Is AI Accelerating Drug Discovery? Data From 2 Molecules

Editorial Note: AI is making parts of drug discovery faster—but faster discovery is not the same as faster proof. We look at two recent examples to separate what AI has already changed from what still depends on experiments, clinical evidence, and human judgment.

AI Is Rewriting Drug Discovery: Two Landmark Molecules—and What They Mean for Pharma Jobs

Part 1 — Two Molecules That Changed the AI Drug Discovery Debate

The argument over whether AI can actually build new drugs comes down to two molecules.

The first is Rentosertib. A generative AI platform identified its target, and AI models designed the compound itself. It is now in Phase III trials for idiopathic pulmonary fibrosis (IPF). The journey from target identification to preclinical candidate took roughly 18 months. The conventional path usually takes four to six years.

The second is RFantibody. David Baker's lab—the 2024 Nobel laureate in Chemistry—designed it from scratch: feed the model a target epitope, and it outputs an antibody that binds with atomic-level precision. The work was published in Nature in November 2025, and the code sits on GitHub, open for anyone to download and reproduce.

These two molecules frame the two questions this article answers:

Is AI actually accelerating drug development?
What does that acceleration mean for the people who work in pharma?

Both stories will get told here in full—the Nature Medicine headlines and the FDA clinical hold alike. The data comes first; the conclusions follow.

Part 2 — Rentosertib: An AI-Discovered, AI-Designed Drug Reaches Phase III

Start with the small molecule, because it is closest to market.

The target emerged from an AI-driven target-discovery process. IPF is a progressive lung disease with substantial unmet medical need. Insilico Medicine used its PandaOmics platform to analyze multi-omics and other biomedical data and prioritized TNIK as a potential therapeutic target. The company then used its generative chemistry platform, Chemistry42, to design a small-molecule inhibitor against TNIK (Source: Insilico Medicine; Nature Biotechnology).

Speed is the first hard piece of evidence. Target confirmation to preclinical candidate: about 18 months, against an industry norm of four to six years for small-molecule discovery. At that magnitude, the front-end efficiency gain no longer needs to be argued.

The efficacy data passed peer review. In June 2025, Phase IIa results appeared in Nature Medicine: 71 IPF patients, randomized, double-blind, placebo-controlled. The 60 mg group showed a mean improvement in forced vital capacity (FVC) of +98.4 mL; the placebo group declined by −20.3 mL (Source: Nature Medicine, 2025). FVC is a standard measure of lung function and an important efficacy measure in IPF. The placebo group showed a decline over the 12-week study period, while the 60 mg group showed a mean improvement of 98.4 mL. The trial was primarily designed around safety, however, so the FVC finding should be treated as an encouraging signal rather than definitive evidence of efficacy.

The dark side belongs in the record too. In May 2025, the FDA placed the U.S. Phase IIa trial of Rentosertib on clinical hold; the hold was lifted in October 2025 (Source: Insilico Medicine filings). The episode illustrates a central limit of AI drug discovery: computational models can accelerate molecular discovery, but human safety still has to be established through clinical development.

Big pharma has already placed a broader bet on Insilico's AI drug-discovery platform. In March 2026, Insilico struck a global licensing and research collaboration with Eli Lilly worth up to $2.75 billion, including a $115 million upfront payment. The deal covers certain preclinical oral therapeutics and additional joint R&D programs—not Rentosertib itself. Rentosertib entered Phase III development in July 2026 (Source: Insilico Medicine; Lilly disclosures).

So the verdict for this section can stay narrow. The strongest evidence for front-end acceleration is the roughly 18-month path from target identification to preclinical candidate nomination reported by Insilico—not the fact that the program later reached Phase III. Its progression into Phase III shows that the resulting candidate survived substantial downstream development, but it does not by itself prove that AI caused the acceleration. Whether the drug works at scale waits on Phase III data. Until then, "AI has conquered drug discovery" is premature—and "AI pharma is all hype" is contradicted by a rapid discovery cycle and a positive Phase IIa signal.

Small molecules had one advantage going in: they can be written as SMILES strings, which makes them, in effect, generatable text—squarely inside the capability zone of generative AI. The rules of the biologics world are different.

Part 3 — De Novo Antibody Design: How RFantibody Changes the Discovery Step

Antibodies are large, complex three-dimensional proteins, and for years they resisted generative modeling. The standard discovery path runs: immunize an animal (usually a mouse) → screen hybridoma or phage display libraries → humanize the hits → validate function. The process takes years, leans heavily on animal work, and depends partly on the luck of the screen.

In November 2025, Nature published "Atomically accurate de novo design of antibodies with RFdiffusion" from David Baker's lab at the University of Washington's Institute for Protein Design (Source: Nature, 2025). Baker won the 2024 Nobel Prize in Chemistry for computational protein design. RFantibody distills to one sentence: give the model a target epitope on an antigen, and RFdiffusion generates an antibody structure that binds it with atomic-level accuracy. De novo design can reduce reliance on animal immunization and shift the discovery process from conventional library-based screening toward computational design followed by experimental screening.

One detail deserves its own emphasis: the code is fully open-source, hosted on GitHub. Any lab with structural biology expertise and GPU resources can reproduce the work. That stands in sharp contrast to pharma's deeply rooted patent-and-black-box culture, and it continues a pattern Baker's lab has set since AlphaFold—the open-sourcing of protein design tooling keeps accelerating. For the technically inclined reader, this may be the most striking development in the entire story: a Nobel-caliber lab has turned antibody design into a repository anyone can clone.

The commercial stakes match the technical achievement. More than 160 antibody drugs have been approved worldwide, spanning oncology and immunology—the largest therapeutic areas—and some of history's biggest blockbusters (Humira, Keytruda) are antibodies. If antibody design gets fundamentally faster, the impact compounds at scale.

The limits deserve equal clarity. RFantibody is at the academic validation stage. Between "an atomically accurate design in silico" and "a developable drug" sit wet-lab confirmation (expression, purification, measured affinity), immunogenicity assessment, process scale-up, and clinical safety. AI delivers a high-quality starting point—compressing years of screening into weeks of validation work—but the finish line has not moved.

With both molecules on the table, the central question can be answered head-on: how much time has actually come out of the cycle?

Part 4 — Is AI Actually Accelerating Drug Development? What the Data Shows

Three layers.

Layer one: front-end acceleration is well documented.

AI-assisted virtual screening can search much larger chemical spaces than conventional experimental screening and can reduce the number of compounds that need to be tested physically.
AI-assisted lead optimization can shorten parts of the design–test–redesign cycle, although the size of the gain varies substantially by program and workflow.
One widely cited comparison claims that AI-designed drugs have Phase I success rates above 80%, versus roughly 52–63% for conventional drugs (Source: BCG, 2024). This figure requires particular caution: most AI-designed assets remain in early-stage development, samples are small, definitions vary, and survivorship bias is a serious concern. Treat it as a directional claim rather than a settled industry benchmark.

Layer two: the back-end ceiling still stands. Overall drug development remains highly attritional. In the BIO/QLS/Informa analysis of 2011–2020 programs, the probability of a program progressing from Phase III to regulatory submission was 57.8%, while the overall likelihood of approval from Phase I was 7.9% (Source: BIO/QLS/Informa, Clinical Development Success Rates). Human biology's complexity, patient enrollment, and long-term safety follow-up still impose constraints that AI cannot simply compute away. As of mid-2026, no AI-discovered and AI-designed drug had reached regulatory approval. Rentosertib is the most advanced prominent example, with its Phase III trial now underway and primary completion currently estimated for late 2029.

Layer three: regulators have moved from watching to rule-making. On January 14, 2026, the FDA and EMA jointly released Guiding Principles of Good AI Practice in Drug Development, a common set of ten principles covering areas including human-centric design, risk-based approaches, data governance, model performance, and lifecycle management (Sources: FDA and EMA websites). The document is guidance rather than binding regulation, but the joint framework signals that AI in drug development has moved beyond isolated corporate experiments into a phase where common expectations for evidence, validation, and oversight are being established. The EMA followed in June 2026 with its AI Observatory Report, tracking the use of AI across the European medicines regulatory network.

One closing observation for this section, and it runs against intuition: compressing the front end may increase pressure on the back end. The logic is direct—if AI makes it cheaper to generate and evaluate candidates, more programs could move into downstream experimentation, and every program carries its own wet-lab experiments, CMC (chemistry, manufacturing, and controls), preclinical toxicology, and regulatory documentation. AI may therefore shift more of the industry's bottleneck toward physical experiments, clinical operations, manufacturing, and compliance verification.

Cycle restructuring always brings a restructuring of labor. The next section maps it onto actual jobs.

Part 5 — Will AI Replace Medical Writers? Mapping the Impact on Pharma Jobs

Two macro datasets first, then the specific roles.

Karpathy's AI occupational exposure scores (March 2026): tasks across 342 US occupations from Bureau of Labor Statistics data were scored for AI exposure. Screen-dependent occupations cluster at the top—medical transcriptionists 10/10, office clerks 9/10—with roughly 60 million American jobs scoring 7 or above. (The project page was taken down minutes after going live, but the underlying BLS data is public and the results were widely reported.)
Anthropic's occupational analysis: Claude's observed exposure covered a large share of the tasks associated with computer programmers, while medical records specialists and market research analysts also appeared among highly exposed occupations. The important distinction is that task exposure measures where AI can perform or assist with work; it does not mean that the corresponding jobs have already been automated away.

Now the pharma-specific map, ordered by impact intensity:

1. Medical writers—among the most exposed. The AI medical writing market stood at roughly $1.4 billion in 2025 and is projected to reach $4.5 billion by 2035, a CAGR of 12.3% (Source: market research). The academic signal is harder to ignore: an arXiv study analyzed 1.19 million open-access biomedical papers in PubMed Central and estimated that 77% of papers in its 2025 sample showed an excess of vocabulary associated with LLM-assisted writing, up from 52% in 2024 and 19% in 2023 (Source: arXiv preprint). The estimate is based on changes in word-frequency patterns, not direct disclosure or detection of AI use in individual papers. On the industry side, Parexel—a major US CRO—has fully adopted structured content authoring plus GenAI for safety and regulatory documents, with internal writer survey data published (Source: EMWA journal Medical Writing). Template-heavy portions of clinical study reports (CSRs), literature reviews, and protocols are already automated. The remaining human value concentrates in scientific judgment, data consistency checks, and regulatory framing.

2. CRAs and clinical operations support roles—partial automation. AI tooling for data entry, patient-recruitment matching, and centralized monitoring is commercially deployed. Medical-records-type roles appearing in Anthropic's top ten exposed list puts clinical operations' document-heavy workflows in the same category. On-site visits, site relationship management, and handling emergent medical events remain human work.

3. MSLs and medical affairs—tooling upgrade. Conference summary generation, guideline comparisons, and first drafts of HCP responses are all AI-completable. Key opinion leader relationships, live medical judgment, and individualized communication stay with people.

4. Pharmacovigilance—the grind layer automates, the judgment layer holds. Automated extraction of adverse events from literature and case reports into safety databases is a mature application. Signal interpretation, causality assessment, and risk decisions remain human territory. The UK MHRA's position is blunt: AI in pharmacovigilance "will always require human judgment" (Source: MHRA public statements).

5. The winners—computational chemists and AI drug scientists. The most sought-after role family in the industry right now. The real bottleneck is hybrid talent: people who understand both wet-lab science and AI are scarce, and teams from the two backgrounds still struggle to speak each other's language—a structural gap the industry openly acknowledges.

The key qualification for this entire map is that exposure will not translate mechanically into replacement. The most exposed zone is likely to be mid-layer templated cognitive work, where outputs are digital, repetitive, and relatively easy to verify. More defensible ground currently sits toward both ends—the physical experiment end and the accountable-signature end—where real-world execution, scientific judgment, and responsibility remain harder to automate.

That "safe at both ends, exposed in the middle" structure leads directly into the final practical section: what individuals should actually do.

Part 6 — Building AI-Era Skills in Pharma: A Career Playbook for Medical Writers, CRAs, and Scientists

6.1 A framework first

A useful way to assess any role's AI exposure is to look at three dimensions:

Screen dependence: does the work product exist 100% in digital form?
Output verifiability: can correctness be judged quickly by machine or rule?
Signature accountability: do regulators or the law require a human to answer for the result?

The Karpathy/BLS scores confirm the framework: medical transcriptionists (pure screen, pure transcription) score 10/10; surgeons (physical operation plus accountability) score very low. Mapped onto pharma: safety lies at the two ends—bench work and accountable sign-off. Danger sits in the templated cognitive middle. Every recommendation below derives from this frame.

6.2 Medical writers: from author to verifier

The data is already on the table—77% of the literature shows AI fingerprints, and Parexel has GenAI embedded in production workflows. The relative value of first-draft production is declining. Verification, scientific judgment, and regulatory framing are becoming more valuable.

Concrete moves, in priority order:

Shift your skill center from "producing documents" to "scientific judgment plus source verification." An AI hallucination in a regulatory document is fatal: one fabricated reference or one mismatched data table can get an entire submission rejected. Line-by-line verification of AI output is the new moat—and it sits precisely in the "signature accountability" dimension that AI cannot enter.
Learn AI output validation methodology in GxP environments. Traditional computerized system validation (CSV) is extending toward AI systems: how do you demonstrate that a generative model's output meets GxP requirements, how do you manage model change control, how do you leave auditable validation records? This methodology is an open market right now. Early practitioners hold pricing power.
Read the FDA/EMA ten AI principles from January 2026 closely. Regulatory texts are dry, but they can be one of the cheapest paths to professional differentiation. People who understand both medical writing and the emerging AI regulatory framework are still likely to be relatively scarce.

6.3 CRAs and clinical operations: from visit executor to process architect

Data entry, centralized monitoring, and patient matching are automated; medical-records-type roles sit in Anthropic's top ten exposed list. What remains human: on-site judgment and process design.

Concrete moves:

Migrate toward the "AI + clinical" hybrid direction. Learn to review AI-generated monitoring plans and to run risk-based monitoring with AI tooling—in effect, manage AI as your junior CRA and move up to overseeing its output.
Build data literacy. The goal is reading AI system outputs and tracing data-quality issues to their source. Note the scope: interpreting model output, not building models.
The transition path already exists. Clinical operations → digital clinical operations / clinical technology roles are being staffed up across major CROs and sponsors. Internal transfers cost far less than job-hopping.

6.4 Wet-lab scientists: your data is becoming the most valuable asset

A counterintuitive claim: wet-lab skills may become more valuable in an AI-intensive discovery workflow.

The logic lives inside lab-in-the-loop: models generate hypotheses → experiments test them → data flows back into the computational system. Roche has publicly described a large-scale AI infrastructure strategy, with more than 3,500 NVIDIA GPUs across its hybrid-cloud AI factory, and has positioned this infrastructure across the therapeutics and diagnostics value chain (Source: Roche). The company also describes a lab-in-the-loop approach in which computational and experimental capabilities work together. Model iteration depends on the return flow of high-quality experimental data. Garbage in, garbage out still applies—and in an increasingly automated discovery system, the quality and structure of experimental data become more consequential.

Concrete moves:

Understand how your experimental data gets consumed by models. Learn structured, standards-compliant recording (FAIR principles: findable, accessible, interoperable, reusable). The same experiment, recorded with proper metadata, can double in data value.
You do not need to write models, but you need to talk to the people who do. Understand the basics of virtual screening, ADMET prediction, and generative chemistry well enough to judge whether a model's chemical suggestions make sense. Translation ability between wet and dry teams is scarce.
Stack bonuses: design of experiments (DoE) plus automation equipment operation. High-throughput data from automated labs is exactly what models crave.

6.5 An industry-level warning: the talent pipeline is breaking

A January 2026 study in Nature points to a related structural risk. Analyzing more than 33 million research teams, the authors found that AI-adopting teams were smaller on average, with the number of junior scientists falling more sharply than the number of established scientists. The study found an average reduction of 1.33 scientists per AI-associated team, with junior scientists decreasing from 2.89 to 1.99 on average. The finding does not show that AI is eliminating junior researchers outright, but it raises a harder question for the industry: if AI absorbs more of the routine analytical work through which early-career scientists traditionally learn, where does the next generation of senior expertise come from?

This is a structural problem beyond any individual's control, but both sides have moves available:

Early-career individuals: deliberately seek out the parts AI does not do—wet-lab rotations, sitting through a complete regulatory submission, attending agency meetings. These experiences are becoming scarce rather than obsolete, and scarcity means bargaining power.
Organizations: reinvest the time AI saves explicitly into junior talent development—mentoring, rotations, real-project exposure. Cutting training costs borrows against your expert supply chain. The bill comes due in a decade, and it always comes due.

AI is changing more than the job list. It is changing how this industry passes knowledge forward. Learning to use AI is the passing grade; rebuilding how people develop under human-machine collaboration is the deep water.

Part 7 — The Verdict Window: What to Watch in AI Drug Discovery

Three things to watch over the next five years. They will decide whether every judgment in this article holds up.

Rentosertib's Phase III results (with primary completion currently estimated for late 2029): a major test of an AI-discovered and AI-designed drug at late-stage clinical development. The result will provide an important data point for judging whether the speed achieved in discovery translates into clinically meaningful outcomes at scale.
The maturation of FDA/EMA frameworks: how quickly ten principles become enforceable detail will determine how deep AI penetrates GxP workflows.
The commercial closed loop of the first AI-native pharma companies: whether any firm sustains itself on AI-derived pipelines rather than continuous fundraising—the only real test of a business model.

The front end has demonstrably accelerated. The back-end ceiling still stands. The division of labor in between is being rearranged now. The molecules are generatable; the evidence will arrive on clinical trial timelines, and it will be worth waiting for.

References

U.S. Food and Drug Administration (FDA) & European Medicines Agency (EMA), 2026 Guiding Principles of Good AI Practice in Drug Development https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development
Bennett, N. R., Watson, J. L., Ragotte, R. J., et al., 2026 Atomically accurate de novo design of antibodies with RFdiffusion https://www.nature.com/articles/s41586-025-09721-5
Ren, J., et al., 2025 A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models https://www.nature.com/articles/s41587-024-02143-0
Hao, Q., Xu, F., Li, Y., & Evans, J., 2026 Artificial intelligence tools expand scientists’ impact but contract science’s focus https://www.nature.com/articles/s41586-025-09922-y

Author Note: The interesting question is no longer whether AI can generate useful drug candidates. It is what happens when generating candidates becomes cheaper and faster. The value of human work may move further downstream—to experiments, verification, judgment, and accountability.