits, stopping rules and data rights are versioned before the final test data is touched.
- 03
Compute with provenance
Every run records its data versions, code, model, parameters, seeds and environment. Failed runs and negative results are kept.
- 04
Review, seal, deliver
A reviewer other than the author evaluates a precise version. The dossier is sealed and time-stamped, and you receive a proof you can verify yourself.
LILI, the AI scientist
An AI that shows its work
LILI is the one brain of NEO AI, shared with every vertical. In DNOVA it prepares protocols, searches sources, launches pipelines and drafts dossiers, and it shows you the plan, the sources, the tools and the cost before and after it acts.
Separate roles
Search, computation, critique and writing are separate agents. Agreement between agents is not an independent validation.
Rights checked on the server
Every tool call is checked against your permissions. Text found in articles and files is treated as untrusted data.
A human signs
A specialist reviewer signs every scientific conclusion that leaves the platform.
Ask LILI to
- Compare three targets for a disease
- Explain contradictory evidence
- Launch an image model benchmark
- Compare cellular responses
- Prepare a reproducible dossier
- Explain why a result is not ready to share
>
Plan
- Pull genetic, expression and intervention evidence for each target
- Deduplicate studies reused across databases
- Map supporting and contradicting links
- Draft a comparison with its limits
Sources consulted
Open TargetsGWAS CatalogGTExChEMBLClinicalTrials.govEurope PMC
Tools called
atlas.evidence_map() · studies.deduplicate() · graph.contradictions()
Proof Ledger
Four statements, never one green badge
Every deliverable carries four separate statements. They are never merged, because a result can be intact and still unreproduced, or reviewed and still unpublished.
Technical integrity
Files match their recorded fingerprints.
Reproducibility
Whether the result was re-run, by whom, and to which tolerance.
Scientific validation
Computational evidence, external experiment or independent replication, stated as such.
Publication status
Internal, preprint or peer reviewed, with its version history.
How a dossier is sealed
Every file of a dossier is fingerprinted. The fingerprints are combined into one root, signed, time-stamped by a third party and anchored on a public network. Anyone holding the files can check them.
A blockchain anchor proves that a fingerprint existed at a date. It does not prove that a result is true, new, or owned.
Illustration. Fingerprints are computed live in your browser.
Scientific integrity
What computation can and cannot establish
| Question | By computation | Needs a laboratory |
|---|---|---|
| Is a gene associated with a disease in the consulted sources? | Yes, with the sources, versions and populations stated | No |
| Does modulating a target change disease biology? | Only as a hypothesis with its supporting and contradicting evidence | Yes, in a recorded protocol with a partner laboratory |
| Does a segmentation model generalise to a new laboratory? | Yes, on held-out public data split by source | Only if no independent public data exists |
| Does a predicted perturbation response hold in real cells? | Agreement with held-out public measurements | Yes, for new cell systems and new perturbations |
We do not operate a physical laboratory. When a hypothesis needs an experiment, it is tested with partner laboratories under a recorded protocol, and the independence of that test is stated.
Data universe
Built on the open infrastructure of biology
Publications, genetics, molecules, proteins, cells, images, trials and ontologies, reached through official interfaces and downloads. Each source is qualified for access, licence and rights before it is used.
Publications
PubMed · PubMed Central · Europe PMC · BioStudies
Diseases and genetics
Open Targets · GWAS Catalog · ClinVar · gnomAD · GTEx · Ensembl · NCBI Gene · RefSeq · NCBI Datasets · Human Protein Atlas
Molecules
ChEMBL · PubChem · BindingDB · ChEBI
Proteins and pathways
UniProt · EMBL-EBI Proteins API · RCSB PDB · AlphaFold Database · InterPro · Reactome · STRING
Cells and perturbations
GEO · SRA · ENA · CELLxGENE Discover · CELLxGENE Census · scPerturb · Tahoe-100M · LINCS · Connectivity Map · DepMap · CCLE · Open Problems · MetaboLights · MGnify
Biological images
BioImage Archive · IDR · BBBC · Cell Painting Gallery · JUMP Cell Painting · RxRx
Trials and public health data
ClinicalTrials.gov · AACT · openFDA · GDC
Ontologies and patents
OLS · Gene Ontology · HPO · MONDO · Cell Ontology · EFO · EPO OPS · Lens
Sources in the connector scope. A name on this list does not mean a partnership, an endorsement or a tested integration. Restricted sources stay restricted.
Built on NEO AI
One brain, separate data, full audit
One LILI
DNOVA uses the same LILI as every NEO AI vertical, through a versioned gateway. No second brain, no provider keys of its own.
Your data stays yours
DNOVA keeps its own database, users, tools and permissions. Partner and client data never feeds research or models without a specific written right.
Central Super Admin
Configuration, users, rights, quotas, budgets, suspension and audit, from the shared NEO AI console. Every exceptional access is justified and logged.
Accounts by invitation
No open sign-up. Partners, reviewers and auditors are invited on defined projects, with roles and permissions per action, project and data category.
Access
DNOVA opens by invitation
For biotech teams, investors, partner laboratories, reviewers and auditors working on defined projects. Tell us who you are and what you work on, and we will come back to you.