Introduction
Data Management
Start-Ups
AI In Pharma
Download PDF
Generative AI in Science—
A New Era of Chemistry?
E
Every October, scientists eagerly await the news of who gets a call from Stockholm. The Nobel Prize is among the highest honors for a scientist, acknowledging the “contributions that have conferred the greatest benefit to humankind.” Unlike other high-profile awards that focus on recent achievements, the Nobel Prize recognizes work that has had a long-term impact. Often, the prizes are awarded for research that began decades earlier.
But 2024 was different. In addition to announcing the laureates, the Nobel committee signaled a transition to a new age in science. Generative artificial intelligence (AI)–related research won not one but two prizes—an unprecedented accomplishment for a technology that achieved mainstream adoption less than 2 years prior.
David Baker, John Jumper, and Demis Hassabis shared the 2024 Nobel Prize in Chemistry. While Baker is a professor who has studied computation protein design and structure prediction for many years, Jumper and Hassabis are not traditional academics. They led a research team at Google DeepMind that developed AlphaFold, an AI tool for predicting protein structure. AlphaFold2—the second generation of the model—achieved over 90% accuracy at the 2020 Critical Assessment of Protein Structure Prediction (a competition for predicting protein structures)—a massive advance over all previous models. Protein structure prediction, a problem that had baffled biochemists for over 50 years, had been essentially solved.
Protein structure prediction, a problem that had baffled biochemists for over 50 years, had been essentially solved.
While few argue with the technical brilliance of AlphaFold, there is disagreement over whether the technology has yet proven itself worthy of a Nobel Prize. Does this innovation count as science, given that the insight is captured in an algorithmic black box rather than a scientific theory? Where are the “contributions” that have come from AlphaFold? Should companies like Google influence scientific progress?
In April of 2025, Hassabis said in an interview with 60 Minutes that all diseases could be cured in the next 10 years thanks to AI. Many scientists, including chemistry commentator Derek Lowe, pushed back on this bold prediction. Reactions like this to Hassabis and AlphaFold reflect the concerns many in the chemical sciences have toward generative AI.
AI technology in chemistry is not new. Software for predicting physicochemical properties, toxicity, and synthetic routes uses a combination of machine learning and AI algorithms. These tools have been widely used since the 1990s and can be found everywhere, from high school classrooms to the largest pharmaceutical companies.
The release of ChatGPT at the end of 2022 created a wave of excitement about generative AI. Instead of being limited to making predictions within the confines of a training set, this technology can generate new content. This could have radical implications on how we work in the chemical sciences. Drug discovery, formulation testing, and material design are time-intensive and expensive research endeavors. Generative AI tools could reduce the amount of “trial-and-error” experiments done in the lab, allowing chemists to focus on synthesizing and fine-tuning a small number of high-potential compounds and materials.
However, the implementation of generative AI in chemistry is still limited. A survey of American Chemical Society members and contacts in the fall of 2024 found that only 16% feel generative AI substantially improves their productivity in their work or research. A mere 25% of respondents are using GenAI chat tools more than once a month. Moreover, 68% of respondents were concerned about the ethical and legal issues surrounding AI. This skepticism was found across age groups and in both industry and academia.
View the Responses
The American Chemical Society surveyed members and contacts in fall 2024 about their opinions about generative AI. Click the tabs to review the data.
“I am concerned about the ethical and legal issues surrounding generative AI.”
All
Industry & Academia
Ethics
Productivity
Use
Age
0%
25%
50%
75%
100%
All
n = 530
10%
5%
17%
25%
43%
1 - Strongly Disagree
2 - Disagree
3 - Neutral
4 - Agree
5 - Strongly Agree
Given the complexity and potential impact of generative AI, now is an opportunity to lay out the current state of this technology in the chemical sciences.
What are the barriers to the adoption and implementation of generative AI in chemistry?
How are start-ups using this technology to develop breakthroughs?
Will generative AI allow us to rethink how we research and commercialize medicine?
This report offers an overview of how chemistry is transitioning to the age of generative AI and describes the challenges and opportunities we can expect in the near future.
Chemistry’s
Data Management Problem
Chemistry data is special. While data, in general, is sometimes regarded as a high-volume, low-value commodity, large high-quality chemistry datasets are rare and valuable. Researchers have long relied on chemistry data for breakthroughs in medicine, chemical engineering, material sciences, and other fields.
The demand for chemistry data is also on the rise. This demand is partly due to an increase in interdisciplinary research, according to Leah McEwen, chemistry librarian at Cornell University. "I work with so many people beyond chemists,” she says. “Chemical information and data are relevant to a lot of fields. It’s not always clear for those other fields where to find that information.”
Unfortunately, chemistry data are not easy to access, both for humans and computational tools.
Files can be large and are often locked in proprietary formats. For example, high-resolution mass spectrometers can produce gigabytes of data per run, often in formats that computers cannot read without specific software.
Even data presented in a journal article or clinical trial are often not in a format that is usable for other researchers.
Unpublished samples, negative results, and methodological details about experiments are often left out of datasets.
Chemistry data is also fussy. Precise quantities can vary wildly based on details of experimental conditions, such as the instruments or equipment used. “Sometimes it’s the glass—did you use glassware or some other material?” McEwen explains.
Challenges in managing chemistry data existed before generative AI, but this novel technology has magnified the issue. Generative AI requires massive training sets of high-quality data. To realize all its potential benefits in chemistry, we will need better data. Experts like McEwen believe chemists need to implement machine-friendly data management.
Playing FAIR
In 2019, C&EN reported that academic productivity was growing—scientists were publishing more articles, books, and datasets—but sharing and reusing data from experiments was not keeping pace. Many chemists began advocating for researchers to implement “FAIR” principles of data management, meaning the data are findable, accessible, interoperable, and reusable.
What is FAIR?
Data with unique, machine-readable identifiers. This data includes metadata, such as the experimental conditions used to gather the data. Data and metadata should also be searchable.
indable
F
Once users find the required data, they need to know how to access it. Data and metadata should be readable by both humans and machines, and metadata should remain available even after data access expires.
ccessible
A
Data that is compatible for use across multiple platforms using common data languages, standards, and terminology (for example, authority constructs).
nteroperable
I
Data that has clearly explained usage licenses and details about its origin. Data should also be described so that it can be replicated if necessary.
eusable
R
Sources: Wilkinson et al., 2016 | National Institute of Health | Go Fair
FAIR data is a must for generative AI algorithms to solve problems in chemistry because these standards create a foundation that’s accurate and functional. Generative AI applications like ChatGPT were trained in part by text and images available on public websites. Not only did this process involve a massive amount of data, but this data was also available in a machine-readable format. By contrast, generative AI tools in chemistry often do not have access to structured scientific data along with the relevant metadata for training.
Since 2019, many changes have occurred in how academic journals manage chemistry data. Some journals, such as Nature, require scientists to upload their data. ACS journals encourage FAIR data for all paper submissions, and three of the journals require statements that describe the availability of the data. ACS plans to require such statements for all journals in the future.
“I’m glad to see the awareness at the journal level and the funder level of how important this is,” says McEwen. And these initiatives also have the backing of major scientific institutions. McEwen recently led a partnership between CODATA (the Committee on Data of the International Science Council), RDA (the Research Data Alliance), and IUPAC (International Union of Pure and Applied Chemistry) called the WorldFAIR Chemistry. Her interdisciplinary team delivered prototypes of FAIR-friendly practices: recommendations for chemistry FAIR data policies, a training resource, and other services.
Implementing these changes could help accelerate innovation, especially in the private sector. “I think a big potential winner out of this is industry,” says McEwen. “Data are so valuable for them to translate it into things they can do. . . and they have to optimize processes. They need it to be reproducible and consistent.”
Implementing these changes could help accelerate innovation, especially in the private sector. “I think a big potential winner out of this is industry,” says McEwen. “Data are so valuable for them to translate it into things they can do. . . and they have to optimize processes. They need it to be reproducible and consistent.”
How is AI reshaping scientific research?
Register for the “Navigating AI in Science” 7-week email series to learn more, featuring insights from CAS experts.
Register Now
Practical barriers still exist, like the cost of implementing the right systems. For example, some journals have data editors who review submissions dataset by dataset, such as ACS’s “help desk” scientists who assist with checking data. “That is good. I’m glad they’re doing that,” says McEwen. “But we also need some semiautomated ways to validate data” to improve efficiency.
Human experts, however, will remain integral to compiling, reviewing, and correcting datasets. CAS, a division of the American Chemical Society, compiles data from published papers and patents for its Content Collection. Industry clients partner with CAS to develop curated datasets that merge in-house data with CAS data and resources for training algorithms as well.
“You can start to look at patterns,” says Molly Strausbaugh, director of scientific content and commercial chemistry at CAS. “We have a huge human-curated reaction set here. There’s a great opportunity for training.”
A key component of the CAS approach to curating scientific data is integrating “dark data” into datasets that can train generative AI. Dark data is the unstructured and unincorporated information that researchers already have, such as data from lab notebooks or figures in academic papers that are not machine-readable—data that are available but inaccessible. Data curators at CAS seek out this dark data and digitize it, which makes it usable for scientific research.
Chemical information has come a long way since the printed paper data books from 100 years ago. Strausbaugh sees AI as a post-digital step in the evolution of CAS—a step that makes chemistry data even more practical. “You have more and more power to query that data. And so that’s becoming more powerful now—all the different ways that you can use search, build data relationships, and derive more insights,” she says.
Overcoming the barriers
McEwen believes that academic chemistry has been able to get away with lackluster documentation for a long time because academic research is often exploratory. But as AI plays more of a role in chemical research, documentation is becoming more important.
This change represents a potential cultural transition with the rise of AI. Will more researchers embrace FAIR data?
This change represents a potential cultural transition with the rise of AI. Will more researchers embrace FAIR data?
Industry could accelerate this cultural shift by funding the required infrastructure. McEwen would like to see data repositories that are “chemically intelligent,” meaning they can make sense of chemical structures, and they know how to represent chemicals. She also mentioned affordable data tools like electronic notebooks, which are common in industry but not academia, and data stewards employed to support student researchers. “More and more research librarians are being trained for data management skills,” says McEwen.
While it is uncertain how these cultural and technological changes will affect chemistry research, researchers must also deploy AI in ways that boost productivity and innovation without sacrificing performance. “We’re very careful about that because we want to be using it ethically and responsibly,” says Strausbaugh. “Sometimes you’ve got to go slow to go fast.”
A cultural phase transition
Tools and platforms leveraging generative AI in chemistry face challenges in accessing the quality and quantity of chemistry data needed, but that hasn’t stopped start-ups from finding innovative uses for AI. Whether it is the automation of laboratory research or designing new materials, generative AI is helping companies accelerate the rate of innovation in the field of chemistry.
Meet the AI Start-Ups
Driving Innovation in the Chemistry Enterprise
Area of Chemistry:
Chemistry technology
Chemify
Glasgow, Scotland
Founder:
Lee Cronin
Founded:
2022
Chemify is a programmable chemistry company focused on digitizing and automating the field of chemistry to accelerate molecular discovery and chemical synthesis.
They achieve this through the Chemify AI-Chemputation Platform which combines AI / ML chemical intelligence, Chemify’s universal chemical programming language, and Chemify’s automated synthesis robots.
The company has automated a diverse range of chemical reactions, which allows for greater exploration of chemical space, which can be directly synthesized into real molecules.
Visit site
Area of Chemistry:
Materials science
Xatoms
Toronto, Canada
Founders:
Diana Virgovicova, Kerem Topal Ismail Oglou, and Shirley Zhong
Founded:
2024
Xatoms is tackling global water pollution by developing photocatalysts that can purify water by treating contaminants such as bacteria, viruses, or heavy metals.
Xatoms’ AI algorithms analyze chemicals to identify stable, cost-effective, and environmentally safe molecular structures that can be used as photocatalysts. Then, the AI model uses quantum chemistry simulations to predict which candidates will have the best light absorption efficiency and redox potential.
Once the most promising candidates are selected, scientists prepare them in the lab and then add the experimental results to the AI training set.
Visit site
Area of Chemistry:
Materials science
Dunia Innovations
Berlin, Germany
Founder:
Alexander Hammer, Ahmed Ismail, and Marcus Tze-Kiat Ng
Founded:
2022
Dunia Innovations uses AI and robotics to accelerate the discovery of new materials essential for clean energy.
Developing effective catalysts for applications like green hydrogen production can take decades. Dunia’s platform combines physics-informed AI algorithms with an automated “self-driving” lab to quickly design, synthesize, and test new catalysts. The AI suggests promising material candidates, and robotic systems create and test them.
Dunia aims to automate the research cycle and bring scalable, carbon-neutral technologies to market more quickly, enabling the sustainable production of fuels, fertilizers, and plastics.
Visit site
Area of Chemistry:
Medicinal chemistry
PostEra
Boston, USA
CEO:
Aaron Morris
Founded:
2019
PostEra is the creator of Proton, an AI platform for medicinal chemistry that combines innovations in generative chemistry and synthesis-aware design to accelerate the discovery of new medicines.
PostEra is advancing an internal pipeline and small molecule programs through partnerships with biopharma companies. They have closed over $1Bn in AI partnerships, including four multi-year agreements with Pfizer and Amgen.
PostEra also leads an antiviral drug discovery center for pandemic preparedness, funded by the NIH.
Visit site
Area of Chemistry:
Medicinal and computational chemistry
Iambic Therapeutics
San Diego, USA
Founders:
Tom Miller and Fred Manby
Founded:
2020
Iambic Therapeutics uses AI to simulate how molecules interact with proteins to design drugs for different diseases. Their technology also aims at targets that were traditionally considered undruggable.
Iambic’s AI models explore multiple drug candidates in parallel, simultaneously optimizing for potency, safety, and clinical performance. Their AI models predict how a molecule binds to a protein, forecast clinical properties, and learns from experimental data.
Iambic plans to make drug discovery and development faster, smarter, and more predictable by combining deep learning with the laws of chemistry and physics.
Visit site
Area of Chemistry:
Chemical technologies
Kebotix
Cambridge, USA
Founders:
Alán Aspuru-Guzik, Christoph Kreisbeck, Semion Saikin, Dennis Sheberla
Founded:
2017
Kebotix builds AI-controlled automated lab to discover new materials and chemical compounds faster.
Their platform acts as a “self-driving” lab: AI models predict what molecules or materials might have the client’s desired properties; robotic systems run the experiments; and the data feeds back into the AI system to improve future predictions.
This closed-loop system helps to reduce the slow, expensive trial-and-error process of materials R&D.
Visit site
Area of Chemistry:
Chemical technologies
Artificial
Palo Alto, USA
Founder:
David Fuller
Founded:
2017
Artificial builds software that connects all the tools in a lab so scientists can design experiments, schedule tasks, and keep track of results in one place.
Their system lets researchers monitor reaction progress in real time, and AI can take over repetitive steps like adjusting temperatures or mixing reagents. It learns from past experiments to help optimize reaction conditions. The AI model can also help predict which experiments are likely to work before running them.
Artificial built their AI to work with whatever equipment a lab uses so chemists can focus on science while the system handles the rest.
Visit site
Area of Chemistry:
Materials and manufacturing
Orbital Materials
London, England
Founders:
Jonathan Godwin, James Gin-Pollock, and Daniel Miodovnik
Founded:
2022
Orbital Materials uses AI to design new materials with tailored properties and integrates them into complete systems.
Their AI model aids material development and optimizes the layout of equipment, piping, and process flows. Once a design is ready, Orbital fabricates these systems, preparing them for real-world application. For example, Orbital has built liquid coolant systems for data centers and is collaborating with Amazon Web Services, Nvidia, and others to test these systems.
By combining chemistry, AI, and engineering, Orbital aims to accelerate the development of technologies that enhance efficiency and sustainability.
Visit site
Area of Chemistry:
Chemical technologies
Basetwo
Toronto, Canada
Founders:
Thouheed Abdul Gaffoor and Thamjeeth Abdul Gaffoor
Founded:
2021
Basetwo uses AI to help manufacturers run chemical processes more efficiently. Their platform learns from the data collected during manufacturing, such as temperatures and pressures, to understand how reactions and material flows behave inside a system.
The AI system can suggest changes to improve yields, reduce waste, or increase consistency. Engineers can test these ideas in a digital twin of their system before applying any changes to the actual process.
This technology eliminates the guesswork of running a complex chemical process and makes better decisions faster.
Visit site
Area of Chemistry:
Chemical technologies
SandboxAQ
Palo Alto, USA
Founder:
Jack Hidary
Founded:
2022
SandboxAQ’s AQChemSim platform uses AI and quantum simulations to model molecular behavior at the atomic level.
Quantum simulations - virtual experiments that follow the rules of quantum mechanics - are used to predict properties such as stability, reactivity, and toxicity. Researchers use this data to train Large Quantitative Models (LQMs), which can quickly generate predictions for new molecules, accelerating the exploration of chemical space.
By combining supercomputer-powered quantum simulations with generative AI, AQChemSim helps accelerate the discovery of materials for applications across many industries.
Visit site
CEO:
Ashish Kulkarni
Processes like drug research and development (R&D) are known to be expensive and time-consuming. A 2020 study published in the medical journal JAMA found that bringing a new medicine to market costs an average of 1.1 billion USD. From a business value perspective, the use of AI can lower costs and speed up discovery when applied the right way, says Shirley Zhong, Xatoms chief operating officer (COO).
Several start-ups are using generative AI to design and simulate test materials before synthesizing them, which can accelerate the discovery and development cycle. Companies such as Orbital Materials, Sandbox AQ, and Xatoms use AI to design novel materials for clean energy, manufacturing, and water purification, among other applications.
Generative AI is not just a tool for making predictions—it can also change how chemistry is carried out in the lab. In a traditional academic laboratory, a researcher prepares, executes, and analyzes experiments manually. Researchers can increase productivity by outsourcing mundane and repetitive tasks to robotic systems.
While laboratory automation has been available for several years, pairing this technology with AI has the potential to expand the work that a single scientist can do. Companies like Chemify are attempting to realize this vision by developing automation and software tools that allow users to plan and run experiments with limited human involvement. Kebotix is focusing on applying a similar approach to the materials sciences, allowing companies to create a “self-driving lab.”
In the field of medicinal chemistry, companies like PostEra are developing AI platforms that can design molecules to interact with drug targets. “Rather than one drug and one pathway to the market, we have a reproducible flywheel that can continuously turn out new scientific advances,” says Alpha Lee, chief scientific officer of PostEra.
Another company looking to tackle challenges in drug discovery is Iambic Therapeutics, which uses generative AI to simulate how potential drug molecules will interact with proteins. Their Enchant model “learns to get better at predicting clinical properties and molecules by being trained on more preclinical data,” according to Fred Manby, Iambic cofounder and chief technology officer (CTO).
Catalyzing change
Developing novel materials has many of the same challenges as drug discovery and R&D. Novel polymers and catalysts are especially critical for meeting the United Nations sustainable development goals and country-specific environmental and energy goals. Yet designing, synthesizing, and testing these materials can take months, years, and sometimes even decades.
Applying generative AI in the chemical sciences offers tremendous opportunities, but there are also obstacles. As with most start-ups, money is a constraint. “I think time and pricing of how much it costs to run AI models are the biggest bottlenecks we see,” says Diana Virgovicova, Xatoms CEO and founder.
Money is not just necessary for human resources and computational power, but also the chemical data needed to train models. “AI needs good data, and good chemistry data—like material sciences data—is very expensive,” says Semion Saikin, COO of Kebotix.
Adam Skiredj, Chemify’s director of business development, agrees that data quality is essential. “Generative AI often relies on the quality of input data. Biased data can result in unreliable outputs and may produce compounds that are not synthetically possible or have a low chance of being active compounds.”
A unique challenge of chemistry AI start-ups is cutting through the hype around large language model (LLM) chatbots such as ChatGPT, Claude, or Perplexity. “While we believe LLMs are very powerful for certain tasks, we believe that they actually fall short in being able to really supercharge our ability to explore uncharted territories in the chemical industry,” says Arman Zaribafiyan, Sandbox AQ head of strategic alliances.
Beyond proving that their AI technology works, companies also need to show that it’s cost-effective. “You’re trying to demonstrate not only that you can actually find drug molecules that can be put into human patients, but also that you can do so in a framework that’s reducing the capital cost of developing medicines,” says Manby. “That’s very challenging.”
Challenges to
Challenging the Status Quo
Today’s chemistry start-ups are not just driving technical innovation, they are also integrating AI into business practices such as operations, marketing, and data management. They are also changing the way industry thinks and collaborates.
Start-ups like Sandbox AQ anticipate that AI will influence “the entire chemistry value chain, whether we are looking at chemicals, industry, pharmaceutical, the energy sector, or automotive,” says Zaribafiyan. This effect will create the space for partnerships that normally may not have existed, especially between industry and start-ups.
PostEra has created a close synergy among chemistry, engineering, and AI. Being a start-up has allowed them to scientifically control what goes into their platform and have a rigorous understanding of the models, explained Lee. “It’s no longer just chemists, biologists, and physicians as a core team, now we’re bringing in machine-learning scientists and engineers working in the same house . . . that is really important, and in a start-up, we are able to nurture and foster that interdisciplinary collaboration, which feeds into the science of everything being built.”
Updating business processes could be especially groundbreaking in the chemical industry. A 2024 McKinsey & Company report found that generative AI could generate between $80 billion and $140 billion in value for the chemical industry. Much of this potential was outside research and development and focused on finding new applications for existing chemicals, customer acquisition, and other support functions.
“Many big companies, they bring an AI group into their existing businesses. We are building a company around AI. We put it in the core,” says Saikin. “It changes the whole process, how you discover and develop.”
While it remains to be seen how these innovative companies will influence the broader chemical sciences, established companies are taking notice. Their impact is perhaps most clear in the pharmaceutical industry, which has been working to adopt and adapt many of these technologies.
Innovation
Beyond the Bench
The promise and the hype of AI for Big Pharma
The tools are here.
The investments have been made.
And now, we wait.
Advancing medicine has been hailed as one of generative AI’s most exciting applications. The technology is especially appealing for drug discovery because of its potential to develop successful drug candidates faster and at a lower cost than traditional drug discovery. With applications such as designing compounds, optimizing lead candidates, and predicting absorption, distribution, metabolism, and excretion (ADME) properties, AI may be poised to hasten the path from bench to bedside.
Chemists have used machine learning to design drug compounds for decades. Until recently, this practice was called de novo compound design. “For those that work in the area, the advances have felt more gradual,” says Robert Abel, executive vice president and chief scientific officer of platform at Schrödinger, a computational chemistry platform company.
Even with generative AI, human input remains critical in de novo design to choose which hit or lead series to optimize, select the right target, and define which in vitro properties to aim for. “No matter how powerful the method, you need humans to decide where to aim it and how to aim it,” Abel says. “A medicinal chemist can look at a molecule and make judgments about how challenging it will be to synthesize that are not fully addressed with computational analysis.”
Recent improvements to the availability of cloud-based computing power and the underlying scientific techniques have dramatically improved the performance of de novo design approaches, increasing the volume, precision, and accessibility of these methods. These improvements have fueled interest and excitement when paired with the general hype around generative AI.
Big Pharma’s
Big Bets
But amidst this flurry of interest in the past few years, where is this excitement yielding results? And what are some of the key challenges preventing AI from fulfilling its full potential in drug discovery?
Following the money, it’s clear that Big Pharma is enthusiastic about generative AI’s potential.
Following the money, it’s clear that Big Pharma is enthusiastic about generative AI’s potential. In March 2025, AstraZeneca announced a $2.5 billion investment centered around “a new state-of-the-art AI and data science laboratory” in Beijing.
Pharma is also investing heavily in partnerships with technology firms offering AI-enabled services, such as Schrödinger, which licenses its de novo design software to pharmaceutical companies. In November 2024, the company announced a multitarget collaboration with Novartis for $150 million up-front, with the potential for up to $2.3 billion in milestone payments and royalties. And in early 2024, Novartis and Eli Lilly and Company announced deals with Isomorphic Labs, a drug discovery firm that spun off from Alphabet’s DeepMind. These deals represent just a handful of the major agreements between AI vendors and big pharmaceutical companies in recent years.
2022
2023
2024
2025
Partners on the
Pharma Pipeline
Atomwise signs a multitarget research pact with Sanofi.
August
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
The pharmaceutical industry's race to make use of generative AI is accelerating. Here's how things have evolved since 2022.
A May 2024 analysis by the Boston Consulting Group (BCG) published in Drug Discovery Today estimated that 67 molecules from AI-native biotechs were in human trials. The study found that the success rates for AI-discovered molecules were substantially higher than for conventional molecules in Phase 1 clinical trials and about average in Phase 2. A March 2025 report from BiopharmaTrend, which looked at a smaller subset of companies, estimated 10 molecules in Phase 2 and two in Phase 3, one of which has since been discontinued.
One AI-derived drug that has made it to Phase 2 is Insilico Medicine’s rentosertib, a treatment for idiopathic pulmonary fibrosis that shows encouraging top-line results from a Phase 2a clinical trial in late 2024. Rentosertib is the first drug for which AI was used both to locate the target and select the compound, and Wired has hailed it as “the world’s first fully AI-generated drug.”
Insilico has a separate partnership with Sanofi to advance candidates for up to six targets that could be worth up to $1.2 billion. And in March 2025, Pfizer began Phase 3 clinical trials for ibuzatrelvir, a next-generation COVID-19 antiviral designed with the use of AI.
Beyond drug discovery, Big Pharma is putting generative AI to work in other aspects of drug development, including optimizing clinical trials and regulatory work. Novo Nordisk recently announced a series of partnerships and in-house projects to reduce time spent on documentation. Both Novartis and Eli Lilly have partnered with the generative AI company Yseop to automate the writing and review of clinical trial and regulatory documents.
AI medicine meets
Human Patients
Theoretically, AI could give drugs an edge in the clinic. “If you have a fundamentally better molecule that’s easier to formulate, has a better property distribution than alternative chemical matter than you would have arrived at conventionally, it should be more likely to succeed in clinical trials,” Abel says.
But so far, evidence for an efficacy advantage remains to be seen. While big-budget deals have been flowing, the timeline for drug discovery is long, and only a handful of AI-discovered drugs have advanced in clinical trials.
Robert Abel
Executive Vice President & Chief Scientific Officer of Platform
Schrödinger
A Dose of Reality
Just as life-changing drugs that zip through the clinic have yet to materialize, not every AI tool on the market will live up to its marketing promises. For example, Abel says users should be skeptical of any model that is tested only on historical datasets but is touted as ready to discover innovative, real-world drugs. Because unintended similarities between a method’s training and test sets may lead researchers to overestimate a method’s prospective potential, only robust prospective validation can reliably establish how well a method works under real-world conditions.
Many barriers to generative AI’s adoption remain, including:
Regulatory challenges such as documenting how a drug was developed or why a candidate was selected (the inner workings of many models are difficult to interpret).
Organizational access to computing power and cross-functional training to make the most of the technology’s potential.
Protecting privacy while handling large quantities of patient data.
Use of the technology also requires a cultural shift—what Abel calls a transition from computer-assisted to computer-driven drug discovery. “You need to be willing to allow the de novo design method to be a coequal partner with the discovery team,” he says.
While generative AI drug discovery tools may be daunting, the prospect of making better medications faster and getting them to patients sooner means these are barriers worth addressing. “These methods are continuing to improve,” Abel says. “Addressing that vast unmet clinical need—that’s what has me most excited.”
Contributors
Editorial lead: Jesse Harris
Project manager: Mariam Hussein Agha
Editors: Sara Cottle, Shane M. Hanlon, and Jordan Nutting
Writers: Sara Cottle, Jesse Harris, Max Levy, Ananya Palivela, Alexandra Taylor
Art and UI/UX design: Cesar Caminero
Web producers: Kay Youn
Copyeditor: Kari Hallenburg
Published by C&EN BrandLab
9%
6%
13%
31%
39%
11%
4%
18%
22%
43%
All
Industry & Academia
18-34
35-44
45-54
55-64
65+
Age
Industry
Academia
0%
25%
50%
75%
100%
n = 530
All
Industry & Academia
Age
All
Industry & Academia
Industry
Academia
18-34
35-44
45-54
55-64
65+
Age
“Generative AI tools substantially improve my productivity.”
0%
25%
50%
75%
100%
n = 530
All
Industry & Academia
Age
All
All
Industry & Academia
Industry
Academia
18-34
35-44
45-54
55-64
65+
Age
“How often do you use generative AI chat tools in your work or research?”
Ethics
Productivity
Use
All
26%
35%
23%
12%
4%
24%
33%
11%
12%
3%
24%
33%
11%
12%
4%
9%
75%
11%
5%
7%
73%
15%
4%
8%
78%
11%
4%
Insilico Medicine's rentosertib (formerly INS018_055) enters clinical trials with human patients.
June
Read Story
Insilico used generative AI to find the target and select the compound, and the candidate has been hailed as the first fully AI-generated drug.
Roche's Genentech unit forms an AI computing agreement with Nvidia.
November
Read Story
Genentech plans to use the semiconductor designer’s cloud platform to host its algorithms, aiming to speed up the drug discovery process.
Novartis and Eli Lilly and Company announce deals with Isomorphic Labs at the JP Morgan Healthcare Conference.
January
Read Story
The deals total $82.5 million up front, with the potential for up to $1.7 billion (from Lilly) and $1.2 billion (from Novartis) in milestone payments plus royalties.
Eli Lilly and Company partners with OpenAI to invent new antibiotics, supporting a prior $100 million commitment to the phama giant's AMR Action Fund.
June
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Demis Hassabis and John M. Jumper of Google DeepMind, of which Isomorphic Labs is a spin-off, share half of the 2024 Nobel Prize in Chemistry for protein structure prediction.
October
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Schrödinger announces a multi-target collaboration with Novartis.
November
Read Story
The pact is worth $150 million up front, with the potential for up to $2.3 billion in milestone payments and royalties.
Pfizer expands its partnership with PostEra and will pay the firm up to $350 million to use its software to design new small molecules and antibody-drug conjugates.
January
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Isomorphic Labs broadens its partnership with Novartis to pursue small-molecule therapies for three more undisclosed targets.
February
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Pfizer begins Phase 3 clinical trials for ibuzatrelvir, a next-generation coronavirus antiviral.
March
Read Story
Artificial intelligence played a role in the drug's design.
Open
indable
F
Open
Once users find the required data, they need to know how to access it. Data and metadata should be readable by both humans and machines, and metadata should remain available even after data access expires.
ccessible
A
ccessible
A
Open
Data that is compatible for use across multiple platforms using common data languages, standards, and terminology (for example, authority constructs).
nteroperable
I
nteroperable
I
Open
Data that has clearly explained usage licenses and details about its origin. Data should also be described so that it can be replicated if necessary.
eusable
R
eusable
R
Data Management
Start-Ups
AI In Pharma
n = 530
n = 542
n = 516
n = 530
n = 441
n = 515
n = 548
n = 452
n = 534
Even with generative AI, human input remains critical in de novo design to choose which hit or lead series to optimize, select the right target, and define which in vitro properties to aim for. “No matter how powerful the method, you need humans to decide where to aim it and how to aim it,” Abel says. “A medicinal chemist can look at a molecule and make judgments about how challenging it will be to synthesize that are not fully addressed with computational analysis.”
Open
Data with unique, machine-readable identifiers. This data includes metadata, such as the experimental conditions used to gather the data. Data and metadata should also be searchable.
indable
F
indable
F
Open
Once users find the required data, they need to know how to access it. Data and metadata should be readable by both humans and machines, and metadata should remain available even after data access expires.
indable
F
ccessible
A
Open
Data that is compatible for use across multiple platforms using common data languages, standards, and terminology (for example, authority constructs).
indable
F
nteroperable
I
Every October, scientists eagerly await the news of who gets a call from Stockholm. The Nobel Prize is among the highest honors for a scientist, acknowledging the “contributions that have conferred the greatest benefit to humankind.” Unlike other high-profile awards that focus on recent achievements, the Nobel Prize recognizes work that has had a long-term impact. Often, the prizes are awarded for research that began decades earlier.
But 2024 was different. In addition to announcing the laureates, the Nobel committee signaled a transition to a new age in science. Generative artificial intelligence (AI)–related research won not one but two prizes—an unprecedented accomplishment for a technology that achieved mainstream adoption less than 2 years prior.
David Baker, John Jumper, and Demis Hassabis shared the 2024 Nobel Prize in Chemistry. While Baker is a professor who has studied computation protein design and structure prediction for many years, Jumper and Hassabis are not traditional academics. They led a research team at Google DeepMind that developed AlphaFold, an AI tool for predicting protein structure. AlphaFold2—the second generation of the model—achieved over 90% accuracy at the 2020 Critical Assessment of Protein Structure Prediction (a competition for predicting protein structures)—a massive advance over all previous models. Protein structure prediction, a problem that had baffled biochemists for over 50 years, had been essentially solved.
Pfizer begins Phase 3 clinical trials for ibuzatrelvir, a next-generation coronavirus antiviral.
March
Read Story
Artificial intelligence played a role in the drug's design.
Isomorphic Labs broadens its partnership with Novartis to pursue small-molecule therapies for three more undisclosed targets.
February
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Pfizer expands its partnership with PostEra and will pay the firm up to $350 million to use its software to design new small molecules and antibody-drug conjugates.
January
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Schrödinger announces a multi-target collaboration with Novartis.
November
Read Story
The pact is worth $150 million up front, with the potential for up to $2.3 billion in milestone payments and royalties.
Demis Hassabis and John M. Jumper of Google DeepMind, of which Isomorphic Labs is a spin-off, share half of the 2024 Nobel Prize in Chemistry for protein structure prediction.
October
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Eli Lilly and Company partners with OpenAI to invent new antibiotics, supporting a prior $100 million commitment to the phama giant's AMR Action Fund.
June
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
Novartis and Eli Lilly and Company announce deals with Isomorphic Labs at the JP Morgan Healthcare Conference.
January
Read Story
The deals total $82.5 million up front, with the potential for up to $1.7 billion (from Lilly) and $1.2 billion (from Novartis) in milestone payments plus royalties.
Roche's Genentech unit forms an AI computing agreement with Nvidia.
November
Read Story
Genentech plans to use the semiconductor designer’s cloud platform to host its algorithms, aiming to speed up the drug discovery process.
Insilico Medicine's rentosertib (formerly INS018_055) enters clinical trials with human patients.
June
Read Story
Insilico used generative AI to find the target and select the compound, and the candidate has been hailed as the first fully AI-generated drug.
Atomwise signs a multitarget research pact with Sanofi.
August
Read Story
The collaboration is worth up to $20 million up front with the potential for $1 billion in milestone payments plus royalties.
2022
2023
2024
2025
Partners on the
Pharma Pipeline
The pharmaceutical industry's race to make use of generative AI is accelerating. Here's how things have evolved since 2022.
Once a month or less
Once every 1-2 weeks
Once every few days
At least once per day
5 - Strongly Agree
1 - Strongly Disagree
2 - Disagree
3 - Neutral
4 - Agree
5 - Strongly Agree
1 - Strongly Disagree
2 - Disagree
3 - Neutral
4 - Agree
5 - Strongly Agree
1 - Strongly Disagree
2 - Disagree
3 - Neutral
4 - Agree
5 - Strongly Agree
1 - Strongly Disagree
2 - Disagree
3 - Neutral
4 - Agree
5 - Strongly Agree
Visit site
Visit site
Visit site
Visit site
Visit site
Visit site
Visit site
Visit site
Visit site
