An Interview with Dr. Joshua Swamidass

Introduction

I recently had the opportunity to interview Dr. Joshua Swamidass from the Washington University in St. Louis. Dr. Swamidass has published over 150 articles, with research on applications of machine learning to problems in chemical biology and medicine. He is the author of The Genealogical Adam and Eve and the founder of Peaceful Science, where he writes on the advancement of the civic practice of science. In 2022, Swamidass became a fellow of the American Academy for the Advancement of Science (AAAS).

Over the course of our hour-long interview, we discussed several interesting and prescient issues. These include the intersections of faith and science (with a focus on origins), the secular concepts of emergence as seen in AI as well as in chemical systems, the innovations of AlphaFold, and how AI is employed in various healthcare settings.

On Faith, Science, and Origins

Marcus S.: Could you tell me about your faith background?

Joshua S.: I was born and raised in California to Indian immigrants who were Christian, and young-earth creationists. I grew up in that context, but I was really drawn to science. For a long time, I tried to work out how to fit faith and science together. I first wanted to be a medical doctor, but then when I did my first rotations in a research lab in college, I decided that I wanted to be a scientist. Then I learned about computational biology, back when it was a new field. I thought I had wanted to be a doctor, but I did not know how to fit these interests together. I found out that I could do a PhD in machine learning and AI while getting a medical degree through a combined degree program. In late 1998 it all clicked, and I ended up going to graduate school for an MD and a PhD, which took nine years. My career is not one I imagined was possible when I was entering college.

Marcus S.: How did your parents respond to your pursuits in a scientific field? Were they concerned about your faith? How did your faith change as you went through that scientific path?

Joshua S.: My parents were really pleased that I was going to school and working hard at it. When I started being more public about my disagreement with young-earth creationism, they were more concerned about my faith and wanted to know more. But they took the time to understand my position and understood that I still held to the core doctrines of my faith. The details about how God created the earth were not nearly as important to them. For other family members this is still an ongoing issue. They do not like that I have no problem with evolution.

Marcus S.: You wrote an opinion piece in the Wall Street Journal responding to a comment from Senator Marco Rubio regarding the age of the earth ¹, how was this received?

Joshua S.: I was a young faculty member when I wrote that piece, and I remember that there was a lot of risk in writing it. I was outing myself to the Christian community in thinking that we should be okay with evolution. I was also outing myself to the secular scientific community by sharing that I was a Christian who believes Jesus rose from the dead. I remember receiving a lot of mail after publishing that piece in 2012.

On Emergent Phenomena in AI and Chemistry

Marcus S.: One popular belief in secular science and philosophy is the idea of emergence; that is, simple phenomena leading to these complex behaviors. In molecular dynamics we can use Brownian motion to accomplish maneuvering very complex molecules with second- and third-order structures. In machine learning, emergence has also taken hold through diffusion models (for example in recent publications by Jing et. al. ²). There are now two fields that scientists can point to where emergence has led to innovations. But we believe in an intelligent creator, who drives these stochastic processes. How do you rectify the practical applications of emergence with your belief in an intelligent creator?

Joshua S.: In a general sense, most religions hold to the basic idea is that a God or gods created everything. It is implied that they created everything out of smaller parts that work in simpler ways to create emergent things. Emergence can take different meanings, but in general it means that structures can emerge which do not easily follow from an understanding the smaller parts. It becomes easier to build a logical framework at a higher level. That is not a particular problem for religion, it just means that God made the world in a very rational way. It works very well with monotheism. It is a bit of a challenge with polytheism, but I am sure that there are ways to work around that.

This a bigger problem for those taking a science-only approach to the world. To illustrate this, we can think of chemistry as an emergent field. The most basic level is quantum theory, built on Schrodinger’s wave equation. Even if you know the wave equation, it does not mean that you can infer much about larger chemical systems. That is why you learn chemical intuition apart from the wave equation, even though the wave equation technically includes all information. Chemistry is a higher level of abstraction that is completely contained within the wave equation. Knowing the wave equation is not immediately sufficient, and that is what emergence implies. Physicists sometimes argue that chemistry is nothing but Schrodinger’s equation. I think emergence undermines that sort of view, because at a higher level, there is a vast area of legitimate knowledge that operates in a distinct and essentially orthogonal way. Emergence is just how things work, and there will always be a higher level of abstraction where real insights can be found.

AlphaFold and AI for Drug Research Innovation

Marcus S.: AlphaFold has proven to be an exceptionally powerful tool. Users input a string of amino acids and an AI will determine likely structures and geometries of peptides and proteins ³. Has your research found ways to augment the insights gained from AlphaFold?

Joshua S.: AlphaFold is an algorithm from Google DeepMind that can predict protein structure from a starting amino acid sequence. Part of why these algorithms work so well is because they are effective at operating in in statistically uncertain domains which lack the clean lines and sharp distinctions found in other areas. Within this framework of probabilistic reasoning the AI tries to make inferences based on context. Proteins have been resistant in a lot of ways to past prediction algorithms, partly because they do not merely operate like machines. They are more in a fuzzy domain, and one where newer algorithms are working quite well. One major innovation of AlphaFold is its use of a transformer architecture. Over the last five years, transformers have been a big advancement in ML and has been applied in a lot of related subfields. Many people are familiar with ChatGPT, which uses several layers of transformers. Many people are also familiar with stable diffusion, and these models use transformers internally too. Transformers allow ML algorithms to tackle harder tasks in a common framework.

When it comes to introducing people to the challenges that advanced artificial intelligence and deep learning pose, a good framework separates present and future issues. There is the AI of here and now, which is raising many ethical questions, such as how to apply it in ways that are fair. That is an area where many have contributed, and where our research group works as well. The second side considers the larger questions of the future of AI. This latter area is more philosophical, and a lot of theological questions come up. With ChatGPT we have a here-and-now set of ethical questions. Nevertheless, these present issues give us insight and experience when considering deeper questions such as whether AI could really have a mind.

Marcus S.: There has been a lot of excitement surrounding AI used for drug discovery. Do you think that an AI will create a drug that can pass FDA approval in our lifetimes?

Joshua S.: AI has certainly already contributed in very important ways. Drug companies were early adopters of AI, especially in cheminformatics. AI algorithms are already being used as part of the process. Some take this idea farther, such that AI will replace the process. I think that is highly unlikely, especially considering the amount of money involved and the degree to which a drug must be clearly understood by humans to pass regulatory approval. Drug development requires making many complicated decisions that have nothing to do with science. I am doubtful that an AI could really supplant this. It is the case that there are many subproblems where AI really produces a large improvement and dramatically helps with the drug discovery process. AI already powers tools that researchers use to make the drug development process safer, faster, more efficient, and less expensive.

AI in Healthcare

Marcus S.: You and your group work on applying AI in the healthcare space. What do you see as some of the biggest hindrances to progress in that area?

Joshua S.: The biggest hindrances lie in translating between AI experts and medical professionals. There are many powerful AI tools out there, but they are often only solving toy problems which are divorced from real clinical decisions or do not properly weight patient risks. In principle we have demonstrated that these algorithms can be helpful, but going after the questions that improve patient care is more difficult. Most AI researchers do not know what those questions are, and how to engineer solutions so they will be safe and understandable to physicians. We are just scratching the surface right now.

Marcus S.: When IBM Watson defeated Ken Jennings, the public was excited about Watson’s potential to improve healthcare. It has been several years since that IBM shut down the program, citing difficulties in acquiring data. How have AI systems moved past some of the issues faced by Watson?

Joshua S.: Watson was a rule-based AI system, and by the time it attracted lots of publicity, it was already shutting down. It had been incubated at IBM for quite some time and leveraged much older technologies that were dominant in the 1980s and 1990s. Even thought it had a lot of marketing, most people in medicine were not impressed at any point. Similarly, most AI researchers were not impressed at any point. I would contrast IBM Watson sharply with ChatGPT, image analysis, and deep learning developments happening right now in medicine. These developments are not owned by a single company either. Are they being over-hyped at times? Absolutely. Will there be disappointment at times? Absolutely. But because there is solid technology behind it, and because it is so decentralized, we can be certain that AI will have a large and growing impact in medicine and many other places.

Marcus S.: How do you think uncertainty quantification and explainability will impact how AI is applied in a healthcare setting?

Joshua S.: Applying these techniques will provide a big advantage. People talk about how AI is not explainable, but this is not entirely true. There are several different ways to make an AI that is explainable. One way is to break a big problem into a set of subproblems that are smaller and more distinct, each of which are solvable by an AI, and that can be recomposed either by an AI or by simpler statistical approaches. In this way you can construct an explainable system. This is especially important in medicine and is why we are often integrating information from disparate sources. We need to know why the AI is predicting something, so that we can effectively integrate it into a larger picture. Many newer ML methods return outputs that are not merely binary but are instead probabilistic claims that can be calibrated. Physicians are already well trained on integrating very complex data sources, and AI provides another indicator that can be incorporated with data outside the purview of the AI. This illustrates why Watson did not work well, because it could not account for uncertainty, and therefore was not the right tool for the job. But current methods can integrate probabilistic uncertainties, which ends up being incredibly valuable within a medical context.

Marcus S.: There is an incredible amount of venture capital for AI in the healthcare space. While there have been many successful healthcare companies built using venture capital, companies like Theranos have taken advantage. As you see start-ups combining AI innovations with medicine, do you have any rules of thumb to judge these entities, what they are claiming to do, and how they might be received by the medical community?

Joshua S.: Well, Theranos was not an AI company, it was a diagnostic company. Their entire premise was based on technology that was not made available to the wider scientific community for scrutiny. It came down to whether you took their word for it. Theranos fell apart because they were not honest about what the technology could do, which can give us insights on how to think about other companies. In general, having academics and physicians collaborating is a positive thing. For physicians that I know that are involved in start-up ventures, the goal is not merely to make money. Their motivations center around helping patients. In all the work that we are doing, we want the inner workings published so that it can be scrutinized by our peers. Even if there are parts that we keep proprietary, we are not hiding anything. Also, most of the technology that in this field right now is not actually proprietary. The deep learning software is open source, and anyone can use it. The challenging piece is finding the right problems to go after, collecting the right data, and building something that will work for that specific set of problems. This is what will allow these companies to sink or swim: their ability to apply an open technology to creatively solve focused questions that have an impact.

Future Work

Marcus S.: Can you tell us about what you are working on right now?

Joshua S.: You can follow more about me at peacefulscience.org, where we primarily engage with the public on questions of origin, but increasingly cover questions about AI as well. I am invited to speak in a variety of contexts, including both secular and Christian colleges, about the interaction of AI and faith. If that is something you are interested in, please contact me.

Marcus S.: What are you most excited about right now from a technical perspective?

Joshua S.: I am excited about recent advances in imaging analysis. These advances have not been fully explored when it comes to applications in medical imaging. We are doing work in this area which, I think, could lead to some big improvements.

Acknowledgments

A big thanks to Dr. Joshua Swamidass for taking the time to speak with me. Thanks also to Emily Wenger for checking the interview questions and proofreading this document.

References

Michael Jones. “The Origins of Young Earth Creationism.” Peaceful Science Prints (2022).

S. Joshua Swamidass. “Rubio and the Age-of-Earth Question.” Wall Street Journal Opinion (2012).

Jing, Bowen, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola. “Torsional diffusion for molecular conformer generation.” arXiv preprint arXiv:2206.01729 (2022).

Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A. and Bridgland, A., 2021. Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), pp.583-589.

High, Rob. “The era of cognitive systems: An inside look at IBM Watson and how it works.” IBM Corporation, Redbooks 1 (2012): 16.

Gaut, Joseph, Jon Marsh, S. Joshua Swamidass, Rick Berry, Victoria Swamidass, and Mike Blackford. “Deployment of a Deep Learning Model to Assist Pathologists with Donor Kidney Biopsy Evaluation.” Journal of Pathology Informatics 13 (2022): 100087.

1. Young-earth creationism is a Christian fundamentalist belief in a literal seven-day creation as well as genealogical reckoning placing the earth’s age at roughly 6,000 years old. This contrasts with the scientific consensus that the earth is roughly 4.6 billion years old ⁴. ↑
2. Molecular dynamics is a method of simulating the motion of atomic and molecular bodies over a fixed period and borrows principles from both classical and quantum physics. In the context of molecular dynamics, Brownian motion is a way of mathematically describing the random motion of suspended particles. ↑
3. Schrodinger’s equation is part of the bedrock of quantum mechanics. It is the quantum equivalent of Newton’s second law of motion. ↑
4. IBM Watson is a question answering system which competed on the popular show Jeopardy! in 2011, where it won first prize ⁵. ↑
5. Theranos was founded in 2003 by Elizabeth Holmes and raised over $700 million from private investors. The company claimed it could rapidly perform many diagnostic tests using only a small amount of blood. In 2015 it was discovered that these claims were false, and the company was dissolved in 2018. In 2022, Holmes and former company president Sunny Balwani were found guilty on charges of wire fraud and conspiracy. ↑
6. Last year, Dr. Swamidass and colleagues published work on a deep learning model to assist biologists with evaluating imaging from kidney biopsies ⁶. We look forward to seeing future work from Swamidass and colleagues in this and other areas of study. ↑