How AI Is Used in Drug Discovery

By Kevin Fernando · Sep 8, 2026 · 10 min read

AI speeds drug discovery by automating three core tasks: identifying disease-causing proteins, screening billions of molecular candidates computationally, and predicting whether compounds will work in the body before synthesis. This compresses timelines from years to months and reduces preclinical costs, though human chemists and clinical trials remain essential.

AI speeds drug discovery by automating three core tasks: identifying disease-causing proteins, screening billions of molecular candidates computationally, and predicting whether compounds will work in the body before synthesis. This compresses timelines from years to months and reduces preclinical costs, though human chemists and clinical trials remain essential.

Drug discovery is slow and expensive. A new medicine takes 10-15 years to move from laboratory discovery to regulatory approval, with preclinical R&D alone costing approximately 1.098 billion USD. Most compounds fail. Only 12 percent of drugs that enter Phase I clinical trials ever reach patients.

AI does not eliminate these constraints. It compresses the early stages, the parts before human testing begins, by automating the work that chemists and biologists do by hand: screening millions of molecular candidates, predicting whether a compound will bind to its target, identifying which proteins cause disease. This saves years and billions in preclinical spending, but it does not replace the wet lab, the clinic, or human judgment. Understanding where AI adds measurable value and where it stops is essential for evaluating biotech companies claiming AI-driven efficiency.

How AI is used in drug discovery: the core techniques

AI accelerates drug discovery by automating three sequential tasks: finding the right target (the disease-causing protein), generating and screening candidate molecules, and predicting whether those candidates will work in the body. Each task traditionally required months or years of human effort. AI does not eliminate the work; it compresses it by replacing manual screening with computational prediction.

The core techniques, molecular docking, QSAR modeling, and generative neural networks, are not new. What has changed is scale and speed. A chemist can test thousands of compounds by hand. A computer can screen billions. The bottleneck has shifted from "Can we test this?" to "Which candidates are worth testing?"

Target identification and validation

Drug discovery begins with a target: a specific protein whose malfunction drives disease. Identifying the right target is critical. A target that does not actually cause disease will yield a drug that does not work, no matter how well optimized.

AI accelerates target identification by analyzing large biological datasets, gene sequences, protein interactions, disease tissue samples, to propose which proteins are likely involved in a disease. Machine learning models can identify patterns in genomic data that suggest a protein's role in disease progression. Once a target is proposed, AI can also help validate it by predicting what would happen if that protein were blocked or modified.

This remains an area where AI supports human biologists rather than replacing them. The computational prediction must be tested in cells and animals before a company commits resources to drug development.

Molecular screening and compound generation

Once a target is chosen, the search begins: which molecules will bind to it and alter its function? The chemical space of synthetically accessible, drug-like small molecules is estimated to exceed 10^60, far too large for any laboratory to test physically. Typical high-throughput screening campaigns screen only hundreds of thousands of compounds, a tiny fraction of what exists.

AI expands the search by using virtual screening: computational methods that predict whether a molecule will bind to the target without synthesizing it first. The Enamine REAL Space library comprises drug-like compounds available for virtual screening. A neural network can score each one against the target in seconds.

Molecular docking is one technique. It computationally optimizes the placement of a ligand (the drug candidate) in a macromolecular binding site of known or proposed 3D structure and provides a score as to the quality of the fit. Docking can discover novel ligands for a target by screening large virtual compound libraries and provides a useful starting point for structure-based ligand optimization. Docking methods were pioneered during the early 1980s and remain a highly active area of research.

QSAR (Quantitative Structure-Activity Relationship) models work differently. They are mathematical relationships linking chemical structure and pharmacological activity in a quantitative manner for a series of compounds. QSAR models first summarize a supposed relationship between chemical structures and biological activity in a dataset of chemicals, then predict the activities of new chemicals. QSAR has served as a valuable predictive tool in the design of pharmaceuticals and agrochemicals for decades. 3D-QSAR has emerged as a natural extension, exploiting three-dimensional properties of ligands to predict biological activities.

Generative AI takes this further. Instead of screening existing molecules, generative models can propose entirely new structures that meet specified criteria: bind to the target, resemble known drugs, avoid known toxins. This expands the effective search space and allows chemists to focus synthesis effort on molecules with higher predicted success rates.

Predicting drug properties and toxicity

A molecule that binds to its target in silico may fail in the body. It might be too large to cross cell membranes, too unstable to survive stomach acid, too toxic to liver cells, or too similar to a molecule that causes side effects.

AI predicts these properties before synthesis. Machine learning models trained on historical drug data can forecast whether a compound will be absorbed, how it will be metabolized, whether it will be toxic at relevant concentrations, and how it will interact with off-target proteins. These predictions are not perfect, but they are good enough to eliminate the worst candidates before a chemist spends weeks synthesizing them.

AtomNet, a convolutional neural network, successfully found novel hits across 318 individual projects spanning every major therapeutic area and protein class in virtual high-throughput screening. This demonstrates that AI-driven screening can work across diverse biological problems, not just a narrow class of targets.

Where AI saves time and money in the pipeline

The economic impact of AI in drug discovery is measurable. Across 38 drug development programs deploying generative AI models with cloud-orchestrated data pipelines, median lead-to-candidate timelines were reduced by 41.3% compared to conventional structure-based methods. Computational costs per candidate were reduced by 62.7% for programs using generative AI plus cloud infrastructure versus conventional structure-based methods alone.

Timeline compression is most dramatic in the earliest stages. Lead identification (hit-to-lead) takes 1-2 years with capitalized R&D costs of 166 million USD. Lead optimization takes approximately 2 years with capitalized R&D costs of 414 million USD. If AI reduces these phases by 40 percent, a company saves roughly 8-10 months and tens of millions in preclinical spending per candidate.

Insilico Medicine reported a typical timeline to developmental candidate using generative AI of 13 months, compared to the traditional benchmark of 4.5 years. The company's record is 9 months. Recursion's REC-7735 (PI3Kα H1047R inhibitor) development cycle took 10 months from target identification to candidate nomination, requiring 242 synthesized compounds, compared to an industry benchmark of four years and 2,500 compounds.

These timelines matter because they compress the period before clinical entry, when a company is burning cash without revenue. Faster preclinical work means faster entry into Phase I trials, where the real validation begins. The total preclinical R&D costs prior to clinical entry are approximately 1.098 billion USD. Reducing that phase by 40 percent saves hundreds of millions per program.

What AI cannot yet do in drug discovery

AI's limitations are as important as its capabilities. A computationally promising molecule may be impossible to synthesize, prohibitively expensive to manufacture, or toxic in humans. Clinical trials cannot be replaced by prediction. These gaps define where human chemists, biologists, and wet-lab work remain essential.

Synthesis and manufacturability

A molecule that looks perfect on a computer may be a nightmare to synthesize. Some structures require reaction sequences that are inefficient, dangerous, or require rare reagents. Others are unstable and decompose during synthesis. Still others are technically possible but require so many steps that the cost per gram is prohibitive for a drug that must be manufactured at scale.

AI can predict some of these problems, neural networks trained on synthetic chemistry data can estimate whether a proposed structure is synthetically accessible, but the prediction is not reliable enough to skip human chemist review. A chemist must evaluate the proposed synthetic route, identify bottlenecks, and redesign the molecule if necessary. This is why AI-driven programs still synthesize hundreds of compounds. The computational screening eliminates the worst candidates, but human judgment filters the rest.

Manufacturability is a separate problem. A compound that works in a small-scale laboratory synthesis may be unstable or impractical when scaled to kilogram or ton quantities. Process chemists must develop scalable synthetic routes, optimize yields, and ensure reproducibility. This work happens after a molecule enters clinical development and cannot be predicted from structure alone.

Clinical translation and human biology

AI cannot predict human toxicity or efficacy with certainty. Animal models and human trials cannot be replaced by in silico prediction. One AI-discovered drug candidate reached Phase IIa with encouraging efficacy; another stalled in Phase I when safety signals surfaced. Both passed computational screening and animal testing. Both entered human trials. Only one survived.

This is the hard boundary of AI in drug discovery. Computational models are trained on historical data, chemical structures, binding affinities, animal study results. They are excellent at interpolating within that training set. They fail when human biology behaves differently than the model predicts. A compound might be toxic to a liver enzyme that was not represented in the training data. It might trigger an immune response that no animal model predicted. It might work in a cell culture dish but fail in a living organism.

AI can support clinical trial design, conduct, and analysis including streamlining information gathering, optimizing eligibility criteria, and predicting trial success. AI-driven methods can support patient recruitment and retention through matching algorithms, digital health technologies, and personalized interactions. During trial conduct, natural language processing and sensor-based monitoring offer opportunities to reduce administrative burden and capture real-world, patient-centered outcomes. AI-based simulations such as digital twins and in silico trials hold potential to complement conventional trial designs. But none of these replace the trial itself. The trial is where computational prediction meets human biology, and human biology often wins.

How to evaluate AI's real impact on a biotech company

Investors and analysts evaluating biotech companies often encounter AI claims that sound transformative but are difficult to verify. Real AI-driven efficiency shows specific signals.

First, look at pipeline progression. Has the company advanced candidates faster than historical benchmarks? Insilico Medicine's AI-designed drug Rentosertib (TNIK inhibitor for idiopathic pulmonary fibrosis) reached Phase II clinical development. Recursion's REC-4881 (MEK1/2 inhibitor) is in Phase 2 development for Familial Adenomatous Polyposis with Fast Track and US and EU Orphan Drug designations. These are not hypothetical AI candidates; they are in human trials. This is verifiable progress.

Second, examine the data infrastructure. Recursion has over 50 petabytes of proprietary multimodal biological and molecular data created to train foundation models. The company generated hundreds of billions of microglia for biological mapping. This is the foundation for AI models. Companies without this data infrastructure cannot train models that generalize across diverse targets.

Third, look for measurable outcomes from AI-driven programs. Recursion's REC-7735 showed 43% median reduction in polyp burden after three months of treatment in the Phase II TUPELO study. Recursion achieved AI trial enrollment improvement of 1.3 to 1.6-fold compared to historical benchmarks. These are concrete results, not promises.

Fourth, assess partnership validation. Recursion received over 500 million USD in cumulative realized inflows from all biopharma partnerships. The company received 216 million USD in upfront and milestone payments from Roche-Genentech collaboration and 134 million USD in upfront and milestone payments from Sanofi collaboration. Major pharmaceutical companies do not commit this capital to AI-driven discovery unless they see results. Partnership revenue is a signal that the approach works.

Beware of claims about timeline compression without pipeline evidence. "AI reduces discovery timelines by 50 percent" is marketing. "We advanced three candidates to Phase II in 18 months, compared to historical timelines of 4-5 years" is a claim you can evaluate.

How AI in drug discovery connects to company valuation

AI's impact on valuation flows through three mechanisms: pipeline value, burn rate, and time to revenue.

Pipeline value depends on the probability that a candidate will reach patients and the revenue it will generate. A candidate in Phase II is worth more than a candidate in Phase I because it has survived more hurdles. If AI advances candidates through preclinical stages faster, it increases the number of candidates in clinical development relative to a company's age and spending. More candidates in the clinic means higher expected pipeline value.

Burn rate is the cash a company spends per month before generating revenue. Preclinical research is expensive: average cost of preclinical research is 5 million to 40+ million USD per program. If AI compresses preclinical timelines by 40 percent, a company spends less cash before entering clinical trials. Lower preclinical spending means lower burn rate, which extends cash runway and reduces the need for additional financing.

Time to revenue is the years between founding and the first product approval and sales. This is the most important metric for early-stage biotech. A company that reaches Phase II candidates in 18 months instead of 4 years reaches clinical validation faster. It reaches potential revenue faster. This reduces the risk that the company will run out of cash before proving its science works. Faster clinical entry also means earlier market entry, which matters for competitive positioning.

These effects compound. A company with AI-driven efficiency has more candidates in the clinic, lower burn rate, and shorter time to revenue. All three improve valuation. But valuation also depends on whether the candidates actually work, whether the clinical trials succeed. AI accelerates the path to the clinic. It does not guarantee the clinic will say yes.

Do this automatically

Screening a single molecular candidate manually takes weeks of laboratory work across multiple chemists, and a typical drug program requires testing thousands of candidates before identifying promising leads, a timeline that stretches discovery timelines across years and consumes substantial resources before any compound reaches synthesis. These three tasks, protein identification, molecular screening, and efficacy prediction, can now be automated computationally, compressing the preclinical phase and reducing costs while preserving the human expertise and clinical validation that remain essential to bringing safe drugs to patients. Healthcare research reads ClinicalTrials.gov records, regulatory filings, and published literature directly, and cites each claim to the record it came from, across as many subjects as you need, with Mimir attaching a citation to every figure. It will not predict a trial's outcome or advise on anyone's treatment. Those decisions stay with clinicians and regulators, and deliberately so. Every plan includes it, including the free one.

Start your first report and run it on something you already know well.