2

Fundamentals of AI in Scientific Research

WHAT CAN AND SHOULD AN AI SCIENTIST DO?

The workshop’s first session was devoted to the question of what an artificial intelligence (AI) scientist would look like—what such a scientist could and should do. In her introduction to the session, moderator Yolanda Gil (University of Southern California) offered a brief history of how scientists have thought about AI and scientific discovery. In 1973, for instance, Herb Simon published the paper “Does Scientific Discovery Have a Logic?” in which he argued that there was such a logic. Simon worked for many years to articulate the structure of this logic. Other scientists took up the same issue from various vantage points, with the sense that if the logic of scientific discovery could be set forth, it might be possible to automate it.

AI is already affecting science, Gil said, and she offered some examples of how that is the case. Many AI technologies can be applied to different aspects of science, such as AI-guided sensors, natural language techniques to process scientific literature, and machine learning systems that can point to explanatory and causal relationships.

Finally, she offered some characteristics that future AI scientists should have if they are to partner with human scientists. These included rationality, initiative, networking, articulation, and ethics. Developing AI scientists with such characteristics will be challenging, she said, but potentially very rewarding.

Nobel Turing Challenge

Hiroaki Kitano (Sony AI, Inc.) spoke first. The Nobel Turing Challenge, he said, is to develop AI systems by 2050 that can make major scientific discoveries that are worth the Nobel Prize and beyond (Kitano, 2016, 2021). That is, the systems would carry out autonomous scientific discovery to create ideas by analyzing data, generating hypotheses, designing experiments to verify or falsify them, and drawing conclusions about their implications. A major question related to this challenge is whether such an AI scientist, capable of doing research at the highest level, would behave similarly to human scientists or behave differently, with a different form of intelligence.

If an AI scientist behaved similarly to a human scientist, then one approach to creating such a scientist would be to map out the approaches that human scientists have used successfully in the past and try to reproduce and augment those in an AI scientist. An important question concerning that approach is whether it is possible to formalize scientific discovery and then automate it to create scientific discovery on a large scale. Today, Kitano said, scientific discovery can be thought of as being at the same level as manufacturing before the Industrial Revolution—it is carried out by individuals and limited by how quickly those individuals can work. Automating it would allow scientific discovery to be done much faster and more cheaply and might cover more topics than is possible by a group of human scientists.

In moving toward such a goal, he said, the first stage will be to automate labs as much as possible, creating connected research labs that serve as AI assistants for scientific discovery. From there, the AI science assistants will need to be given the capability to operate autonomously, making discoveries themselves. He mentioned certain steps along the path toward this goal, such as Adam–Eve, the first closed-loop laboratory robotic system (Sparkes et al., 2010). The strategy is to automate the lab and give autonomy (at appropriate levels) to AI scientists. The ultimate goal, he added, is to automate scientific discovery at scale, and achieving that will require success in several areas, most notably vision and leadership, technology platforms, and project strategy/management.

AI for Accelerating Scientific Discovery for a Sustainable Future

Carla Gomes (Institute for Computational Sustainability, Cornell University) spoke about advancing AI and computational methods that can be used to address sustainability challenges. She defined “sustainable

development” as development that meets the needs of the present without compromising future generations, with “needs” referring to anything that affects human well-being.

She began by offering examples of computational sustainability research challenges, such as using AI to accelerate materials discovery for solar fuels and other forms of renewable energy or using AI to combat the rapid decline in biodiversity.

Speaking in another direction, she commented that there has been tremendous progress in AI over the past decade, but it has mainly been in data-driven AI such as image and voice recognition and ChatGPT. But pure data-driven AI is not suitable for scientific discovery, since there are several requirements for scientific discovery that data-driven AI cannot satisfy. These include understanding and explaining phenomena, beyond mere prediction; identifying causation rather than just correlation; dealing with small and incomplete datasets; and making choices.

Ultimately, Gomes said, what will be required is combining data-driven AI with scientific reasoning. This sort of data- and knowledge-driven AI will be able to reason from first principles, factoring in prior knowledge; reason about uncertainty; and carry out symbolic and combinatorial reasoning such as in dealing with combinatorial search spaces.

As an example, Gomes spoke about how a deep reasoning network, which integrated prior knowledge and logical and constraint reasoning into deep learning, was able to solve a problem involving crystal structures that human researchers had not been able to solve. She also described her system, the Scientific Autonomous Reasoning Agent, or SARA, for materials discovery, which controls the annealing synthesis and X-ray diffraction analysis of various materials, deciding what new materials to synthesize according to what has been learned about the previous materials.

In concluding, Gomes emphasized that it will be important to design more ethical AI systems that consider various objectives instead of a single objective, as in working toward sustainable development, which requires balancing environmental, economic, and societal goals.

How Can AI Contribute to Scientific Understanding?

Mario Krenn (Max Planck Institute for the Science of Light) began his talk by pointing to a paper he and colleagues had published a year earlier (Krenn et al., 2022), “On Scientific Understanding with Artificial Intelligence.” To gain insight into how AI can contribute to scientific understand-

ing, his group reviewed philosophies of science and interviewed more than 50 computational physicists, chemists, and biologists.

The researchers structured computer-assisted scientific understanding into three classes. The first was “computational microscopes,” or using AI to carry out computational experiments that could not be done in the real world and thus gaining unique insights into a phenomenon. The second was AI as a source of inspiration, providing fresh ideas that can lead to new understandings. The third was AI as an agent of understanding, with AI gaining new scientific understanding autonomously and passing it along to human researchers. Their survey turned up many examples of the first two classes but not a single example that could be classified as AI acting as an agent of understanding. For the remainder of his presentation, he focused on examples in the second class.

He began by offering an example from his work in computational quantum physics. To explore a specific type of quantum correlation that had never been observed, he tried to figure out an experimental setup that would work, but nothing succeeded; in frustration, he turned to a computer algorithm, which found an answer within a few hours.

Ultimately, his team was able to build the setup and run the experiment successfully. In another case, AI developed a way of overcoming a limitation that no one had thought of by using a technique from a much earlier paper in a new way. The results of that theoretical study were published in Physical Review Letters, and as Krenn noted, the idea that was described in the paper was not generated by any member of the team but by an algorithm.

Krenn said that AI can serve as an artificial muse—that is, a source of original ideas—in various ways. AI can detect anomalies or find patterns in large collections of data, and the way it ends up solving problems can be studied for insights into new ways of thinking about a topic.

Krenn concluded that if AI is to be effective in scientific discovery, it will need to have the characteristics of successful human scientists, such as creativity, curiosity, and understanding. But to get to the point where AI has these qualities, it will be necessary first to gain a clear understanding of what creativity, curiosity, and understanding mean in a human context.

Discussion

There was much discussion in response to a follow-up question from Patrick Riley (Relay Therapeutics), who asked about Krenn’s comment that he found no instances where AI served as a source of understanding and

what the others thought about that. The panel spoke about cases where AI and machine learning provided information that helped scientists improve their understanding of a situation—and thus, in a way, served as a “source of understanding”—but the AI itself did not develop an understanding of the situation.

In response to a question from a remote participant, Gomes said that there is great potential in AI being used to create and work with digital twins in studying such things as the human body. In response to a question from I. Glenn Cohen (Harvard Law School), Gil said that AI could play a role in helping scientists do their jobs better—report their findings better, put them in context better, verify them better, and so on. In response to another question from an audience member, Krenn said it would be very important to reach a point where an AI model has developed an understanding of a phenomenon and can explain it to human scientists. That explanation could come in various forms, from natural language to illustrations or equations, but it is the communication that will be important.

EXAMPLES OF HOW AI HAS ALREADY ASSISTED IN SCIENTIFIC DISCOVERY

The second panel was moderated by Keith Brown (Boston University). This panel was devoted to offering examples of how AI has made contributions in various disciplines, such as materials research, chemistry, climate science, biology, and cosmology, to identify where AI is currently in the process of scientific discovery. The panel was also tasked with identifying technologies that the community requires to advance the use of AI in research.

Use of Machine Learning in Biomedical Imaging

The first presenter, Steve Finkbeiner (Gladstone Institutes), described how he has used machine learning to gain insights from biomedical images. He is interested in neurodegenerative diseases such as Huntington’s disease and Alzheimer’s disease, which are exceptionally complex and difficult to study.

Initially, he was using robotic microscopy to carry out high-throughput longitudinal single-cell analysis to observe how cells changed over time and spot pathologies in those cells. He employed an array of more than 270 biosensors to visualize the biology of the cells and quantify disease pheno-

types. Originally, the images were interpreted by human scientists, but with deep learning (an AI technique), he was able to discover information and reveal insights that are more difficult for humans to detect. In one study, an algorithm learned to identify with perfect accuracy which cells were going to die, when the best that humans had been able to do was only 70–80 percent accuracy. The algorithm also made it possible to identify which cells were going to die days before death—an improvement that could make it possible to intervene effectively in the neurodegeneration process.

Finkbeiner then offered several other ways that AI might be harnessed in studying neurodegenerative diseases. It could, for example, help answer the question of whether diseases such as Alzheimer’s disease and Parkinson’s disease are one disease or many. Deep learning could potentially be used to find novel pathologies. In time, he suggested, closed-loop machine learning might be applied to rapidly create predictive models of complex biological spaces and to discover biological paths to desired destinations, such as how to transform a sick cell into a healthy one.

AI for Protein Design

Proteins can be used in a wide variety of applications, according to Lynda Stuart (Institute for Protein Design, University of Washington). Some examples of uses for proteins include vaccines and drug delivery systems, cancer immunotherapy, bio-based computers, nanoscale manufacturing, protein-silicon devices, carbon sequestration, and degradation of plastics. The number of proteins found in nature represents only a miniscule percentage of all possible proteins, and researchers are interested in learning how to design novel proteins that will have desired properties.

AI is already playing a major role in this process, Stuart said. As an example, she described RoseTTAFold, a neural network (a particular type of AI) that has been trained to predict the structure of a protein from the sequence of amino acids that makes up that protein (Baek et al., 2021). A second technique, RFdiffusion, generates new protein structures that satisfy particular design criteria via a process called “progressive denoising,” which was inspired by deep learning methods for generating synthetic images (Watson et al., 2023). By combining these techniques with ProteinMPNN, a technique for assigning amino acids to protein structures, it is possible to design proteins from scratch that have particular properties, such as binding to specific molecules. This AI-enabled process has many applications in medicine, such as building antibodies to viruses.

Stuart identified two areas as “next frontiers” in the use of AI for protein design. The first was developing standardized datasets for the training and optimization of AI models. By using semiautomated protein production and yeast surface display, it is possible to produce and characterize 192 proteins a week. These new proteins can then be used in datasets. The second frontier she identified was rapid iteration for active learning. By going through several iterations of designing a pool of proteins, using an AI model to choose the most promising, and testing the designs, the models can be sharpened quickly. The major barrier to this, she said, is that for it to be practical, the cost of gene synthesis will need to be cut by a factor of 10.

Using Foundation Models for Science

Shirley Ho (Simons Foundation/Flatiron Institute and New York University) spoke about how foundation models can be used to improve the performance of AI in scientific discovery. Foundation models, she explained, are large models that are pretrained with task-agnostic objectives on massive, diverse datasets. These models extract features that can be used as bases for task-specific fine-tuning, leading to performances that are better than can be achieved with supervised training on many types of problems.

In particular, Ho is involved with the development of Polymathic AI, a tool with similarities to ChatGPT but that is aimed specifically at developing models for use in various areas of science. Part of the reason that tools like ChatGPT perform so well on specific tasks such as language translation is that they have already inferred shared concepts such as grammar and causality. Since there are many shared concepts across scientific disciplines, it makes sense to develop models with such shared concepts. As an analogy, she said that it is easier to teach physics to a biologist than to a toddler because the biologist already has an understanding of how the world works and can transfer knowledge to new areas. But in most cases where machine learning is used in science, she said, people start with an untrained network—that is, the toddler. “Our initiative aims to help change this,” she remarked.

One potential use of this approach, Ho said, is in tackling scientific questions in which there are limited data to work with. By pretraining models on large quantities of simpler, more easily simulatable systems that only partially capture the target physics, they have found that it is often possible to get better performance in analyzing the systems of interest that have limited data.

The goal of the Polymathic AI initiative, she concluded, is to usher in a new class of machine learning for scientific data, building models that can use shared concepts across disciplines by developing such foundation models, releasing them for use by researchers worldwide, and training those researchers in their use.

Using AI for Scientific Discovery for Weather and Climate

In the panel’s last presentation, Amy McGovern (University of Oklahoma) talked about how AI is being used for scientific discovery in the areas of weather and climate. In particular, she described work applying AI to increase understanding of the evolution of tropical cyclones (e.g., hurricanes and typhoons) over their lifetimes, including their structure and intensity. The ultimate goal, she said, is to improve the forecasting of hurricanes and other tropical cyclones.

At present, McGovern continued, the scientific understanding of physical tropical cyclone processes such as convective structure, intensification, and rainband structure is limited by the capabilities of current satellite instruments. In particular, visual imaging cannot see into the interior parts of a storm, while microwave imaging—which can see inside a storm—is available from satellites only intermittently. To address this issue, McGovern’s group used AI to create an “AI sensor” that uses visual images of tropical cyclones to generate synthetic images that reveal the storms’ internal structures. These synthetic images—which have been shown to closely mirror what the interiors of the storms look like—can then be used to learn how the internal structure of a tropical cyclone evolves over time.

This work is already being used to produce real-time predictions for the behavior of tropical cyclones, with an updated prediction being generated every 10 minutes. McGovern listed many science questions related to the AI sensor. Can they be used, for instance, to identify the rain bands in tropical cyclones, which are key to understanding the associated rain and flooding? Can they better identify secondary eyewalls and main eyewalls that are obscured by higher clouds? Can the increased temporal resolution that AI offers be used to study processes involved in convective rain bands and eyewall replacement cycles more closely?

In closing, McGovern said that AI is already facilitating novel scientific discovery in meteorology and has the potential to transform the foundational understanding of many high-impact phenomena in the field.