We have received the following perspectives about recent AI developments in science, including Large Language Models (LLMs) and specialized tools like AlphaFold, from computational biologists and bioengineers.
Andrew McCulloch is Director of the Institute of Engineering in Medicine and Distinguished Professor, Bioengineering, at UC San Diego. His lab uses multi-scale engineering approaches, including experimental and computational models, to help understand, diagnose and treat heart diseases.
More transparency for AI models can help to avoid bias and to improve the ability to judge reliability of inferences and of generated data. LLMs have a valuable role in basic science; the single most important repository of scientific knowledge exists in the form of published papers, and the rate at which they are being published is increasing exponentially. Our ability to keep up with them is decreasing, and tools that allow us to read and search the scientific literature more quickly are going to make a difference. Clearly LLMs have the ability to parse and interpret more complex prompts that the searches we have been used to. One of the biggest promises of AI comes from what it can do for computational models, at multiple scales, from atomic resolution to population scales. I can give a few examples of how AI can assist computational modeling to advance life science.
There is a class of models that represents biology at the critical mesoscale of intercellular and cell-matrix interactions in tissues, the scale at which much of pathology and physiology takes place. Different cell types in different tissues occupy different states, communicate with each other, and regulate the state and phenotype of the tissue. One class of agent-based models treats each cell as an agent that can talk to its neighbors and can change its state or position as a function of the behavior of its neighbors. These types of models lend themselves very naturally to the use of AI, because they are Markov chains, and they are well suited to machine learning. One of the limiting factors in making good agent-based models is that we have to extract rules from the literature about how different cells respond to their environments. In principle, we might be able to model every molecular mechanism by which cells respond to different stimuli from first principles, but in practice we need empirical information too. To achieve this, we read papers, and we translate them into quantifiable logical or mathematical statements. Now we can start to use existing agent-based models to train LLMs to discover these rules from very large swaths of the literature, for many more cell types and conditions.
At the population scale, the most compelling examples of AI applications are built on large datasets where we have data on many individuals. But when we look at particular pathologies, often we do not have so much data on the specific physiology or pathology in individuals with that particular disease. Here we get value from unsupervised machine learning strategies that we can use to generate virtual populations. We can learn a lot about heart disease by knowing the shape and structure of the heart. We have datasets of cardiac images, for example by MRI, and we use them to create statistical shape models. We use machine learning to automate the process of making these patient models, so we can model tens of thousands of patients fully automatically. We then use unsupervised machine learning strategies like principal component analysis or latent vector methods to create statistical atlases. We next discover the main components of shape variation across a population, and we use it to discover correlates with clinical outcomes and to synthesize new virtual populations, helping to provide mechanistic models of their electrophysiology or mechanics that are not as readily measured. For example, we make predictions from non-invasive measurements, like the ECG, of parameters that can only be measured using invasive methods like the pattern of electrical activation in the heart. These results have clinical applications, there is an FDA cleared product on the market based on this idea. It is helping cardiac electrophysiologists to perform ablation procedures. They can learn from the 12-lead ECG where to look for the source of the arrhythmia, even before the invasive mapping procedure in the EP lab. The procedures are faster, with less risk for the patient and a higher success rate.
In summary, we and others are using deep convolutional neural networks to process the images, LLMs to process the electronic health records, unsupervised principal component analysis to generate virtual populations. We can combine these with mechanistic computational models that incorporate prior scientific knowledge and physico-chemical constraints. The virtual patient populations are used to train new machine learning algorithms to precisely target therapies using less invasive measurements.
There is undoubtedly a good case for public investment in large scale AI development in health. LLMs have problems with bias and inaccuracy, difficulty in replication and validation. NIH has a program called Bridge to Artificial Intelligence (Bridge2AI) that is helping to address many of these challenges, and I think it is very valuable. AI can also help to improve the collaborative culture of life sciences. AI comes from computer science and computer science has a much stronger tradition of open access and open source sharing than medicine. The computational modeling community has also been at the forefront of developing standards for reproducibility and sharing.
Bjoern Peters is a Professor at the La Jolla Institute for Immunology, and he is an expert in computational tools to address fundamental questions in immunology. His lab helps to maintain the Immune Epitope Database.
In the field of epitope predictions we have machine learning tools developed from the 1970s and onwards, and the current ‘deep learning’ set of tools is an important but incremental advance. For the broader field of biology research, LLMs are proving groundbreaking by enabling access to data stored in free text such as journal articles. The need to convert free text into machine-readable presentations has previously been recognized with the advent of the genomic era, and led to the development of the Gene Ontology that was developed to ensure annotation and compatibility of data and databases. This concept has been generalized to provide a machine-readable representation of biological knowledge. An ontology is a system of carefully defined terminology, connected by logical relationships, and designed for both humans and computers to use (see for example obi-ontology.org). It supports the building of knowledge graphs. Data quality is very important for machine learning applications. We are considering the development of LLMs as user interfaces, front ends that can be queried by scientists using natural language, but where the underlying answers come from database queries or from other models. Answers from existing LLMs like ChatGPT are still quite superficial, they might be used by scientists in some cases as an introduction to topics with which they are not familiar. We also appreciate the usefulness of AlphaFold and we have included it in recent versions of the Immune Epitope Database. Another point to consider is that it would be helpful to see in the literature more validations of AI models, including comparisons with simpler methods.