October 14, 2026 · By DataTecnica

How to Run AI-Assisted Drug Due Diligence on a Target or Asset

A practical workflow for AI-assisted drug due diligence in biopharma: test target validity, the landscape and safety signals with sourced evidence.

Drug due diligence is the work of finding out whether an asset or a target is as good as its pitch says. It happens when a company considers licensing, acquiring or investing in a program, and again internally when a team decides whether to commit to one. The scientific half of it is evidence gathering under time pressure: many sources, little time, and a conclusion that will be questioned.

This post sets out a workflow for the scientific side of that work and shows where AI agents help. It does not describe a feature called "due diligence". It describes how to use an AI research platform for biology, with sourced evidence and an inspectable trace, to do that job faster and more defensibly.

What the scientific side of diligence asks

Whatever the deal structure, the evidence questions are consistent:

Target validity. Does human biology support this target for this disease?
Landscape. What drugs, trials and competing programs already exist around it?
Safety signals. Does the genetics suggest the target is safe to modulate, or is there a reason for caution?
Biology and mechanism. Does the protein's known function fit the proposed mechanism?
Evidence quality. How strong is the published support, and who produced it?

Each question maps to specific sources, which is what makes it tractable to automate the retrieval.

A workflow, step by step

State the claim being tested. Write the asset's central claim as a sentence: "Inhibiting X reduces Y in disease Z." Diligence is an attempt to break that sentence.
Test target validity with human genetics. Query the GWAS Catalogue and SMR (Mendelian randomization) gene–trait associations for the target and the disease, and for neurodegenerative indications the full GWAS summary statistics for Alzheimer's, Parkinson's, ALS, FTD and Lewy body dementia. Genetic support is not proof, but its absence is a finding worth recording.
Look for safety signals in the population. gnomAD reports gene constraint and the frequency of loss-of-function variants. A gene that is strongly constrained against loss of function may warrant caution. A gene where people carry loss-of-function variants and are healthy can be reassuring. Interpret either with expertise; the data is the starting point.
Check the clinical variant record. ClinVar shows which variants in the gene are classified and for what conditions, which can reveal disease links you were not asked about.
Map the landscape. OpenTargets brings together targets, diseases, drugs and supporting evidence, so you can see what already exists for the target and the indication.
Confirm the mechanism. UniProtKB/Swiss-Prot gives protein function, subcellular location, modifications and disease associations to compare against the claimed mechanism.
Read the primary literature. Run the key questions against PubMed (through PubTator) and Semantic Scholar, and upload the papers or documents the seller or partner has provided so they are read alongside everything else.
Write the summary with the evidence attached. The output should be a document in which each claim links to its source, so a reviewer can check it.

Why traceability matters more here than anywhere

A diligence conclusion is read by people who were not in the room and may disagree. Two properties of the process make it easier to defend:

Each claim keeps its citation. The argument is about what the evidence shows, not whether the evidence exists.
The trace is inspectable. You can see which agent queried which source with which question, so a colleague can reproduce or extend it.

This is the main difference from asking a general chatbot, which answers from memory and gives you nothing to audit.

What it will not do

Say this plainly in the room:

It is not a verdict. It assembles and organises evidence. Whether the asset is worth the price is a judgement for experienced people.
It cannot see what you do not give it. Proprietary data-room documents are only included if you upload them. Public databases do not contain unpublished trial results or confidential preclinical data.
It does not assess commercial or regulatory factors. Its built-in sources are genomic and biomedical evidence.
Public data has gaps and lags. Treat an absence of evidence in a database as a question to ask the counterparty, not as proof.

Making it repeatable across a team

Diligence is rarely one person. In codon.bio a project holds the question, the sources, the documents and the findings, and can belong to a team with admin, manager and member roles, or be shared with individual collaborators. A staged research plan lists what remains to be checked, and tasks move to done as the work happens in the project's chats, so the open items are visible instead of living in someone's notes.

Questions people ask

Can AI do drug due diligence?

It can do the evidence retrieval and first-pass synthesis far faster than manual searching, with citations. The judgement and the decision stay with your team.

What sources does it use?

GWAS Catalogue, neurodegenerative disease GWAS summary statistics, SMR associations, gnomAD, ClinVar, OpenTargets, UniProtKB/Swiss-Prot, PubMed (through PubTator), Semantic Scholar and the CRISPRbrain, CRISPRlipid and CRISPRvirus Data Commons, plus the open web with a trust tier for each domain. It also reads web pages and PDFs you point it to, citing PDFs by page. You can pick which to favour, and each one you pick is either queried or explicitly marked as not applicable.

Can I include confidential documents?

You can upload documents to a project. Review your organisation's own data-handling requirements before uploading anything confidential.

How is it priced?

There is a free Individual plan, and paid plans for individuals, labs and enterprises. See the pricing page.

Try it on a past decision

The easiest test is a target you have already assessed. Run it through a codon.bio project and compare what it surfaces with what your team found, and what it missed.