October 14, 2026 · By DataTecnica

AI Drug Discovery for Biotech: From Target Evidence to a Research Plan

How biotech and biopharma teams use AI in early drug discovery: building target evidence from human genetics, OpenTargets and screens, with sources.

AI in drug discovery covers a wide range, from generating molecules to predicting binding. This guide is about the earlier, less glamorous step that decides whether any of that effort is worth spending: building the evidence that a target is worth pursuing for a disease. That is where an AI research platform for biology, rather than a chemistry tool, earns its place.

What "AI drug discovery" means here

Most drug programs fail for reasons that were knowable early: the target was not as tightly linked to the disease as hoped, or the biology was less tractable than it looked. Target evidence work asks questions such as:

Is there human genetic support linking this gene to this disease?
What does the protein do, where is it, and how is it modified?
Which drugs and trials already exist around this target or indication?
Does functional screening data point the same way?
What does the published literature say, and how good is it?

Each question lives in a different database. The slow part of the work is moving between them. That is the part AI agents can do in parallel, while keeping every claim tied to its source.

The evidence a target case draws on

codon.bio queries these sources directly, so each part of the case comes from a named database rather than from a model's memory:

Question	Source
Genetic association with the disease	GWAS Catalogue, full GWAS summary statistics for Alzheimer's, Parkinson's, ALS, FTD and Lewy body dementia, SMR (Mendelian randomization) gene–trait associations by tissue
How tolerant is the gene to loss of function	gnomAD gene constraint and population frequencies
Known clinical variants in the gene	ClinVar
Target-disease links, known drugs, supporting evidence	OpenTargets
Protein function, location, modifications, disease links	UniProtKB/Swiss-Prot
Functional evidence from pooled screens	CRISPRbrain, CRISPRlipid and CRISPRvirus Data Commons
Published support	PubMed (through PubTator), Semantic Scholar

You can pick the sources to favour per project and per question, and each one you pick is either queried or explicitly marked as not applicable.

A target-evidence workflow

Fix the target and the indication. Be specific: a gene and a disease, not a pathway and a therapeutic area.
Start with human genetics. Ask for GWAS and SMR (Mendelian randomization) associations, and for gnomAD constraint. For neurodegenerative indications, the full summary statistics let you look past the lead hits at a gene's whole locus. Strong human genetic support is one of the more useful signals available at this stage. A gene that tolerates loss of function in healthy people tells a different story from one that does not.
Add the protein biology. Pull function, localisation and modifications from UniProt, and the known variant landscape from ClinVar.
Map what already exists. OpenTargets brings together targets, diseases, drugs and the evidence behind the associations, so you can see what is already known and who has tried what.
Look for functional evidence. If your target has been perturbed in a pooled CRISPR screen, the screen data can show whether knocking it down or up changes the phenotype you care about.
Read the papers, in context. Run the literature question against the same project, so the published evidence sits next to the database evidence.
Turn it into a plan. The point of the exercise is a decision and a next step. codon.bio keeps a staged research plan with dependencies and a critical path, and tracks each task's progress from the work done in the project's chats.

What this is, and what it is not

Being direct about scope saves time:

It is an evidence and analysis platform. It retrieves, compares and cites, and it runs real Python and R analysis on your data.
It is not a generative chemistry, docking or ADMET-prediction tool. Its built-in sources are genomic and biomedical evidence. If your question is "which molecule should we make", that is a different category of software.
It does not replace the people who judge whether the evidence is strong enough to commit resources.

For teams in biotech and biopharma, the useful framing is that it shortens the time between "we have a hypothesis" and "we have a sourced evidence summary we can argue about".

Making the case defensible

A target decision gets challenged. Two habits help:

Keep the trace. You can inspect which agent called which tool with what query, so a colleague can see how a conclusion was reached and rerun it.
Keep the citations attached. Every claim in the summary links back to its source, so the discussion is about the evidence, not about whether it exists.

Questions people ask

Can AI discover drugs on its own?

Not in the sense the headlines suggest. AI can speed up the evidence-gathering and analysis that precede a program, and some tools help design molecules. A platform for target evidence, like codon.bio, supports the first of those.

How is this different from a general chatbot?

Answers are retrieved from named biomedical sources rather than recalled from training data, every claim keeps its citation, and the full trace is inspectable. When a question needs computation, it writes and runs real Python or R.

Does it work for a team?

Yes. Teams are invitation-only, with admin, manager and member roles and a shared workspace. A project can belong to a team or be shared with individual collaborators, and chats and files can be shared with one or more teams.

Can I add my own data?

Yes: PDFs, CSVs, Excel workbooks, VCFs, code and documents, or a linked Google Drive or Dropbox folder, GitHub repository or Slack channel.

Try it on a target you know

Pick a target you already understand well, open a project in codon.bio and ask for the evidence case. Checking the answer against what you know is the quickest way to judge whether it earns a place in your workflow.