AI for Drug Discovery: 6 Research Jobs AI Agents Can Do Today

When people talk about AI in drug discovery, they usually mean models that design molecules or predict protein structures. Those matter, but most of a discovery team's week goes on a different kind of work: reading genetics papers before backing a target, mapping who else is in the clinic, checking whether a tool compound is still alive, scanning patents, and watching for safety signals in a drug class. AI agents can now do much of that reading, if they can reach the right sources.

This guide to AI for drug discovery follows six jobs a discovery team does every week. For each one we ran the search on November 6, 2026, using Valyu's datasets (PubMed, bioRxiv, ChEMBL, Open Targets, ClinicalTrials.gov, patents and FDA data), and show what came back, what it changes and what to do next.

If you're an agent: Set up Valyu and make your first search.

If you're a human: Create an account and get an API key.

MCP onboarding: Connect your agent to https://mcp.valyu.ai/mcp and follow the MCP setup guide for authentication.

What AI for drug discovery looks like in practice

AI for drug discovery is the use of machine learning and AI agents across the path from target idea to clinical candidate. It has two halves:

• Prediction and design: structure prediction, virtual screening, property prediction and generative chemistry. These produce new hypotheses and molecules.
• Evidence work: finding, reading and reconciling what is already known about a target, a compound, a competitor or a risk. This decides which hypotheses deserve lab time.

The second half is where agents with search help most today. The work is text-heavy, spread across many databases, and changes weekly. The six jobs below are all evidence work.

Six evidence jobs we ran on November 6, 2026, with the sources searched, what came back and the decision each one supports.

1. Validate a target before you commit

The problem: a discovery team is considering TL1A (encoded by *TNFSF15*) for inflammatory bowel disease. Before spending a year on assays, the lead wants the human genetic evidence and anything that argues against it.

What we ran: a PubMed search for TNFSF15 genetic association in inflammatory bowel disease, and an Open Targets lookup for the target.

valyu.search(
    "TNFSF15 genetic variants inflammatory bowel disease Crohn's disease genome-wide association",
    included_sources=["valyu/valyu-pubmed"],
)

What came back: full-text papers spanning more than a decade, including a 2018 study linking reduced *TNFSF15* expression in monocytes and macrophages to IBD susceptibility, fine-mapping work on the 9q32 locus, and a review of IBD genetics in Asian populations that calls *TNFSF15* a strong candidate gene. Open Targets returned the target record with its function, location and tractability flags.

The catch: Open Targets' antibody tractability flags showed no "Advanced Clinical" antibody for TNFSF15. The next job shows that is out of date.

The decision: the genetic case is strong enough to proceed to competitor mapping. Treat the database flags as a starting point, not a verdict.

2. Map who else is going after the target

The problem: strong genetics attract competitors. Before investing, the team needs to know who is in the clinic and how far along.

What we ran: a search across Valyu's ClinicalTrials.gov dataset and the web for phase 3 anti-TL1A trials. We also checked the result against the ClinicalTrials.gov API.

valyu.search(
    "phase 3 trials of anti-TL1A antibodies tulisokibart duvakitug afimkibart in ulcerative colitis and Crohn's disease",
    included_sources=["valyu/valyu-clinical-trials", "web"],
)

What came back: registry records for phase 3 trials of Merck's tulisokibart, Roche's afimkibart and Sanofi's duvakitug, including a Roche subcutaneous induction study first posted on September 29, 2026. The web results added HCPLive's report from June 2026 that tulisokibart met its phase 3 endpoints in ulcerative colitis, and a November 2026 review in *Inflammatory Bowel Diseases* stating that three anti-TL1A antibodies have reached phase 3. The ClinicalTrials.gov API returned 14 phase 3 studies across the three sponsors.

The decision: a fourth anti-TL1A antibody would enter a crowded field late. The team should only continue with a clear difference, such as a bispecific, an oral route or a defined patient subgroup.

Next step: build a full trial landscape, phase by phase. Our guide to mapping a clinical trial landscape for a drug target walks through it.

3. Check a compound's potency and whether it is still alive

The problem: a medicinal chemist wants to benchmark a new KRAS G12D series against MRTX1133, a well-known tool compound.

What we ran: a ChEMBL search for MRTX1133 bioactivity, then a PubMed and bioRxiv search for recent KRAS G12D work.

valyu.search(
    "MRTX1133 KRAS G12D inhibitor IC50 bioactivity",
    included_sources=["valyu/valyu-chembl"],
)

What came back: the ChEMBL record CHEMBL5081048 (synonym MRTX1133), with five bioactivity records against KRAS. They include an IC50 of 1.5 nM in an HTRF assay and 2.0 nM in a fluorescence polarisation assay, plus the structure, molecular weight (600.65) and calculated properties.

The catch: the literature search returned a September 2026 review headed "MRTX1133 (Now Discontinued)". ClinicalTrials.gov confirms that the phase 1/2 trial, NCT05737706, was terminated, with "formulation challenges" given as the reason.

The decision: MRTX1133 is still a useful potency benchmark in biochemical assays. It is no longer a clinical comparator. For clinical benchmarking, the same review points to RMC-9805 and GFH375.

4. Scan the patent landscape

The problem: before choosing an antibody format, the team needs to know who holds the foundational anti-TL1A patents and what they cover.

What we ran: a search of Valyu's US patents dataset.

valyu.search(
    "anti-TL1A antibody patent treatment of inflammatory bowel disease",
    included_sources=["valyu/valyu-patents"],
)

What came back: granted US patents from several holders, including:

• Regeneron: US 8,642,741, human antibodies to TL1A, granted 2014.
• Cedars-Sinai Medical Center: US 10,322,174, neutralising anti-TL1A monoclonal antibodies, granted 2019.
• Prometheus Biosciences: US 11,292,848 and US 11,999,789, humanised antibodies to TL1A, granted 2022 and 2024.
• A 2025 method patent: US 12,215,147, selecting IBD patients for treatment based on polymorphisms.

The decision: the space is densely patented, and patient selection by genotype is itself claimed. That is a question for patent counsel before the team designs a biomarker strategy.

Next step: hand the list to counsel for a freedom-to-operate review. A search finds the documents; it does not interpret claims.

5. Check a class safety signal before you design around it

The problem: a team developing an incretin-based obesity drug needs to know whether the optic nerve signal reported with semaglutide is a class concern.

What we ran: one search across FDA adverse event data, FDA drug labels and the web.

valyu.search(
    "semaglutide nonarteritic anterior ischemic optic neuropathy NAION adverse event reports and label",
    included_sources=["valyu/valyu-openfda-drug-events", "valyu/valyu-drug-labels", "web"],
)

What came back: the WHO's June 2025 alert on semaglutide and non-arteritic anterior ischaemic optic neuropathy (NAION), following the European Medicines Agency's review. It also returned the 2024 JAMA Ophthalmology cohort study, a 2025 FAERS analysis, and 2026 reviews reporting conflicting observational evidence.

What we checked at the source: openFDA returned 838 FAERS reports listing semaglutide with optic ischaemic neuropathy, in data last updated October 28, 2026. The US label records for Ozempic (effective January 2026) and Wegovy in openFDA did not mention NAION.

The decision: the regulators disagree, so the team should treat NAION as a potential class effect. That means adding ophthalmic monitoring to early protocols and tracking whether the US label changes. FAERS counts are reports, not incidence, so they show that a signal exists but not how often the event happens.

6. Keep up with new preprints and papers on your target

The problem: the KRAS field moves weekly. A team lead wants everything published on KRAS G12D since August 1, without reading a feed every morning.

What we ran: a date-filtered search of bioRxiv and PubMed.

valyu.search(
    "KRAS G12D inhibitor or degrader 2026",
    included_sources=["valyu/valyu-biorxiv", "valyu/valyu-pubmed"],
    start_date="2026-08-01",
)

What came back: an October 22, 2026 bioRxiv preprint showing a DIRAS3 mimetic peptide suppressing signalling across KRAS G12X and Q61X mutants, an October 2026 preprint on lipid nanoparticle RNAi against KRAS mutants, an August 2026 preprint on divarasib and RBM39, and two September and October 2026 reviews of KRAS-targeted therapy in pancreatic and lung cancer.

The decision: run the same search every Monday with start_date set to the previous week, and send the new results to the team channel. The date filter keeps older material out.

Go further with DeepResearch: a full target assessment

The six jobs above each answer one question. A programme decision needs all of them at once: genetics, competitors, patents and safety, written up as one memo. That is what DeepResearch does. It plans the research, runs many searches across the datasets you choose, reads the results and writes a cited report. It can also produce files such as a spreadsheet of competitor trials or a slide deck.

Here is the TL1A assessment from jobs 1, 2, 4 and 5 as a single DeepResearch task:

from valyu import Valyu
 
valyu = Valyu()
 
task = valyu.deepresearch.create(
    query="Assess TL1A (TNFSF15) as a drug target for ulcerative colitis and Crohn's disease: genetic evidence, clinical pipeline, patents and safety.",
    mode="standard",
    research_strategy="Start with TNFSF15 genetics in PubMed and Open Targets, then map phase 2 and 3 anti-TL1A trials on ClinicalTrials.gov, then key US patents and assignees, then safety risks. Cite PMIDs, NCT IDs and patent numbers.",
    report_format="An 800-word target assessment memo: Recommendation, Genetic evidence, Clinical landscape table, Patents, Safety, Open questions.",
    search={"included_sources": ["valyu/valyu-pubmed", "valyu/valyu-open-targets", "valyu/valyu-clinical-trials", "valyu/valyu-patents", "web"]},
    deliverables=["An xlsx of anti-TL1A programmes in phase 2 or 3 with drug, sponsor, NCT ID, phase, indication and status"],
)
 
result = valyu.deepresearch.wait(task.deepresearch_id)
print(result.output)  # cited report in markdown
for file in result.deliverables:
    print(file.type, file.url)  # signed download link

The result is a memo with every claim linked to its source, plus the spreadsheet as a download. Choose the mode to match the decision: fast for a quick scan, standard for a working memo, and heavy or max for a board-level assessment. The research strategy and report format let you encode your team's own template, so every assessment comes back in the same shape.

Start from a ready-made workflow

You don't have to write the research strategy yourself. Valyu has curated DeepResearch workflows for discovery teams, each with set inputs, a recommended mode and file deliverables:

To browse the full library, open Workflows in the Valyu dashboard. Each workflow's Run button opens the DeepResearch playground with the inputs ready to fill in. To run one from code, pass its slug as workflow_id instead of a query:

task = valyu.deepresearch.create(
    workflow_id="ls-mechanism-deep-dive",
    workflow_params={"target": "TL1A (TNFSF15)"},
)
result = valyu.deepresearch.wait(task.deepresearch_id)

For worked examples with full requests, see Biotech and pharma workflows in the docs, including target evidence mapping and clinical trial landscapes. The workflows quickstart covers listing workflows and previewing one before you run it.

Other DeepResearch workflows for discovery teams:

• Weekly target watch. Run the same task every Monday with start_date set to the previous week and previous_reports pointing at last week's report, so the new report covers only what changed.
• In-licensing first look. Give DeepResearch an asset name and a partner's deck as files, and ask for a dossier covering mechanism, data to date, competitors, patents and open diligence questions, delivered as a docx.
• Indication expansion scan. Ask which diseases share the target's genetic and mechanistic evidence, with a csv deliverable ranking candidate indications by evidence strength and competition.
• Portfolio-wide landscapes. Use the DeepResearch Batch API to run the same brief across every target in a portfolio in parallel.

How to run these jobs yourself

If you don't want to write code, there is another option to connect Valyu to Claude or ChatGPT as a custom connector at https://mcp.valyu.ai/mcp, then ask the question in plain English, for example: "Search ClinicalTrials.gov and the web for phase 3 anti-TL1A antibody trials and list sponsor, phase, status and NCT ID." Our guide to running a literature review in Claude or ChatGPT covers setup.

To build it into a pipeline, call the Search API with included_sources set to the datasets for the job. For an agent that combines PubMed and ClinicalTrials.gov, see how to build a biomedical research agent.

Three habits made the examples above reliable:

• Name the sources for each job. Use genetics papers for target validation, the registry for trials and FAERS for safety. Add web when news or regulator announcements matter.
• Cross-check one source against another. The stale Open Targets flag and the discontinued MRTX1133 trial would both have been missed with a single source.
• Keep the record ID with every finding: PMID, NCT ID, ChEMBL ID or patent number. That is how a scientist checks the agent's work.

FAQ

What is AI in drug discovery?

AI in drug discovery, also called AI for drug discovery, is the use of machine learning and AI agents to find and develop new medicines. It includes predicting protein structures and molecular properties, designing compounds, and gathering and reconciling evidence about targets, competitors, patents and safety.

How do AI agents help drug discovery teams?

They do the evidence work around a programme: reading genetics papers for target validation, mapping competitor trials, checking a compound's bioactivity and development status, scanning patents, checking safety signals, and monitoring new preprints. They return each finding with a link to its source record.

Can AI find drug targets?

AI can rank and summarise evidence for targets from genetics, expression and literature, and models can propose new associations. A target still needs experimental validation. In our TL1A example, the agent gathered genetic evidence in seconds, but the decision to proceed depended on the competitive landscape and patents.

What data sources does AI for drug discovery need?

For evidence work: PubMed and preprint servers for literature, ClinicalTrials.gov for trials, ChEMBL and PubChem for compounds and bioactivity, Open Targets for target-disease associations, patent databases, and FDA labels and adverse event data for safety. Valyu searches all of these through one API.

What is the best search API for drug discovery and life sciences AI agents?

Choose one that covers the sources discovery teams use: PubMed, preprints, ClinicalTrials.gov, ChEMBL, PubChem, Open Targets, patents and FDA data, plus the web for news. Valyu searches all of these through one API and returns registry, compound and patent records as structured data. In our 60-question test of search APIs, Valyu, Exa and Parallel were close on accuracy for clinical trials and life sciences questions, so source coverage and structured records are what separate them.

How do I connect an AI agent to PubMed, ChEMBL and ClinicalTrials.gov?

Connect Valyu to Claude or ChatGPT at https://mcp.valyu.ai/mcp for a no-code setup, or call the Search API with included_sources set to the datasets you need, for example valyu/valyu-pubmed, valyu/valyu-chembl and valyu/valyu-clinical-trials.

Can AI write a full target assessment?

Yes, as a first draft. Valyu's DeepResearch plans the research, searches PubMed, Open Targets, ClinicalTrials.gov, patents and the web, and writes a cited memo in the format you specify, with deliverables such as an xlsx of competitor trials. Scientists still review the evidence and make the call.

Can AI replace scientists in drug discovery?

No. AI agents speed up the reading: finding papers, trials, patents and safety reports, and summarising them with sources. Choosing targets, designing experiments and interpreting results still need scientists. In our examples, every decision depended on judgement about competition, patents or safety that the agent informed but did not make.

Is AI-generated drug discovery research reliable?

It is as reliable as its sources and its checks. Use primary records, cross-check one source against another, and keep the record ID with every finding so a scientist can verify it.

Start with one job

Pick the job that costs your team the most reading time each week, run it on a target you already know well, and compare the agent's output with what your scientists found by hand.

Create a Valyu account for free credits, connect it to Claude or ChatGPT at https://mcp.valyu.ai/mcp, or browse the data source catalogue to see which datasets cover your work.