Proceedings of a Workshop—in Brief

Convened February 20–21, 2025

Building Access to Tomorrow’s Medicines
Bringing Together Humans, Robots, and Artificial Intelligence

On February 20–21, 2025, the Chemical Sciences Roundtable of the National Academies of Sciences, Engineering, and Medicine held a workshop, Building Access to Tomorrow’s Medicines: Bringing Together Humans, Robots, and Artificial Intelligence. It examined how artificial intelligence (AI) could potentially revolutionize drug discovery and development, shorten the time for drug development, and eventually personalize medicine at scale. This Proceedings of a Workshop—in Brief summarizes the presentations and panel discussions that occurred at that workshop.¹ Contained herein is a high-level summary of the meeting; recordings from the workshop contain more in-depth discussions on the human-driven interface of AI, data science, and robotics for drug discovery.²

DRUG DISCOVERY: THE FUTURE IS NOW

The initial panel, moderated by roundtable member Andrew White, focused on how AI, machine learning (ML), robotics, and automation are currently being used to progress drug discovery. The presentations did not focus on specific technologies or case studies for the evolving landscape, but a general overview as presented in the proceedings.

AI in Drug Discovery

Alexandra Snyder, head of research and development at Generate Biomedicines, opened the panel by describing a framework her company uses to determine whether AI is being used appropriately in drug discovery. For a medicine to help people, Synder presented it as a word problem. The probability of success is the product of the right target, the right medicine for that target, and be given to the right people. Generate Biomedicines’ framework for discovery focuses on the middle part of this problem.

Different companies focus on different pieces of this framework, and many focus on only one piece, such as identifying new targets for drugs. Generate Biomedicines concentrates on finding the right drug for a specific target, Snyder said. Researchers are using AI to assist in each of these areas—searching for appropriate targets, finding the right drug for a given target, and identifying people who could benefit from a particular drug—and sometimes when a project fails, it is interpreted as a failure of AI when really it may have been researchers’ failure to align the various pieces, she said. For instance,

__________________

AI may have found a potentially effective drug, but the wrong target was chosen.

Two drugs designed by Generate Biomedicines, with assistance from AI, have already been through phase I trials, and a third is about to start, Snyder said. She emphasized that while AI makes it possible to speed the process up, it is still vital to go through all the usual steps of testing and evaluation.

Using AI to Speed Up the Search for Therapeutics

Next, James Li, president of GondolaBio, discussed how his company uses AI in developing therapeutics for genetic diseases with high unmet need, specifically in orphaned and rare diseases. There are more than 7,000 different diseases that affect humans, he said, and most of the search for cures has focused on about 500 diseases that affect relatively large numbers of people. Still, the remaining 6,500 together affect many millions of people, and it is these that are the focus of GondolaBio.

The vision of the company, Li said, is to cure these rare diseases at scale—if 10 diseases can be cured, can it be scaled to 30, then increased to 500? To do this the company uses a “hub-and-spoke model,” where a central platform provides scientific guidance as well as some central logistics and operations support to many individual affiliates that focus on a specific disease or a specific molecule and aim to advance as quickly as possible on that one target.

One of the company focuses is finding the right target, which Li described as “the most important thing in drug development,” the first part of the problem discussed by Synder. One way the company does this is by using AI to interrogate large databases containing genetic and omics data to look for evidence that modulating a particular target would modify a disease. They also look for evidence that modulating a particular target might have negative effects. Sometimes AI produces insights that a human researcher could not predict, Li said, but in most cases AI does only about 80–90 percent of the job. Although, AI does it very fast and provides hypotheses that can be pursued.

One use of AI, for example, is for virtual screening—combing through a database to find chemicals to test for their efficacy against a target. It enables a small research team to do the work that used to take 50 or 100 people, Li said. AI is also used to predict the pharmacokinetic properties of a drug. A chemist may have 50 ideas for which molecules might have useful properties, and AI makes it possible to prioritize five or 10 of them. Yet, another use of AI is to suggest possible synthesis routes and possible catalysts for making a particular molecule.

Using AI to Automate Scientific Discovery

Nicolo Fusi, the general manager of Microsoft Research in Cambridge, Massachusetts, spoke about using AI to automate and accelerate scientific discovery. From an information-theoretic view, he said, there are only two things one can do to make scientific discovery faster and better: Speed up scientific experiments so that the discovery loop from hypothesis generation to experiment to analysis and back goes faster, and draw more information from each experiment. The latter is what Fusi’s group is working on.

“To both accelerate the loop and to get more bits of information per interaction with nature,” he said, “we are designing models that . . . learn to speak the language of nature, meaning they can understand and produce the readouts that we expect, directly in the right language.” For instance, one might build a ChatGPT-like model that not only can understand natural language, but can also design a protein.

One important lesson that the AI community has learned, Fusi said, is that over time, general computation tends to outperform specialized models that have been carefully crafted. “If you build a general-purpose method, and you select your data carefully, and you just let it train, it will do a lot better in the long run.”

Another lesson, he continued, is that AI models don’t need more data in order to perform better; rather, they need less data, but better data (ie. accurate, reliable, expansive). A paper from his lab shows that there are particular ways to design one’s data gathering that will lead to better performance of the AI model.³

Finally, Fusi talked about how Microsoft Research uses AI generative models in its biological research. Whereas most people tend to think of generative models as “very

__________________

fancy databases that you can retrieve knowledge from,” they are actually more like simulators. The latter is how his group uses them. To accelerate the discovery loop, they use generative models to simulate outcomes in nature, and when the outcomes are uncertain, the group runs an experiment to check and ground the model.

Integrating Robotics and AI in Drug Discovery

Jesse Kirkpatrick, co-director of the George Mason University Autonomy and Robotics Center, discussed some of the issues that arise when integrating robotics and AI to carry out drug discovery. Kirkpatrick began by talking about responsible AI.

Responsible AI involves acting responsibly in the design, development, deployment, and use of AI, he said. However, when one wants to put these principles into practice in a specific situation—in drug design, for instance—it is not always clear what is required. Does the AI need to be secure? To be safe? To be reliable? Governable? How does one measure AI performance to determine if it is meeting certain benchmarks?

Kirkpatrick suggested that there are key enablers to designing responsible AI systems and building them toscale across organizations. The first is interdisciplinary thinking and partnerships, since various disciplines are involved in this field. Deep stakeholder engagement is another important enabler, as is thinking about the broader ethical implications of the work. Finally, he said stable, predictable, nonpartisan funding sources are a critical enabler.

Kirkpatrick then spoke about self-driving labs in which AI, robotics, and automation are integrated to create experimental workflows that can proceed with minimum human intervention. Potential advantages of these labs include considerably speeding up workflows, optimizing and reducing chemical use, reducing the number of failed experiments, and enabling breakthroughs in drug discovery and materials science.

Many of the challenges are related to robotics. Robots typically function well with preprogrammed, repetitive motions, but they struggle with novel and unexpected situations. There are also data limitations, which can lead to AI limitations. These systems are also very expensive, so smaller companies generally cannot afford to work with them.

Discussion

Following the opening remarks, a panel discussion was moderated by White to better understand AI and drug discovery from the perspective of the speakers and tie their research areas together. White began by asking what new advances in AI have been particularly exciting, highlighting a very recent publication on using AI as a co-scientist.⁴ Kirkpatrick pointed to the development of DeepSeek, a large language model that performs nearly as well as ChatGPT but whose development cost was only a small fraction of ChatGPT’s development cost. This suggests new possibilities in other areas of AI, such as AI-assisted drug design and development. Li cautioned, however, that so far there has been no easy way to adapt the AI models to drug discovery. Fusi pointed to the introduction of test-time compute into AI models, with test-time compute methods allowing the models to spend more computing power on more complex tasks. This specific AI method will lead to major improvements in the performance of these models, he said. Snyder said she was heartened by a change in emphasis from simply getting more and more data to figuring out which data are most important for solving a particular problem.

Snyder said that at her company, AI plays a major role in identifying and designing molecules that could be useful for specific purposes, but it plays much less of a role in the later steps of narrowing down those possible molecules to a few that will be tested. Those steps follow more of a conventional drug-development process. A big part of the reason, she explained, is the closer a molecule gets to the regulated space, the more thorough and more conventional the process needs to be. She added that it seems likely that AI will be used in the preclinical space at some point, which could benefit the field tremendously, but that has not happened yet.

Li agreed and added that AI is already being used in certain ways in the later stages of development, such as in using AI for improving clinical trial enrollment, identifying patients who may have certain clinical diagnoses, and writing significant portions of trial protocols. Earlier in

__________________

the development process, he added, AI is assisting with safety databases and identifying molecules that seem likely to have unacceptable safety risks.

Next, Fusi addressed the issue of interpretability of AI models. The models often make recommendations for reasons that are not clear to humans, and even when the models are queried about the reasons, they may not be able to explain it. “Models that now are getting to the point in noninterpretable areas where they know better than we do,” he said, “so it’s kind of tricky. They kind of have to dumb it down for us.”

Snyder said that some companies are already working to address the potential uses of AI in assessing safety by modeling the behavior of cells and organ systems; but before it will be possible to skip large-animal testing, it will be necessary to “close the loop” between preclinical data and the attendant large-animal data. “The AI part in some ways is the easy part,” she said, “it’s getting the data all together that I think is one of the major challenges.”

One reason that collecting data is difficult, Li said, is that much of the data are private. This is compounded by the presence of publication bias in the published data—researchers do not generally publish the data from their failures. The most valuable data may be clinical trial data collected on disease manifestations, which are cross-sectional and longitudinal. The granularity of the data is also important—one wants the right cell types and the right stages of development.

Fusi added that the heterogeneity of the data is also important: One wants data from patients of different backgrounds, with different treatments and different responses.

Snyder offered an example of how a large, diverse dataset was collected to answer safety-related questions. Friends of Cancer Research got data from the Food and Drug Administration, diagnostic companies, and pharmaceutical companies and had the data aggregated and analyzed by a third party. The various parties agreed to submit data because they were assured that efficacy data would not be shared. “Everybody benefited from that question being answered that could not have been answered except by that aggregation,” she said.

White spoke about the problems that arise when AI models are trained on scientific articles that may have wrong data or conclusions. Generally speaking, AI models cannot tell good science from bad. One way of filtering scientific papers is by limiting the papers used in training an AI model to those from high-quality, peer-reviewed journals. Another filter is citation count, which is imperfect but still useful. Li added that one can look at whether a paper’s results have been reproduced, but this does not work for very recent work. White added that some researchers deposit fake primary data into databases to support their fraudulent papers,⁵ so relying on primary data is not necessarily an answer.

To conclude, White asked the panelists to explore important challenges and unsolved problems. Li pointed to how interdisciplinary the field is and said the major challenge is finding people with the necessary training and background to be familiar with all of its aspects. Kirkpatrick shared ideas to “create, foster, facilitate, and sustain a competitive robust research ecosystem in the United States and globally.” One way to ensure this, he mentioned, is creating and training future researchers with a robust and predictive ecosystem. Snyder discussed the “translatability problem,” i.e. difficulties involved in the knowledge and understanding gained from cell lines or animal models, to what actually goes on in humans. Fusi said that as powerful as the current AI models are, it will likely take at least another generation of development until AI has the ability to come up with genuinely new ideas.

EXPANDING ACCESS: REMOVING THE BIASES FROM DRUG DISCOVERY

The second panel was chaired by Nicola Pohl, co-chair of the Chemical Sciences Roundtable. The three panelists addressed the question of what it will take to remove biases from AI-assisted drug discovery to ensure that the benefits of that drug discovery will be shared as broadly as possible across society.

Making AI Work for Everyone

Sanmi Koyejo, director of Stanford Trustworthy AI Research at Stanford University, began by explaining how AI has the potential to reduce health care disparities. This

__________________

is despite the fact that more attention has been paid to how AI often amplifies these disparities because of biases in health care datasets.

First, he said when AI systems make biased decisions, they often reflect patterns in historical data, which calls attention to such things as populations that were historically underrepresented in clinical trials. Second, unlike the case with human decision making, AI models can be rigorously tested and refined, allowing for systematic improvement. “I think this allows us to proactively address disparities,” he said. Third, intentional design makes it possible to program AI systems to account for population differences and parameters, such as efficacy and side effects. This may involve recognizing that the impacts of AI can vary dramatically by stakeholder, developing metrics to assess the effectiveness of AI models, choosing appropriate targets, and assembling teams with diverse expertise and backgrounds, which reduces the risk of blindness to outcomes that differ across subpopulations.

Looking to the future, he said that optimizing the effectiveness of AI-assisted drug discovery includes finding a productive balance between innovation and safety, establishing rigorous testing protocols to check for effects across different subpopulations, and investing in collecting representative data across different groups. “We have an opportunity to ensure that AI in drug discovery ends up working for everyone.”

Using Real-World Data to Guide Therapeutic Development

Lynda Chin, the chief executive officer of Apricity Health, spoke about the importance of having the right kind of data to guide research and treatment and, in particular, to develop more precisely targeted treatments for individuals. She raised two key points. First, data is fundamental to progress. “Without the clinical data to annotate any kind of molecular data we generate,” she said, “we really don’t understand what that genomic data is telling us.” Although second, getting high-quality clinical data is very difficult.

As a result, researchers generally do not have good enough data to effectively guide decisions concerning drug discovery and development, which in turn has a negative effect on patients, who end up getting drugs that might work better in others. “We are going to need a whole lot more data,” she said. As a result, Apricity Health was created to collect useful data from routine-care patients for use in drug discovery and development, she continued.

Human data are important in drug discovery and development for a number of reasons. Chin noted, “we can identify the target and we can make molecules, and we can cure cancer in the mouse […] but doing so for humans is a different challenge.” Much of drug discovery and development has traditionally been done in a mouse model, but there are enough differences between laboratory mice and human patients that one cannot assume drugs that work in mice will work the same way—or at all—in humans. Furthermore, there is so much variation in how humans respond to drugs that it is important to have data from many humans to understand in which human populations a drug is likely to work and in which it will not.

One approach to collecting enough data is to use the health care delivery system to generate the right sort of data to be able to understand the response, resistance, and toxicity of different drugs in the general population. Chin discussed that it could offer tremendous benefits to AI-driven drug discovery because the AI models could have large amounts of the necessary data that are representative of the general population. The end result could be a chance to develop individualized therapy if AI technology can accelerate the drug discovery process so that it is possible to make new drugs faster and cheaper.

The bottom line, Chin said, is that while the heterogeneity of the patient population is a challenge, it is also an opportunity to create a situation in which AI technology can really excel—but only if one captures that heterogeneity in the datasets.

Market Obstacles to Increasing Representation of Minoritized Populations

Nick Davis, managing partner of Changer and a scientist and investor in various technologies, spoke briefly about increasing the representation across all demographics in minorities in the development of new medicines. Historically, he noted, minorities have been underrepresented in research, staffing, hiring, and medication development.

One significant issue, he continued, is that incentives drive markets and sufficient incentives are not available to include minorities when developing new drugs. Increasing

the breadth and access of drugs to the entire population has been a topic in clinical trials for 30 or 40 years, he said, but the needed changes have not happened. Addressing the issue will likely involve new approaches.

Discussion

The discussion period began with questions about technical challenges and later moved to social, political, and market challenges.

Pohl asked Koyejo about the technical problems and hurdles in this field. Koyejo answered that that the biggest issues are with the measurement gap, being able to optimize mis-specified or incorrect problems. This major issue manifests because the incentives on the technology side often do not align well with what matters most in real-world terms. Expertise is another problem, he said. In AI-assisted drug development, much of the expertise seems to be in the external issues of data curation, measurement, and specifying targets and “much less so on the internals of what the actual models are.”

Chin added that it is common to see people using whatever they have to answer a question even when it is not the right model, the right data, or the right approach “because that is what they have, or that’s what they are good at or what they are comfortable with.” Drug discovery involves more than the typical data collected by a clinician who is practicing routine standard of care, she said. It also includes such things as genomics information and details of the cellular and molecular responses to a drug or drugs, none of which are found in a standard electronic health record. She noted that Apricity Health works with patients to get their consent for it to run additional tests on blood that is already being collected for clinical purposes. A key challenge is building the platform and capability that makes it possible to collect such data over time—the right data—so that it can be used effectively for drug discovery and development.

Chin said that the data collection and analysis platform she is talking about differs from what an academic researcher might need. To collect data that reflects the diversity and heterogeneity of an entire population, it is important to obtain the participation of many more patients and the submission of much more data, which means one must consider the needs and requirements of the patients to maximize the number who participate. “You have to design a very different research protocol,” she said. “I call it patient-centered research versus researcher-centric research.”

Davis echoed Koyejo’s point by saying there tends to be a mismatch between the AI models being developed and the needs of researchers in the drug discovery and development area. Meanwhile, he said, private companies developing AI tools have increased amounts of funds available, while academic researchers are struggling to find funding.⁶

Pohl then asked the panelists to discuss some of the human challenges to using AI in drug discovery and development. Koyejo said that when people are working to improve AI, they are focusing on things like building the biggest model trained on the most data, but they are not looking to create models able to make the most difference in the real world. This is because the incentives, in the form of recognition and rewards, tend to favor the easily measured aspects of AI models instead of the more difficult-to-quantify aspects such as effectiveness in directing drug discovery and development. The question is, what incentives would be better, and how can institutions be convinced to change to those incentives?

Chin pointed out several human problems. One obstacle is the current incentive system for academic researchers and, in particular, the lack of incentives for carrying out research on and collecting data from small, less studied groups. For example, studying a known mechanism and how it can be generalized to a less studied population will generally not be considered novel enough to get a researcher published in a top-tier journal. Another issue, Chin said, is that patients are generally not given enough incentive to participate either in clinical trials or foundational clinical research.

Pohl then asked Davis to suggest some incentive structure systems that could be changed. Davis expressed concern over the disconnect between technical advancements in AI and their practical utility for researchers, suggested a looming crisis in academic funding and hiring practices, and criticized the lack of focus on meaningful scientific inquiry amid a competitive race towards

__________________

larger models. As Davis noted earlier in the workshop, “it is not the responsibility of private investors or privately funded organizations to ensure that public goods are available.” He continued that, in his opinion, some private institutions might sufficiently drift away from their contract with society and deprioritize communicating what they are doing clearly to the people expected to use their products—such as drugs or AI models. This could result in “a deterioration of the fundamentals because we get a de-prioritization by regulators, by government.”

Chin said that fundamental restructuring may be useful to replace the current incentive systems, such as academic incentives for publication or the incentive for patients to engage in data sharing or research, with a better system that keeps in mind the lessons learned from the previous systems.

At this point, the discussion shifted to a new topic focused on immunity. In response to a question about the use of the immune systems in fighting cancer, Chin said a major hurdle is researchers have not been able to assemble a full picture of how the immune system actually operates in the real-world human being. Instead, much of the research has been done on model systems at different stages of cancer. A key feature missing is data tracking how the immune system evolves in people who have cancer and are being treated in a certain way.

From the drug development perspective, it will be important to assemble the bigger picture of available pharmaceuticals to identify combinations of drugs, she said, because there will be no magic bullet. “Multiple processes are at play, and the relative contributions of these hallmarking cancers are not identical in different patients at different times,” she said. Thus, to beat cancer researchers may use combinations of drugs and it will be important to know which drugs should be administered at which particular stages and in a given patient. Getting to that point likely involves collecting large amounts of data from a wide variety of patients.

Next, Alexandra Snyder asked what would be needed to make clinical trials more representative of the populations that will subsequently be treated. Pohl suggested that cheaper diagnostics could be part of the answer, if they would allow people to take the tests at home. Koyejo added it is important to increase trust of the scientific and medical communities among marginalized populations, and this in turn may rely, in part, on how well the purposes and benefits of research can be communicated to these populations.

Audience member Judith Burstyn, the director of the chemistry division at the National Science Foundation, asked for suggestions on how funding agencies could better support transformative science. Chin said that the current funding structure actively disincentivizes high-risk research because any failures are viewed negatively. There are obstacles in peer review, she continued, because “people who end up serving on the study section often are not the people that are really the right peers.”

Koyejo added that a similar practice is at work in hiring at major universities. The schools tend to look for “safe choices” who had already proven themselves rather than taking chances on people who might be amazing but are early in their career with shorter track records. Something similar is going on in the AI community, he said, as the new models being built are variations of models that have already proved to be successful. One possible solution, he suggested, would be to revise reward structures so that people who take chances get rewarded.

FORESIGHT ACTIVITY: IMAGINING THE FUTURE

In the final session of the first day, the workshop speakers and attendees were divided into three breakout sessions to envision three potential scenarios for the future of AI-assisted drug discovery and development: a positive scenario, a negative scenario, and a middle-of-the-road scenario that was intended to illustrate very little progress from the present. Directions for this foresight activity were provided by Apurva Dave, the director of the Climate Security Roundtable at the National Academies, and Erran Carmel, a futurist at American University. The goal of the exercise, they explained, was to develop plausible scenarios for what life might be like in 2050, depending on choices made now and in the near future about AI-assisted drug discovery and development. Generative AI was used to draw up three possible futures and attendees were distributed between the same number focus groups to discuss these futures. Conversations in the groups were open-ended and unprompted.

Positive Scenario

Nicola Pohl and Ben Wright, a Washington, D.C. area researcher in data science and AI, described the posi-

tive future their group envisioned. Pohl said that in this future, 2050 would be much more centered on individuals, with a goal of each person being completely in charge of making decisions about his or her own health. In this scenario, the drug discovery paradigm would be shifted toward “healthspan”—the amount of healthy life—than on lifespan. The big uncertainty, she said, was whether human institutions could evolve fast enough to keep up with the changes in science and technology, as these institutions are important to achieve the future goals the group sketched out.

Wright added that the group discussed privacy concerns and the building of trust, as two important issues to enable the positive future. Part of that future, could be wearables that not only monitor individuals’ health but also deliver necessary treatments.

Attaining such a future, Pohl said, likely involves accelerating processes such as chemical sensing and diagnostics to provide hyper-individualized, on-demand delivery of therapies and treatments. Would it be possible, for instance, to produce chemicals on demand and carry out a basic safety profile of a new molecule on the timescale of just a few days?

Wright said this future could also include major improvements in data acquisition fortraining AI models and the necessary infrastructure, including energy, communications, and research.

Ultimately, Pohl noted, the key question is, “How do we organize ourselves as human beings to get to this future?” Several members of the group suggested that researchers, engineers, and the broader community could evolve to bring this positive future into reality.

Negative Scenario

The negative-future group presented their vision of 2050 via an imagined conversation between two people facing extreme economic struggles. The two discussants were Lara Campbell, former executive director of the President’s Council of Advisors on Science and Technology, and Vikram Venkatram an analyst from Georgetown University.

In their scenario, they envisioned a decline in overall health in the United States with lower life expectancy. Apart from those with economic security, the scenario included lack of access to health care for most people. The group discussed the ramifications of a future in which the best technology, AI, and health care could be found outside of the United States.

Venkatram explained that it will be important for the United States to take advantage of the molecular industrial revolution, which could be “the next huge technological advancement, the chance to take control of human health and improve it for everyone.” Several reasons were offered by group members. First, intellectual property issues and strict regulations could limit the amount of data that could be used by AI in discovering and developing cutting-edge medications. Failure to shift market incentives could mean that entrenched pharmaceutical companies retain their existing model of drug development, without innovative personalized therapies, and investors may not invest enough into AI-driven drug discovery. Other missteps that might lead to this “negative” scenario include not doing an effective job of building trust with communities, failing to develop a workforce with the necessary AI and manufacturing skills, and not constructing effective supply chains.

Middle-of-the-Road Scenario

Linda Nhon, former director of the Chemical Sciences Roundtable, listed examples of a middle-of-the-road future as envisioned by her group. In this scenario, inadequate funding and lack of incentives to embrace change can leave regulatory agencies, the pharmaceutical industry, and academia stagnant even if AI continues to improve exponentially. The group discussed the potential expenses associated with the future of publishing in peer-reviewed journals, open-access models and the publication business structure, and AI technologies that may have unintended consequences. These factors may contribute to the breakdown of the reward structure for a traditional tenured track professor, which could disincentivize new talent from entering academic research.

In the scenario where deliberate efforts have not been made to change the status quo, this stagnation could in turn paralyze the AI-assisted drug discovery, drug validation, drug evaluation process, and the benefits of AI as an enabler for productive research. In addition, access to human health data, with which to train AI models, can be limited. Cheaper and individualized medicines could be

made, but some group members noted that it is important to allow for broad access to these advancements.

Another topic discussed in the context of this scenario is the emergence of nutraceuticals. AI could be widely used to create supplements using organic and all-natural resources plus FDA-approved sources by individuals in their homes.

Roundtable member Andrew White provided some details on how this scenario was developed. The group envisioned a number of hypothetical policy decisions that could impede progress on AI-assisted drug discovery. It was noted a hypothetical policy change of only requiring a drug’s safety data instead of efficacy, could reduce the number of drugs covered by insurance. Expanding on the thought-exercise, increased amounts of regulatory burden and paperwork for clinical trials could also increase their cost, potentially pricing out academics and many federal agencies from funding clinical trials. The result could be a sharp decrease in the number of new drugs.

The group also discussed the importance of affordably training generative AI models, and the benefits and drawbacks of open access research. The potential for companies to market their health treatments directly to consumers was also explored.

Discussion

In the discussion period led by roundtable members Michael Janicke and Laurel Royer, the three breakout groups reviewed how their open-ended ideation ended up with the scenarios that they described. Pohl said her group, the positive future, assumed that 25 years would not be enough time to completely rebuild trust overall, so they assumed it would have to be community-based trust with relatively small communities. Different aspects of health care, such as the chemistry to create new, made-to-order drugs would have to become fast and cheap enough that they are essentially commodities. At that point the care would be affordable, she said, “and then it can be community-controlled and, therefore, slowly rebuild trust with overlapping communities where it’s clear that you have agency as an individual to be able to share your data for how this therapy worked in your particular case.”

Judith Burstyn, whose group considered the stasis scenario, said that they assumed the path to that future would start with disinvestment in the federal research enterprise, which “essentially pulled the rug out from under things like clinical trials and other research.”

James Li, who was in the positive-future group, offered two observations. First, if it is possible to get to that future, it could cost less than anticipated because there would be a healthy, more productive population and the money and effort spent on health care could be less. Second, the positive and negative futures are not as far apart as one might think. “Because the technology is more advanced, it’s easy to fall into the cracks and get into the bad scenario, and you can do that in a more accelerated fashion,” he said. So it is important to put some guardrails, such as decentralization and local community trust, in place to keep the negative future from happening.

Lynda Chin agreed with Li that the futures are not that far apart. “There is not an on and off switch,” she said. “It’s a matter of picking the right balance.” So seemingly small, isolated decisions may have broad-reaching impacts. Second, she said, the degree of support for the ecosystem—the workforce, the supply chain, and so on—will play a major role in determining which future arises.

Venkatram commented it was interesting that all three groups seemed to have assumed that the technology itself will work—the negative-future group predicted that at least the wealthy would have access to lifespan-extending drugs, and the middle-of-the-road group assumed there would be nutraceuticals. So, the difference in outcomes depended mainly on social, political, and economic issues.

Burstyn said that the middle-scenario group’s focus was on the siloing of information. “It wasn’t that there wasn’t progress in certain spaces, but that it wound up siloed and, therefore, not broadly accessible,” she said. So, any use of AI technologies ended up reinforcing the status quo.

The session ended with a brief discussion about how to build trust after Campbell noted that all the groups had touched on the importance of trust. Royer said that lack of positive communication about chemistry with the general public is a major gap which weakens trust. She noted that the onus is on chemists to figure out how to help the general public understand the contributions that

chemistry is making and to address negative misperceptions about chemistry.

Pohl agreed, and added, “I think we also need to get our own house in order as chemists, engineers and scientists.” When the public reads about irreproducible results in science, it damages their trust in science.

BUILDING THE COMMUNITY: SUCCESS STORIES

The workshop’s final panel, chaired by roundtable member Martin Burke, offered a look to the future, with a special emphasis on how AI might help open the search for new medicines to people beyond the traditional community of researchers in academia and pharmaceutical corporations. Today, he noted, there are 350 million people in the United States and 8 billion people in the entire world, yet the number of people who can go into a lab and create medicine could fit into a single building. Thus, “99.99999 percent of all the human imagination and creativity potential that exists on our planet, arguably our greatest natural resource, the thing that AI cannot yet do, is not being tapped.”

But what AI can do is to help open the way to a “world where anyone can make medicines,” he said. So, to explore methods to encourage people from outside typical scientific communities to become involved in a science- or math-related project, he explained, the workshop organizers invited people from outside chemistry and pharmaceuticals to talk about “other places and spaces where they have very successfully leveraged kind of uniquely human super-powers to accomplish extraordinary things.” Could these examples be a precedent for engaging citizen scientists in successfully enabling AI for drug discovery? This final session tapped three inspiring individuals to talk about success stories achieved by engaging citizen scientists in tackling challenging diseases such as cystic fibrosis, teaching millions of students computer coding, and involving amateur astronomers to explore outer space.

Emily’s Entourage

First, Emily Kramer-Golinkoff, the founder of the nonprofit foundation Emily’s Entourage, described how her foundation has advanced research on rare types of cystic fibrosis (CF) that had not received much attention from the mainstream medical research community. As someone living with a rare form of CF, she said, she was driven to form Emily’s Entourage in 2011 by the development of mutation-targeted therapies for CF. These therapies were proving to be game-changing for the 90 percent of people with CF who had the most common CF mutations. Because of the lack of financial incentives, there has been little work to address the rare CF mutations that has affected the remaining 10 percent of the CF population.

As someone with advanced CF, Kramer-Golinkoff said she realized that “time was of the essence and that we didn’t have time for drug development to unfold on the traditional timeline.” So, despite having no background in fundraising or in science, she worked with family and friends to launch Emily’s Entourage. Its mission, she said, is to bring life-saving advances to people in the 10 percent for whom available treatments are not effective and, most importantly, do it quickly because time matters. “Desperation can be an incredibly powerful and effective motivator and change-maker,” she said, “and I think my story is an example of that.”

Since its inception, the foundation has raised over $15 million and has awarded 38 research grants to investigators around the world, who in turn have secured more than $52 million in follow-on funding. Using a venture philanthropy model, the foundation spun off Spirovant Sciences, Inc., which develops gene therapies for rare forms of CF and is now carrying out phase 2 clinical trials of one therapy. It also built a patient database and a clinical trial matchmaking program. This expedites clinical trial recruitment of people with rare forms of CF, she said, “because it turns out clinical trial recruitment is one of the really big barriers for fast drug development for a rare disease.” Emily’s Entourage also carries out fundraising and awareness-raising activities, hosts conferences, and does regulatory and advocacy work.

Scratch Foundation

Margaret Honey, president and chief executive officer of the Scratch Foundation, described how her foundation has made coding accessible to millions of young people so that they can create programs that will do whatever they can imagine.

Two decades ago, Honey, a developmental psychologist interested in how people use new tools, learned about the Scratch program, created by Mitch Resnick at the

Massachusetts Institute of Technology (MIT) Media Lab. What makes Scratch so valuable, Honey explained, is that it offers an accessible approach to coding so that young people with little coding knowledge can use it as a means of creative expression, while also making it possible for people who know more about coding to get quite sophisticated results. Scratch was made available online in 2007, and by 2010 it had 1 million users, she said, and by the end of 2024 it had been used by nearly 115 million young people, most of them between the ages of 9 and 15.

Noting that Scratch uses no marketing and no advertising, Honey said that Scratch is popular because it is not overtly educational: It is not designed to teach things. Instead, it simply makes it possible for young people to use code to realize their ideas and things they are passionate about. Thus, the users see Scratch as “their space” where their creativity can bloom.

A second factor, she added, is that Scratch is accompanied by an online community where the users can learn and share ideas with one another, comment on one another’s projects, share code, and set up discussion forums on particular topics.

Other factors behind Scratch’s success, Honey said, include that it has always been open-source—so its code has been used in a wide variety of other projects—and that there is a growing emphasis on and interest in computer science in education.

As for the challenges that Scratch has faced, Honey pointed specifically to funding issues and the difficulties in scaling the technology platform as it became increasingly popular.

NASA Planet Hunters

The final panelist Rob Zellem, an astronomer from the NASA Goddard Space Flight Center, described a citizen science project called Exoplanet Watch, which has involved thousands of amateur astronomers in the search for exoplanets, planets in other solar systems.

Exoplanets are generally detected either by observing the slight wobble in a star’s path caused by the gravitational pull of a circling planet, or by observing the slight dimming of a star as a planet passes directly in front of it, blocking some of its light. Zellem said even a 6-inch telescope, something that many amateur astronomers own or have access to, is enough to observe such dimming and note the passage of an exoplanet in front of its star. There are many millions of stars in our solar system that might harbor exoplanets and examining them all in detail takes an incredible amount of observing time, as well as time spent analyzing the resulting data, so Zellem encouraged amateur astronomers to get involved in the search.

Today hundreds of amateurs with telescopes are working with Exoplanet Watch to find and track exoplanets, and thousands more without telescopes are working to analyze data supplied from a 6-inch telescope run by NASA. With the software provided by Exoplanet Watch, they can analyze the data on their own computers or even on smartphones. One valuable task they carry out is to determine the exact timing of exoplanets crossing in front of their stars. This is important to NASA because it makes it possible to know exactly when to train the James Webb telescope or other major telescope on the exoplanet to get data that can only be retrieved at the time of the crossing. The amateur astronomers have also discovered dozens of exoplanets on their own.

These citizen scientists will be even more valuable, Zellem said, when the new Nancy Grace Roman Space Telescope is launched, as it will generate data at a rate 1,000 times greater than the Hubble Space Telescope does, offering tremendous opportunities to amateur astronomers who will be given access to that data to help discover and track many more distant planets.

Discussion

Burke pointed out that each of the three panelists had identified the driving force that helped explain the success of their projects: the urgency in seeking treatment options for Kramer-Golinkoff’s CF foundation, the drive for creative agency for Honey’s Scratch Foundation, and the capacity for pattern recognition among the amateur astronomers looking for exoplanets that Zellem spoke about. So what collective strength, Burke asked, might be tapped into by citizen medicinal chemists searching for tomorrow’s medicines? Kramer-Golinkoff answered that what drives change is people who care and connect

with others and people who dream of better futures and are driven to create them. Honey pointed to the ability of people to be thoughtful, intentional, and passionate about the problems they are trying to solve. Zellem echoed Kramer-Golinkoff’s answer by saying that establishing a strong community whose members feel connected and help one another is vital to the success of such projects.

Burke next asked about lessons the panelists had learned from missteps or mistakes they had made. Kramer-Golinkoff said she had learned the importance of thinking ahead and figuring out the most important obstacles to success and addressing them early on. Honey singled out the difficulty of moving a technology from an academic environment to the real world and said it was important to have leaders who truly understand what makes the technology valuable. Kramer-Golinkoff added that mistakes will inevitably be made and what is important is to be able to learn fast and pivot quickly. Zellem spoke about the importance of making a technology that has been designed for specialists accessible to a broader group of users if those users are going to be able to make contributions.

In response to question from the audience about how AI can be integrated with the roles of the citizen scientists in looking for new medicines, Honey said that one possibility would be to use AI as a creative learning assistant, augmenting people’s creativity but not replacing it. Zellem said that Exoplanet Watch has demonstrated the importance of keeping humans involved in the search for exoplanets to do things that AI cannot.

Vikram Venkatram asked how one should go about building a community of people who will be involved in a task such as searching for new medicines. Kramer-Golinkoff emphasized the importance of building trust as well as the value of working with advocacy organizations and showing investment in the community. Zellem said that in the case of Exoplanet Watch, much of the community building came about through word of mouth, giving talks, and having an annual conference that was advertised with a press release. Honey agreed that word of mouth is important, especially among young people.

In response to a question about fund-raising, Kramer-Golinkoff said that the first $100,000 was the hardest, but that raising funds “always requires being a hustler and an ambassador.” Zellem said that projects like these generally start off small and then organically grow as needs change and increase. Honey emphasized the importance of thinking through a long-range sustainability strategy.

Roundtable member Judith Burstyn asked the panelists how their experiences could be applied in the chemistry and pharmaceutical world. Zellem said it would be valuable to find a chemistry equivalent of the 6-inch telescope in astronomy—a tool that amateurs can operate that can produce useful data and even discoveries that professional astronomers can build on. Kramer-Golinkoff said it is important to remove barriers to entry and to make sure that the right problems are being targeted so that amateurs can get involved and not waste their time. Honey echoed that, saying, “If you think about the design of tools in such a way that they create an accessible pathway to entry, then amazing things can happen.”

DISCLAIMER This Proceedings of a Workshop—in Brief has been prepared by Robert Pool as a factual summary of what occurred at the workshop. The statements made are those of the rapporteur or individual workshop participants and do not necessarily represent the views of all workshop participants; the planning committee; or the National Academies of Sciences, Engineering, and Medicine.

COMMITTEE Martin Burke (Co-Chair), University of Illinois Urbana-Champaign; Nicola Pohl (Co-Chair), University of Indiana Bloomington; Milad Abolhasani, North Carolina State University; Lynda Chin, Apricity Health; Carlos Gomez, National Institute of Standards and Technology; Sanmi Koyejo, Stanford University; Laurel Royer, Carinalis Consulting and Research. The National Academies’ planning committees are solely responsible for organizing the workshop, identifying topics, and choosing speakers. Responsibility for the final content rests entirely with the rapporteur and the National Academies.

REVIEWERS To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by Lara Campbell, Petrichor Consulting, LLC; Savanah Shumaker, United States Naval Academy, and Jonathan Wylde, Solugen. We also thank staff member Sheena Posey Norris for reading and providing helpful comments on the manuscript. Lauren M. Everett served as the review coordinator.

SPONSORS This workshop was supported by the National Science Foundation and the Department of Energy Office Basic Energy Sciences Program. Any opinions, findings, conclusions, or recommendations expressed in this publication do not necessarily reflect the views of any organization or agency that provided support for the project.

STAFF Michael Janicke, Senior Program Officer; Kayanna Wymbs, Research Assistant; Darlene Gros, Senior Program Assistant; Thanh Nguyen, Senior Financial Business Partner; and Charles Ferguson, Senior Director, Board on Chemical Sciences and Technology.

SUGGESTED CITATION National Academies of Sciences, Engineering, and Medicine. 2025. Building Access to Tomorrow’s Medicines: Bringing Together Humans, Robots, and Artificial Intelligence: Proceedings of a Workshop—In Brief. Washington, DC: National Academies Press. https://doi.org/10.17226/29162.

For additional information regarding the workshop, visit https://www.nationalacademies.org/event/44130_02-2025_building-access-to-tomorrows-medicines-bringing-together-humans-robots-and-artificial-intelligence-a-workshop.

Division on Earth and Life Studies Copyright 2025 by the National Academy of Sciences. All rights reserved.