AI in drug development is the use of machine-based systems to support decisions at every stage of a drug’s life cycle, from discovery through clinical research and postmarketing to manufacturing1. The catch is that the decision changes from stage to stage. A ranking that tells a lab which experiment to run next needs different evidence than a call that decides what happens to a patient or a production batch. That is why AI in drug development is not one platform but a set of tasks, each with its own data, risk and owner. FDA reports submissions with AI components in all of these phases1, and reviews cover discovery, clinical trials and bioprocessing23. Neither shows that a given model works. This guide shows where AI appears in each phase and what evidence to ask for before you rely on it.
I’m Julia Pizzolato, PhD, Commercial Scientific Liaison at evitria in Zurich. My background spans protein engineering and antibody design as well as data management and analytics, including at CSL Behring, and biomanufacturing analytics. That gives me two views of AI in drug development: the data on a screen and the protein that has to come out of the lab. At evitria, I support customers and project teams in the life sciences through technical consultation, feasibility assessments and training.
In this article, I explain where AI can support decisions across the drug development lifecycle and why the intended use of a model sets the evidence it needs. I also show how teams can assign ownership for data, validation and the final decision. I use antibody discovery as my worked example, because antibody engineering and recombinant expression are where my work at evitria sits, and the other phases get context. The questions and checklists here are editorial practice, not regulatory advice or an FDA requirement.
Throughout this article, “intended use” is what a model’s output is used for: the decision it informs, who acts on it and what happens if it is wrong. “In silico” means computational.
What do artificial intelligence, machine learning and generative AI mean in drug development?
The terms get used loosely across the pharmaceutical industry, so here is how this guide uses them. FDA defines artificial intelligence (AI) as a machine-based system that makes predictions, recommendations or decisions for human-defined objectives. It builds a model from data and uses that model to formulate options for information or action1. Machine learning (ML) is the part of AI that learns from data to get better at a task, and FDA describes it as commonly used across the drug product life cycle1. Generative AI is a mode that proposes something new, such as a protein design, instead of scoring candidates you supply4.
For your evidence plan, the useful distinction is the output, not the model family. A predictive model ranks the candidates you give it or predicts a property for each one. A generative model hands you candidates you did not supply. Either way, you get a suggestion, and someone still has to judge what it means for the program. That is usually a domain scientist working with the data science team that built the model. Automation is a different thing: a fixed script or a liquid-handling robot runs set steps and formulates no options from a model, so it is not AI by this definition.
Why does AI in drug development matter, and what should you expect?
The clearest sign is volume. FDA’s drug center reports a significant increase in submissions with AI components over the past few years, and says the scope and impact of AI use in drug development are expanding1. Its 2025 draft guidance draws on experience with more than 500 such submissions from 2016 to 20231.
What changes is where teams can get decision support. Reviews describe uses from drug discovery to clinical-trial design2 and from strain selection to process control3. One CHO study below shows the upside in a single process: a neural network suggested cultivation settings that raised antibody titer by about 48%5.
Expect narrower results than the headlines. The Med review calls AI’s capacity for innovation transformative2, but a review maps opportunities and does not prove that each use works. The bioprocess review is titled “from promise to practice” for a reason: it documents the gap between the two3. The Med review makes a similar point about approvals: it notes that no AI-conceived drug had yet received approval and proposes large language models (LLMs) and diffusion models as one way to change that2. A realistic expectation is better-informed choices at specific decision points, each tested against evidence that fits its use.
Where does AI support the drug-development lifecycle?
FDA reports drug application submissions with AI components across the nonclinical, clinical, postmarketing and manufacturing phases of the drug product life cycle1. The table sets each phase beside what AI may support, the evidence question to ask and the function that stays accountable. Presence in a phase shows where AI appears, not that a use works.
| Drug-development phase | What AI may support | Evidence question | Accountable function | What the sources show |
|---|---|---|---|---|
| Drug discovery and nonclinical research | Research-stage prediction or candidate ranking for a stated task | Does it beat a task-relevant comparator on relevant data? | Program owner with computational and domain scientists | A review covers AI across drug discovery2; FDA names the nonclinical phase1 |
| Clinical development and clinical trials | Review-level discussion of AI for clinical-trial design | What validation fits the clinical question, the user and the consequence? | Clinical program owner, with quality or regulatory input where applicable | A review discusses trial design as a research area2; FDA names the clinical phase1 |
| Postmarketing | High-level analytics and decision support after launch, including real-world data analytics | Who owns data provenance and monitoring, and do the data represent the intended population? | Postmarketing owner with data and quality functions | FDA names the phase and real-world data analytics as an area of active engagement1 |
| Manufacturing | Process-level ML for optimization, scale-up, monitoring and control | Do process data, comparator and monitoring fit the host, scale and decision? | Manufacturing and quality owners | A review maps these uses3; one CHO study tests one cultivation process5 |
Drug discovery and nonclinical research
AI in drug discovery usually hands the team a ranked list or a prediction: which drug candidates to test first. In silico methods narrow the field, and the experiment that fits the decision then supplies the next piece of evidence. A 2024 review in Med covers AI across drug discovery, development and clinical trials, with a focus on small molecules, RNA and antibodies2. A ranking can prioritize a candidate, and the test that follows decides whether it advances.
Clinical development and clinical trials
The Med review discusses AI for clinical-trial design as an area of research2. Once a model’s answer feeds into a decision about patients, ask what the answer is used for, who relies on it and what happens if it is wrong. The more a wrong answer costs, the more proof you need.
Postmarketing and real-world data
After approval, FDA points to real-world data (RWD) analytics as an area of active engagement where AI use is growing1. RWD are data on patient health status or the delivery of health care, routinely collected from sources such as electronic health records, medical claims data and registries6. An analytics use depends on whether the data represent the intended population and on who owns provenance and monitoring.
Manufacturing
A review in Trends in Biotechnology follows machine learning through strain engineering and selection, process optimization, scale-up, monitoring and control3. The clearest numbers come from one CHO study. In 2025, Richter et al. trained a neural network on historical and newly generated runs of a CHO cell cultivation process for a therapeutic monoclonal antibody5. The network then suggested better settings5. In validation runs, the best setting raised the final antibody titer from 3.15 to 4.65 g/L, about 48% above standard conditions5. That was within 0.06 g/L of what the model had predicted5. Peak cell density, another target, did not improve5. The authors also name the method’s limits: it needs large, diverse data, it cannot extrapolate beyond the data ranges it has seen, and its exact actions are hard for people to follow, the so-called black box5. The model works on cultivation settings, not on the antibody sequence. The authors validated it in small ambr 15 bioreactors and name scale-up as a next step, so the 48% belongs to that one cell line and process and not to antibody yield in general5.

Why should intended use determine the evidence plan?
Ask one model two different questions. First: which designs should the lab express next? Second: is this production lot within specification? The model is the same, but the evidence you need before acting on its answer is not. Intended use is the name for that difference, and it is how FDA approaches AI as well. Its drug center plans to keep developing a risk-based regulatory framework1, and the 2025 draft guidance covers AI used to produce information that supports regulatory decisions on the safety, effectiveness or quality of drugs17. FDA also expects newer tools such as LLMs to widen its own use of AI1.
Start with the forgiving case. A research suggestion only tells you which experiment to run next, and the experiment then decides whether the work moves forward. A model in that role has to be good in aggregate. Its job is ranking, so the question is enrichment: does the top of the list hold more good candidates than chance would?
An illustration, not a result: suppose a model ranks 500 designs, and its top picks contain 40 soluble expressors where the same number of random picks would contain 8. That model is worth using. It still tells you nothing trustworthy about any single design on the list. Only the wet lab shows which designs are good, and how good.
A clinical or process decision is the opposite case. It is the terminal decision, so no further experiment is there to correct it, and the model needs a reliable claim about this one case: is this lot in specification, does this patient receive this drug? That takes calibration on the actual decision, defined failure modes and a monitoring plan, because the risk is much higher. A model that is good enough for the first job can be far from good enough for the second.
To turn this into an evidence plan, ask five questions, in this order. What decision will the output inform? Who uses the output, and who stays accountable for the decision? Which data represent the intended context, and what gaps do you already know of? Which comparison and failure criteria apply? And which change in model, data, workflow or intended use triggers review or revalidation?
How do drug development teams operationalize AI?
An AI use works when it has a named owner and a defined path from model output to decision. The program owner owns the decision and its acceptance criteria. The data-science team implements the model, versions it and runs the technical tests. The domain team tests whether the output holds up biologically, clinically or in the process. Where it applies, a quality or regulatory function assesses the requirements for the stated use, and an external partner supplies the defined material, data or service and returns the agreed records. Your organization may split these roles differently, but each one needs a name against it.
The order of work matters as much as the roles. The program owner fixes the intended use and the acceptance criteria before testing starts. The data-science team picks the comparator before the model sees test data. The domain team validates the output for the stated use and sets when a human reviewer steps in. A named owner then monitors performance and decides which changes to model, data or workflow trigger a review.
A score on its own says little. Two questions help before anyone quotes one.
What does the simplest baseline score? Compare the model with the simplest way you can think of to select or rank candidates, even random picking, and ask how much better the model did. A performance score is hard to interpret until you know how a “stupid” model would have scored.
How closely related are the new designs to the training data? Ask how the team split train from test. The answer shows how well the model generalizes beyond what it has already seen.
What should cross-functional handoffs contain?
A handoff works when it fits on one page. The handoff card names the decision the output informs, the versioned input and output, the limits of use, the acceptance criteria, the responsible owner and the next test. Records that an external partner returns, QC data included, become part of the provenance, so the card lists them too. The same card serves a handoff to discovery, clinical development or manufacturing. The acceptance criteria change, and the six entries do not.
How does antibody discovery illustrate the evidence boundary?
Antibody discovery shows the boundary in one drug-discovery setting. In silico methods rank or propose antibody sequences, and a 2025 review describes ML models trained on sequence, structure and property data that aim beyond affinity at specificity, stability, viscosity and manufacturability4. Deep learning plays a part here: the review calls protein- and antibody-specific language models valuable tools for antibody design and optimization4. The benchmark authors point to the CASP assessments, which led to the Nobel Prize-winning advances of AlphaFold in protein structure prediction, as the template for this kind of prospective, blinded evaluation8. A blinded, prospective benchmark found that AI can optimize antibodies in defined, biologically grounded regimes, and it highlighted gaps in cross-task generalization8. Its authors read the results as an upper bound for the target class they tested, not as evidence of performance across arbitrary antigens8.
That is why every sequence you want to decide on needs expression and QC. Those data supply the evidence for what happens to each design next: advance it, redesign it or drop it. For the full path from model output to measured protein, see AI Antibody Discovery.
What should a team decide before an AI project moves forward?
Nine questions are enough to start, and an item without an answer is the next piece of work.
- Intended use: which decision does the output inform, and what does a wrong output cost?
- Owner: who is accountable for that decision, and who runs the model?
- Data: where do the training and test data come from, and do they represent the context of use, with known gaps recorded?
- Comparator: which simple baseline, even random picking, will you test the model against, and did you choose it before testing?
- Validation: which test shows that the output holds up for this use, and when does a human reviewer step in?
- Failure criteria: which result counts as a failure, and what happens then?
- Downstream action: which result advances the work, and which one sends it back or stops it?
- Documentation: are the model version, the limits of use, the handoff and any partner records, QC data included, written down?
- Change trigger: which change to model, data, workflow or intended use triggers a review or revalidation?
From ranked list to purified antibody: how evitria helps
A ranked list starts to pay off the day its top candidates exist as protein. Getting you to that day is the part we take on at evitria, and the one we enjoy most. It is where a row in a table becomes something you can hold up to an assay.
You send us the antibody sequences you have chosen, whether a model ranked them or not. We express them by CHO-based transient expression, purify them and measure what we made. Standard QC covers purity, quantification, concentration and endotoxin level. Depending on your project, we add titer measurement, HPLC-SEC for purity and aggregation, and CE-SDS for structural integrity. Your antibodies come back with that QC data, so your first experiment starts from measured quality and not from a score. That matters downstream: screens often measure how a candidate interacts with a purified drug target, and the readout depends on the fidelity of the physical protein.
When a model hands you a long list, our high-throughput antibody production service expresses up to several hundred constructs per project in parallel. Every construct runs through the same standardized workflow, so you can trace differences between candidates to the sequence and not to the process. What you measure can then feed your next model round. Every project is different, so tell us how many candidates you have and what you need to learn next, and we will think it through with you.
In the lifecycle above, our work sits at the discovery and preclinical end. Our services cover the research and preclinical stages of drug discovery, and our products are not for human use.
Key takeaways
- AI in drug development covers decision-support uses across discovery, clinical development, postmarketing and manufacturing, not one system.
- A phase, a review or a submission with AI components shows where AI appears, not that a model works.
- A research suggestion has to rank well in aggregate. A terminal clinical or process decision needs a reliable claim about one case.
- Before you trust a score, ask what the simplest baseline scores, even random picking, and how the team split train from test.
- Every handoff carries the decision, the versioned input and output, the limits of use, the acceptance criteria, the owner and the next test.
- evitria expresses and purifies client-supplied antibody sequences, including model-ranked ones, and delivers them with QC data for research and preclinical work.
FAQs about AI in drug development
Reviews describe applications in drug discovery, clinical trials and bioprocess development23. In bioprocess development these include strain selection, optimization, scale-up, monitoring and control3. At FDA, submissions with AI components span the nonclinical, clinical, postmarketing and manufacturing phases1. Which use is appropriate depends on its intended use.
No. It means several decision-support uses across research and development (R&D) and the drug product life cycle, not one platform. Examples range from drug discovery and clinical development2 to process monitoring and control in manufacturing3. Each use has its own data, its own failure consequence and its own evidence need.
Artificial intelligence is a machine-based system that makes predictions, recommendations or decisions for human-defined objectives1. ML is the subset of AI that learns from data to improve at a task1. Generative AI is one mode that may propose new candidate content or designs instead of scoring ones you supply4.
Start from the decision the output informs and the person who stays accountable for it. Then check that the data represent the intended context, that the team chose a simple baseline before testing and that validation matches the use, and agree which change triggers a review. The nine questions above turn these points into a written gate.
Antibody discovery is one drug-discovery example: AI can rank or propose sequences, and expression and QC data decide what happens to each design next. Prospective evidence shows that performance varies by task and data regime8. The antibody section above points to the full workflow.
- U.S. Food and Drug Administration. Artificial Intelligence for Drug Development. https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/artificial-intelligence-drug-development (2026). ↩︎
- Zhang Y, Mastouri M, Zhang Y. Accelerating drug discovery, development, and clinical trials by artificial intelligence. Med 5(9), (2024) 1050-1070. https://doi.org/10.1016/j.medj.2024.07.026 ↩︎
- Helleckes LM, Hemmerich J, Wiechert W, et al. Machine learning in bioprocess development: from promise to practice. Trends Biotechnol 41(6), (2023) 817-835. https://doi.org/10.1016/j.tibtech.2022.10.010 ↩︎
- Matsunaga R, Tsumoto K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J Biomed Sci 32, (2025) 46. https://doi.org/10.1186/s12929-025-01141-x ↩︎
- Richter J, Wang Q, Lange F, et al. Machine Learning-Powered Optimization of a CHO Cell Cultivation Process. Biotechnol Bioeng 122(5), (2025) 1153-1164. https://doi.org/10.1002/bit.28943 ↩︎
- U.S. Food and Drug Administration. Real-World Evidence. https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence (2026). ↩︎
- U.S. Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft guidance). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological (2025). ↩︎
- Erasmus MF, Bedinger D, Hopkins E, et al. A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03238-6 ↩︎

