TL;DR
AI in drug discovery already works in six applications that narrow what scientists test, and what is ahead is clinical proof and clearer rules. The six are target identification, virtual screening, generative molecular design, ADMET prediction, drug repurposing and patient stratification. They attack a slow process, where discovery and development take 10 to 15 years and about 90% of clinical candidates fail. The evidence is real but narrow, led by AlphaFold’s 2024 Nobel Prize and a 2025 phase 2a trial of the generative-AI-discovered drug rentosertib. FDA issued draft guidance in January 2025 and agreed shared principles with EMA in January 2026, yet higher clinical success rates are not yet shown. Teams that start with one measurable use case and clean, governed research data are best placed for what comes next.
Key Takeaways AI in drug discovery helps at six steps today, so it touches everything from picking targets to choosing patients. Its real job is to rank and filter options so scarce lab time goes to the most promising compounds. Protein structure prediction is the clearest proven win, recognized by the 2024 Nobel Prize in Chemistry. A generative-AI-discovered drug has completed a randomized phase 2a trial, but no study yet shows AI raises clinical success rates. Under FDA’s January 2025 draft guidance and the joint January 2026 principles, regulators expect a defined context of use, credible models and documented data. What is ahead is clinical proof, biological foundation models and closed-loop labs, and all of them depend on governed research data. Watch on YouTube
Can AI Transform Healthcare: Trends and Technologies to Watch
Kanerika walks through where AI is changing healthcare and life sciences, which technologies are maturing and what leaders should watch next.
Five Thousand Compounds, Two Hundred Lab Slots Picture a discovery team with 5,000 candidate compounds and enough assay capacity to test 200 of them this quarter. The chemists have instincts and the database has years of past results, but the budget will not stretch further. Someone has to decide which 4% of the list reaches the bench.
That choice is exactly where AI helps. A model trained on past assay data can rank all 5,000 by predicted activity and risk. The 200 slots then go to the best bets. It cannot replace the assay itself, because a prediction is a hypothesis until the lab confirms it.
This guide covers where that ranking works today, what the evidence shows, how regulators treat it and what comes next.
What Is AI in Drug Discovery and Where Does It Fit? AI in drug discovery means using machine learning to make research choices faster and with better odds. Those choices run from which protein to target to which molecule to make next. It works mainly on the early half of the pipeline, before any human receives a dose, although one application, patient selection, reaches into trial design.
That boundary matters because the two halves face different problems. Discovery runs from target identification to a preclinical candidate, while development covers clinical trials, manufacturing and approval. Many articles blur the two, so claims about AI “speeding up drug development” often describe a discovery step and stop there.
In practice, each stage asks a different question, and each question suits a different kind of model. A target team asks which gene drives the disease, while a chemistry team asks which of a million molecules might bind to it. A safety team then asks which survivors are likely to be toxic.
Stage Question asked Typical AI method Target identification Which protein or pathway drives the disease? Omics analysis, knowledge graphs, literature mining Hit finding Which compounds might bind the target? Virtual screening, docking with ML scoring Lead optimization How do we improve potency and safety together? Generative models, QSAR, multi-objective optimization Preclinical safety Will it be absorbed, and is it toxic? ADMET prediction models Bridge to the clinic Which patients are most likely to respond? Biomarker discovery, patient stratification
The table also shows why one “AI platform” rarely covers everything. Instead, teams usually combine several methods, each trained on different data and checked in a different way.
The Main AI Methods Five families of methods do most of the work. Knowing which one sits behind a tool helps you judge its claims, because each has different data needs and failure modes. If the terms are new, our explainer on machine learning versus AI covers the basics.
QSAR and classical ML predict activity or properties from molecular descriptors, and they work well with modest, clean datasets.Deep learning and graph neural networks treat molecules as graphs of atoms and bonds, so they learn structure directly.Generative models propose new molecules that meet stated goals, an extension of the same ideas behind generative AI for text.Protein structure prediction estimates a protein’s 3D shape from its sequence, so chemists can design drugs around that shape.NLP and knowledge graphs pull relationships out of papers, patents and databases to suggest targets and repurposing ideas.Buy, Build or Partner Once a team knows the method it needs, the next decision is how to get it. Commercial platforms are quick to start but keep the models opaque. Open-source models are free and inspectable, yet they still need engineering. Custom models fit your data best but take the longest.
Option Best when Watch for Commercial platform You need results fast and lack ML staff Data control, lock-in, limited validation detail Open-source models You have ML engineers and want transparency Integration effort, support, model upkeep Custom or partner-built Your proprietary assay data is the advantage Cost, time to first result, ownership of the code
Whichever route you pick, four criteria decide the outcome. Ask who controls the data and how the tool connects to your ELN and LIMS. Then ask how its predictions are validated and what it costs per validated result.
Kanerika Service
AI and Machine Learning Services
Kanerika builds, validates and runs machine learning models on your own data, with the engineering, monitoring and governance that research teams need to trust the output.
Explore AI and ML Services Why Traditional Drug Discovery Is Slow and Expensive Drug discovery is slow because biology is uncertain and most bets fail late, after the money is spent. AI targets that waste, so it helps to see its size first.
A 2022 review in Acta Pharmaceutica Sinica B puts discovery and development together at 10 to 15 years. It also reports that about 90% of candidates entering clinical trials fail . The main reasons are poor efficacy and unmanageable toxicity, so early models try to predict both.
Estimates of what a new drug costs vary widely by method. The Tufts Center for the Study of Drug Development estimated $2.558 billion per approved drug in 2013 dollars , including the cost of failures. A later study in JAMA, using public data, found a median of $985 million and a mean of about $1.3 billion , in 2018 dollars.
Some disease areas fare far worse. Alzheimer’s candidates tested between 2002 and 2012 had a 99.6% failure rate , a reminder that better prediction cannot rescue a weak biological hypothesis.
Fragmented data also adds a quieter cost. Assay results sit in one system, compound records in another and literature notes in a third, so scientists rebuild context by hand. That gap is also where most AI projects stall, which is why pharma data analytics comes before any model.
The 6 Key Applications of AI in Drug Discovery The six applications below are where AI already does useful work in discovery. Each one narrows a large set of options into a short list that scientists can test, and each still ends at the bench.
Read them as a chain rather than six separate tools. A target chosen in step one shapes the screen in step two, and safety predictions in step four decide which designs from step three survive.
1. Target Identification and Validation Target identification asks which protein, gene or pathway to act on. Models combine genomics, transcriptomics and network analysis with mined literature to rank candidates by their link to the disease.
For example, a knowledge graph can connect a gene that is overexpressed in fibrotic tissue to pathways already tied to scarring. The model then ranks it higher than genes with weaker evidence. Biology still has to validate the target, often with knockdown experiments, before chemistry starts.
2. Virtual Screening and Hit Finding Virtual screening ranks large compound libraries by their predicted chance of binding a target, so the lab tests the top slice first. In effect, it replaces some of the brute-force work of physical high-throughput screening.
Docking tools estimate how a molecule fits a protein pocket, then ML models trained on past assays score those poses. The output is a ranked list, the same 5,000-to-200 cut from the opening scenario at a much larger scale. The hit rate in the lab is the only measure that counts.
3. Generative Molecular Design and Lead Optimization Generative models propose new molecules instead of searching a fixed library. A chemist sets the goals, such as potency, solubility and novelty, and the model suggests structures that try to meet them all.
Still, the model will happily propose molecules that are hard or impossible to make. Chemists therefore filter the output for synthetic feasibility before anything is ordered. The same generative ideas reach beyond chemistry, as our guide to generative AI for healthcare shows.
4. ADMET and Toxicity Prediction ADMET stands for absorption, distribution, metabolism, excretion and toxicity, so it covers how a drug behaves in the body. Models trained on historical data flag compounds likely to fail on these properties, so teams drop them before costly animal studies.
This is one of the most practical uses because the cost of a late safety failure is so high. The same 2022 review that gives the 90% failure figure attributes about 30% of clinical failures to unmanageable toxicity. These models are only as good as the assays behind them, and assay protocols differ between labs.
5. Drug Repurposing Drug repurposing looks for new uses for approved or shelved compounds. Since their safety profiles are already known, a successful match can move faster than a brand-new molecule.
AI helps by matching drug signatures, such as gene expression changes, to disease signatures. It can also mine literature for overlooked links between a compound and a pathway. The best-known case is baricitinib, an arthritis drug. In February 2020, a team at BenevolentAI used its knowledge graph to propose it as a COVID-19 treatment in The Lancet . After phase 3 trials, FDA approved it for certain hospitalized COVID-19 patients in May 2022. The AI made the suggestion, and the clinical trials proved it.
6. Biomarkers and Patient Stratification The last application bridges discovery and development. Models search omics and imaging data for biomarkers that predict who will respond, so trials can enroll the right patients.
Choosing the right patients matters because a drug that helps only a subgroup can look like a failure when everyone is enrolled. If a biomarker picks out likely responders before enrollment, a trial can show a real effect with fewer patients. Much of this work happens once trials begin, which is covered in our guide to AI in clinical data management . Related methods appear in predictive analytics for healthcare .
Application Input data Output How it is validated Target identification Omics, literature, pathway networks Ranked target list Knockdown and functional assays Virtual screening Compound libraries, protein structures, past assays Ranked hit list Lab hit rate on tested compounds Generative design Property goals, known actives New candidate molecules Synthesis and potency assays ADMET prediction Historical ADMET and toxicity data Risk flags per compound In vitro and animal studies Drug repurposing Gene expression signatures, literature New indication matches Preclinical models, then trials Patient stratification Omics, imaging, clinical data Biomarker and responder profiles Prospective clinical trials
How an AI-Driven Discovery Workflow Runs An AI-driven discovery workflow is a loop, not a one-off prediction. Scientists design, make and test compounds, then the results train the next round of models, a cycle chemists call design-make-test-learn.
Watch on YouTube
Business Ontology on Microsoft Fabric: Global Pharma Case Study
See how a global pharma company connected scattered business data into one model on Microsoft Fabric, the kind of groundwork any AI loop needs before it can learn.
The loop only turns as fast as data moves through it. When assay results take weeks to reach the modeling team, each cycle slows to match. Five steps keep it moving.
Connect the data. Pull assay databases, LIMS and ELN records into one governed store, using an AI data pipeline rather than manual exports.Set a baseline. Record how the team picks compounds today and its current hit rate, so you can compare fairly.Train and evaluate. Build models on historical data and test them on held-out compound series, not random splits. Split by chemical scaffold, so no test molecule has a close analog in training. Report the hit rate in the top-ranked set the lab can actually test. Divided by the hit rate across the whole library, it gives the enrichment factor.Test in the lab. Send the top-ranked compounds to the bench under a protocol agreed in advance.Feed results back. Add new assay results to the training data and retrain, tracked through MLOps practices .Step three hides the most common trap. A random train-test split lets near-copies of a molecule sit on both sides. As a result, the model looks better on paper than it will in the lab.
What AI Has Proven So Far The fair way to judge AI in drug research is by evidence rather than promise. Some uses are proven at scale, some show early clinical signals and some remain unshown.
Grading matters because one breakthrough can make a whole field sound proven. Protein structure prediction has a Nobel Prize behind it, while claims about cheaper or faster drugs still rest on vendor reports and forecasts. The three examples below run from strongest to weakest evidence.
Protein Structure: AlphaFold and the 2024 Nobel The clearest proof is in protein structure. The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design and to Demis Hassabis and John Jumper for protein structure prediction.
Their AlphaFold2 model has predicted the structures of virtually all of the roughly 200 million known proteins. For drug hunters, that means a starting structure for targets that were once too hard to study, although a predicted structure still needs experimental checks.
A Generative-AI Drug in a Randomized Trial The strongest clinical signal so far is rentosertib (ISM001-055), a TNIK inhibitor for idiopathic pulmonary fibrosis. Both its target and its molecule were found with generative AI.
In June 2025, Nature Medicine published its phase 2a trial . The double-blind, placebo-controlled study enrolled 71 patients for 12 weeks. Because its primary endpoint was safety, the main finding was similar adverse-event rates across arms.
That is a real milestone, but its scope is narrow. Rentosertib is not an approved drug, and a 12-week safety trial in 71 patients cannot show that it works. The efficacy test has now begun. In September 2026, Insilico Medicine dosed the first patient in GENESIS-IPF-3 , a 52-week phase III trial planned for about 320 patients in China. The paper itself also notes that few AI-discovered drugs have reached human trials.
Where the Evidence Is Still Thin The broader picture, however, is more sobering. A Perspective in Nature Reviews Drug Discovery from August 2026 calls the evidence of clinically relevant impact from AI “so far, disappointingly limited.”
The authors argue that benchmarks should measure whether AI leads to better decisions, rather than how well a model scores. That is also a useful test for any vendor claim you hear.
Proven Promising Not yet shown Protein structure prediction at scale Generative design reaching randomized trials Higher clinical success rates Ranking and filtering large compound sets Faster target-to-candidate timelines Lower total cost per approved drug Mining literature and omics for hypotheses ADMET models that cut late safety failures End-to-end time saved from target to approval
Challenges and Limits of AI in Drug Discovery AI models in drug research fail for ordinary reasons, mostly data and translation problems rather than weak algorithms. Knowing these limits early saves a pilot from looking better than it is.
Data quality comes first. Assay results vary between labs, instruments and protocols, so a model can learn the lab’s habits instead of the chemistry. Negative results are also underreported, which leaves models with a skewed view of what fails. Our guide to AI data quality explains how to measure and fix this.
Checklist
Enterprise AI Checklist: Readiness, Governance and Adoption
Check data readiness, governance and adoption gaps before you fund an AI pilot, so the model learns chemistry rather than inconsistent lab habits.
Get the Checklist → Translation is the second wall. A molecule can look perfect in a model and in a dish, yet still fail in people. That happens because cell assays and animal models are imperfect proxies for human disease. Alzheimer’s shows the cost, since 99.6% of candidates tested from 2002 to 2012 failed.
Train-test leakage inflates accuracy when similar molecules appear in both training and test sets.Chemical domain shift means a model trained on one chemical series may not transfer to a new scaffold.Interpretability limits trust, so chemists need to see why a model ranks a compound highly, the focus of explainable AI .Talent gaps persist, since few people combine medicinal chemistry, biology and ML engineering.Adoption friction appears when scientists see AI as a black box imposed on them, a pattern covered in AI adoption challenges .Even so, none of these limits is a reason to wait. They are reasons to start small, measure honestly and keep the lab in the loop.
How Regulators Treat AI in Drug Development Regulators judge the evidence that AI produces, asking whether it is credible for the specific decision it supports. Both have set out how they make that call, FDA in January 2025 and the two agencies jointly in January 2026.
FDA’s drug center has real experience to draw on. It reports over 500 submissions with AI components from 2016 to 2023 . It also published a discussion paper in May 2023, which drew more than 800 comments.
FDA: Context of Use and Model Credibility In January 2025, FDA issued draft guidance titled Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products . It remains a draft.
The guidance sets out a risk-based credibility assessment. A sponsor defines the model’s context of use, the specific question it answers, and then shows evidence that the model is credible for that use. Higher-risk uses need stronger evidence.
EMA and the Joint FDA-EMA Principles In Europe, the CHMP and CVMP adopted EMA’s reflection paper on AI in the medicinal product lifecycle in September 2024.
Then, in January 2026, FDA and EMA published 10 guiding principles of good AI practice in drug development . In particular, they call for a human-centric, risk-based approach, a clear context of use, multidisciplinary teams and sound data governance .
What This Means for Research Teams Exploratory discovery work rarely reaches a regulator, so these rules bite hardest once AI output supports a submission. Even so, good habits are cheaper to build early than to retrofit. Broader context sits in our guide to AI regulation .
Document each model’s context of use, training data and version before results inform decisions. Keep data lineage so any prediction can be traced to its source records. Keep a named scientist accountable whenever a decision relies on model output. Apply GxP and 21 CFR Part 11 controls to regulated records, not to every exploratory notebook. Set up AI governance before models move from research into regulated work. How to Choose Your First AI Drug Discovery Use Case The best first project is a narrow, measurable decision your team already makes often, not a moonshot. A good pilot proves value in months, then builds trust for the next one.
Ambitious programs tend to stall because they try to fix data, models and culture at once. A single ranking task with clear before-and-after numbers avoids that trap, and it also shows where the data is weak.
Five Questions Before Funding a Pilot Which decision will the model improve? Name it, for example which compounds go to the next assay.Do we have the data? Check volume, consistency and whether negative results are recorded.What is the baseline? Measure the current method’s hit rate or cycle time first.What evidence counts as success? Agree on the lab result that would prove value before you start.Who owns it? Assign one scientist and one engineer who share the outcome.If two or more answers are unclear, fix those gaps first. An AI implementation roadmap helps sequence that work.
AI Assessment
How Ready Is Your Team for an AI Pilot?
Score your data, skills and governance in a short self-assessment and see which gaps to close before you test a first discovery use case.
Start Your AI Assessment → Worked Example: Ranking 5,000 Compounds for 200 Lab Slots Return to the team from the opening scenario. Here is how a fair pilot would run.
To begin with, train a supervised ranking model on the team’s historical assay results for similar targets. Then agree a protocol up front, in which 100 slots go to the model’s top picks and 100 to the chemists’ usual method. After that, run both sets through the same assay, blind to which method chose each compound. Then compare hit rates, the cost per validated hit and the time from list to result. Finally, feed all 200 results back into the model, whichever arm wins. The split design matters because it gives an honest comparison. Write the pass mark down before any compound is tested. The model arm must beat the chemists’ arm on confirmed hits at the same assay cutoff, by a margin agreed in advance. If the model’s 100 picks clear that bar, the team has evidence, not a vendor claim.
One failure pattern is worth planning for. A single noisy assay reading can look like a hit, and a handful of false positives can swing a 100-compound comparison. The fix is a replicate rule, so a compound counts as a confirmed hit only when a repeat assay reproduces the result.
Measuring Value Without Overstating It Track operational metrics during the pilot and keep scientific outcomes separate. Fewer wasted assays is a real gain, but it is not the same as a better drug.
Operational metrics (pilot) Scientific outcomes (long term) Hit rate per tested compound Candidates that pass preclinical safety Cost per validated hit Candidates entering clinical trials Cycle time per design-make-test-learn loop Clinical success rates Share of predictions confirmed in the lab Approved medicines
What’s Ahead for AI in Drug Discovery The next phase of AI in drug discovery will be judged on two things: whether AI-derived drugs succeed in larger trials, and whether models fit into everyday lab work. Four trends will decide whether today’s narrow wins grow into clinical results.
Biological Foundation and Multimodal Models Researchers are training large models on proteins, cells and omics data, in the same way language models learned from text. These foundation models aim to capture biology broadly, then adapt to a specific target or disease.
Combining data types is the next step, since a model that reads sequence, structure and cell images together may catch signals that each misses alone.
Closed-Loop Labs and AI Agents Automated labs that run design-make-test-learn cycles with little human handling are emerging. In these setups a model picks the next experiment, then robots run it and results flow straight back.
AI agents are also starting to help with literature review and assay planning. These remain early tools that need human review, as our look at AI agents for healthcare and agentic AI shows.
Clinical Proof and Regulatory Maturity The decisive question is whether AI-derived drugs succeed in larger, longer trials. Rentosertib’s phase III, which began dosing in September 2026, will be one of the first real answers. Until results like that arrive, claims about higher success rates remain unproven.
Regulation is maturing too. FDA’s January 2025 guidance is still a draft, so its final form and its fit with the joint FDA-EMA principles are worth tracking.
Market Outlook Investment keeps growing even though the clinical questions are still open. Grand View Research sized the AI drug discovery market at $2.3 billion in 2025 and forecasts $2.9 billion for 2026. It projects $13.8 billion by 2033, a compound annual growth rate of 24.8% from 2026. These figures measure what companies spend on AI tools and services, not how many drugs those tools deliver. Read them as a sign of interest rather than proof.
How Kanerika Helps Pharma Teams Build the Data Foundation for AI Kanerika does not design molecules. It builds what discovery AI depends on, which is connected, clean and governed data that models and scientists can trust.
In practice, that starts with data engineering. Our data engineering team connects research and operational sources, such as LIMS, ELN, assay databases and ERP, into a governed data lakehouse . Quality checks catch duplicates, missing values and unit mismatches before they reach a model.
Lineage and access controls then follow every record from source to report, which supports the documentation regulators expect. Our AI governance services and data governance work set those rules, while our AI and ML services handle model engineering on top.
A US pharmaceutical company shows the foundation in action. It manages over 500 SKUs across 55+ countries and 30+ plants. The team unified data from Model N, SQL Server, SAP and SAP Vistex in a Microsoft Fabric lakehouse, with Power BI for reporting. The result was 78% fewer data-related errors and 41% faster data processing.
Case Study
78% Fewer Data Errors for Pharma with Microsoft Fabric
A US pharmaceutical company unified Model N, SQL Server, SAP and SAP Vistex data in a Microsoft Fabric lakehouse, cutting data-related errors by 78% and speeding data processing by 41%.
Read the Case Study → That project was commercial and operational analytics, not drug discovery, so treat it as proof of the data layer only. The same discipline applies to research data. It also underpins our AI for pharma work in commercial and manufacturing settings and our Power BI for pharma reporting.
Bottom Line on AI in Drug Discovery: Key Applications and What’s Ahead AI already earns its place at six discovery steps, from target identification to patient stratification, where it narrows what scientists test. The proof so far is strongest in protein structure and early-stage ranking, while clinical gains remain unshown.
What is ahead is larger trials of AI-derived drugs, biological foundation models, closed-loop labs and final regulatory rules. Each of these depends on the same thing, which is trustworthy research data and honest lab validation.
For the team in the opening, with 5,000 compounds and 200 lab slots, the next step is the same as for any pharma group. Run one measured pilot on clean, governed data, and let the lab results decide how far to go.
Frequently Asked Questions
How is AI used in drug discovery? AI is used to rank and filter options at six research steps. It helps pick disease targets, screen compound libraries, design new molecules, predict toxicity, find new uses for existing drugs and select likely responders. Each prediction narrows what scientists test, and the lab still confirms every result before a compound moves forward.
What are the main benefits of AI in drug discovery? The clearest benefit is better use of scarce lab time. Models rank thousands of compounds so teams test the most promising first. AI also predicts protein structures, flags toxicity risk early and surfaces repurposing ideas from published research. Gains in clinical success rates or total cost per approved drug have not yet been shown in published studies.
Can AI design new drugs on its own? Generative models can propose new molecules that meet goals such as potency and solubility. A proposed molecule still has to be made, tested and proven. Chemists filter proposals for synthetic feasibility, labs test the survivors and clinical trials decide whether they help patients. Scientists remain responsible for every decision along the way.
Can AI reduce the cost of drug discovery? It can cut waste at specific steps. Ranking compounds before testing means fewer assays spent on weak candidates, and early toxicity flags help teams drop risky molecules sooner. A lower total cost per approved drug has not been demonstrated yet, so measure savings per step, such as the cost per validated hit.
How much time can AI save in drug discovery? Savings show up at individual steps rather than across the whole pipeline. Virtual screening can shorten hit finding, and generative design can reduce the number of design cycles a team needs. No published study yet shows a reliable end-to-end time saving from target to approval, so track cycle time for each step you automate.
What are the limitations of AI in drug discovery? Models depend on data that is often inconsistent across labs and missing negative results. Random train-test splits can inflate accuracy, and models may not transfer to new chemical series. Many predictions are hard to interpret. The largest limit is translation, since a compound that looks strong in a model or dish can still fail in people.
How is AI used in drug target identification? AI ranks possible targets by their link to a disease. Models combine genomics, transcriptomics, protein networks and mined literature to find genes or pathways that drive the condition. The highest-ranked targets then go to biologists, who validate them with experiments such as gene knockdown before any medicinal chemistry work starts on the target.
How can pharmaceutical companies start using AI in drug discovery? Start with one narrow, frequent decision, such as which compounds go to the next assay. Check that historical data is consistent, measure the current hit rate, agree in advance on what result proves value and name an owner. Then run a split test against the usual method and feed every result back into the model.
How is AI drug discovery different from traditional drug discovery? Traditional discovery relies on expert judgment and large physical screens to choose what to test. AI-driven discovery adds a prediction step, ranking or generating candidates from past data before the lab spends time on them. Both still depend on lab experiments and clinical trials, so the change is in how candidates are chosen.
Will AI improve clinical trial success rates? That has not been shown yet. About 90% of candidates entering clinical trials fail, and a Nature Reviews Drug Discovery Perspective from August 2026 called evidence of AI’s clinically relevant impact disappointingly limited. Better early filtering may help over time, and larger trials of AI-derived drugs will provide the evidence needed to judge it.
What is AI structure-based drug discovery? Structure-based discovery designs drugs around the 3D shape of a target protein. AI now predicts those shapes from amino acid sequences, and AlphaFold2 has predicted structures for virtually all of the roughly 200 million known proteins. Chemists then use docking and machine learning to find molecules that fit the predicted pocket, followed by lab confirmation.
What data do you need to use AI in drug discovery? You need consistent, well-labeled experimental data, including assay results, compound structures and negative outcomes. Records from LIMS, ELN and assay databases should be connected, with clear units, protocols and lineage. Omics and literature data help target work. Without clean, governed data, models tend to learn lab quirks and miss the underlying biology.
Has a drug discovered with AI reached clinical trials? Yes. Rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis whose target and molecule came from generative AI, completed a phase 2a trial published in Nature Medicine in June 2025. That trial enrolled 71 patients for 12 weeks and focused on safety. A 52-week phase III trial in China dosed its first patient in September 2026. It is not yet approved.
Which companies use AI for drug discovery? Three groups use it. Large pharmaceutical companies run internal AI teams, AI-native biotechs build their pipelines around it, and technology firms supply platforms and models. Insilico Medicine is one example, having taken the generative-AI-discovered drug rentosertib into a phase 2a trial. Many academic research groups also use open-source models for screening and property prediction.
How is the FDA approaching AI in drug development? FDA reviews whether evidence produced with AI is credible for the decision it supports. Its January 2025 draft guidance asks sponsors to define a model’s context of use and show proportionate evidence of credibility. In January 2026, FDA and EMA also published ten shared principles of good AI practice in drug development, covering governance and oversight.
What does the future of AI in drug discovery look like? The next phase centers on proof. Larger trials of AI-derived drugs will show whether early signals hold. Biological foundation models trained on proteins, cells and omics data are growing, and automated labs are testing closed design and test loops. Regulators are still developing their rules, and FDA’s January 2025 guidance remains a draft. Documented, well-governed research data will matter more.
What is the role of AI in drug repurposing? AI matches existing drugs to new diseases. It compares how a drug changes gene expression with the molecular signature of a disease, and it mines published research for overlooked links. Because approved or shelved compounds already have known safety profiles, a strong match can reach testing faster than a new molecule, though trials still decide.
How does machine learning help in drug discovery? Machine learning learns patterns from past experiments and applies them to new compounds. Models predict binding, solubility and toxicity from molecular structure, so chemists can prioritize what to make and test. Each new round of lab results then retrains the model, which helps predictions improve over time when the underlying data is consistent.
What AI methods are used in drug discovery? Five method families do most of the work. QSAR and classical machine learning predict properties from molecular descriptors. Graph neural networks learn directly from molecular structure. Generative models propose new molecules. Protein structure prediction estimates 3D shapes from sequences. Natural language processing and knowledge graphs extract relationships from papers, patents and biological databases.
Is AI new to drug discovery? No. Computational chemistry and QSAR models have supported drug research for decades. What changed is scale and capability. Deep learning, generative models and accurate protein structure prediction now handle problems that older methods could not, and AlphaFold’s creators shared the 2024 Nobel Prize in Chemistry. The idea of predicting before testing is long established.
