Behind the Manuscript: Developing Algorithmic Psychiatry with Michael Halassa MD, PhD

Journal of Clinical Psychiatry: psychiatrist.com/jcp/

“Developing algorithmic psychiatry via multi-level spanning computational models”

Dr. Halassa’s Substack: michaelhalassa.substack.com

The Halassa Lab: https://halassalab.tufts.edu/

Ben Everett, PhD, is the Senior Scientific Director for The Journal of Clinical Psychiatry and Psychiatrist.com, where he oversees editorial strategy, content development, and multimedia education initiatives. He is the creator and host of The JCP Podcast, a series that brings together leading voices in psychiatry to explore the latest research and its clinical implications. Dr. Everett earned his PhD in Biochemistry with an emphasis in Neuroscience from the University of Tennessee Health Science Center. Over a two-decade career spanning academia, publishing, and the pharmaceutical industry, he has helped launch more than a dozen new treatments across psychiatry, neurology, and cardiometabolic medicine. His current work focuses on translating complex scientific advances into accessible, evidence-based insights that inform clinical practice and foster meaningful dialogue among mental health professionals.

This transcript has been auto-generated and may contain errors. Please refer to the original recording for full accuracy.

00:00 – Why Algorithmic Circuit Psychiatry Could Modernize Mental Health Care

Ben Everett: Hello and welcome to the JCP Podcast, where we explore the science and stories shaping mental health care today. I am your host, Ben Everett, Senior Scientific Director with Physicians Postgraduate Press, which is the publisher of the Journal of Clinical Psychiatry. On this podcast, we speak with clinicians, researchers, and thought leaders who are advancing the field of psychiatry. Our focus is not just on what is new, but on what is meaningful for our listeners and their actual clinical practice.

Today’s episode is part of our “Behind the Manuscript” series where we take a deeper look at a paper that was recently published. Today, we are going to talk to Dr. Michael Halassa about this paper that he had published in Cell Press with some collaborators, as well as some related writings on his Substack, to discuss his vision of algorithmic circuit psychiatry. We will also discuss why psychiatry feels so different from other fields of medicine and how a different framework for thinking about how we approach psychiatry might transform mental health care.

Dr. Halassa is a physician scientist with dual appointments. He is a Professor and the Director of Translational Research in the Department of Neuroscience at Tufts University, as well as a Professor of Psychiatry in the Department of Psychiatry, also at Tufts. His lab is interested in understanding how the brain controls thoughts and actions based on an internal model of the world. His ultimate goal is coming up with a circuit-based computational description of inference and belief updating that would explain how psychotic states arise and provide a clear path towards intervention.

Clinically, he is an expert in treating schizophrenia. He only works in the inpatient setting where he often manages patients who are acutely ill or decompensated. He has been an early adopter of xanomeline-trospium in this setting, and we will discuss some of his early learnings and clinical pearls from that experience as well. So with that introduction, Dr. Halassa, it is a pleasure to have you with us today.

Dr. Michael Halassa: It is great to be here, Ben. Thanks for having me.

Ben Everett: Yeah, absolutely. So we start every one of these with pretty much the same set of icebreaker questions. We try to let people get to know you a little bit on a personal level. You have got dual degrees; you have the PhD and the MD. Listeners have heard me say this before, but one “D” degree was enough for me. I do not envy all the extra training and time that you guys put in, but I am very thankful for all of your studies and the work that you are doing. From a young age, did you know you were going to do science, or medicine, or some combination of the two?

02:36 – From Physics to Psychiatry: Building a Scientist-Clinician Lens

Dr. Michael Halassa: I knew I was going to be a scientist from a young age because it just came naturally to me. I actually wanted to be a physicist when I was growing up. Part of the reason is that you are doing math and you are doing physics in middle and high school, and this stuff comes easy. You gravitate towards it, and that is kind of what happened.

I wanted to be a physicist, but I grew up in a country called Jordan. It is in the Middle East. It is a small country and both of my parents were healthcare providers. My father was a neurologist. They really encouraged me to go into medicine. Medicine is a degree that you get into right after high school. So there was this tension where I wanted to be a physicist while my parents wanted me to be a doctor, and ultimately I just did what they wanted. That is how I got into medicine.

In medical school, there were things that I liked and things that I did not like. After I finished, I figured I was going to do something that was more “sciency,” and that is exactly what I did. After graduating from medical school, I went into a PhD program at the University of Pennsylvania in Philadelphia. That was great; it was very experimentally heavy. Towards the end of my PhD, I was thinking about what I wanted to do afterwards. I was fortunate to meet many MD-PhD psychiatrists who encouraged me to pursue psychiatry. Now, my father never approved. He thought it was not a real thing and always teased me about going and doing neurology instead, but I am pretty happy with where things have landed.

Ben Everett: Well, I am glad you all can joke about that now, because I can tell you I am very glad that you are in this field for the work that you are doing, and your patients probably think the same. I have actually heard very similar things from other psychiatrists that we have had on for different reasons. Part of it, which we will get into today, is just the approach of medicine and how much medicine has changed in other specialties based on biomarkers.

A lot of times, it is just easier. It is more objective, kind of like your physics. I did chemistry for the same reason. I just loved chemistry; it made sense to me. It was very rational, and that really appealed to me to be able to think about something theoretical, but then do some math and end up with a concrete answer. We do not have that at all in psychiatry. It is very subjective. The DSM really loves the structured clinical interview, and we do not have biomarkers to guide decision making. We do not have an A1C or a measurement of millimeters of mercury for blood pressure or anything like that. We will get into all that later. It is really exciting.

So, psychiatry, you kind of explained why you went into that field. Now you have got your MD training, you have done your psychiatry residency, and you have done your PhD training at Penn. After you did both of these, how did you think you were going to put them both together? Was it just research, or were you trying to find a clinical appointment? I know oftentimes it is difficult to find an appointment that lets you do research and clinic at a rate that is comfortable. A lot of times it is 90% research and 10% clinic, like one half-day of clinic a week, which a lot of people just do not feel is enough.

05:52 – Decoding Brain Circuits With Computational Models and Modern Tools

Dr. Michael Halassa: When I was done with my residency, I did one year of a fellowship in psychotic disorders. I did my residency at Mass General Hospital in Boston. At the same time, while I was doing residency, I was in a research track. I was doing my postdoctoral fellowship at MIT with Matt Wilson.

Matt’s work is very basic and has nothing to do with clinical research. He is an electrical engineer by training. He was the person who discovered a phenomenon called “replay” in the brain. It is the brain replaying awake experience when it is sleeping. That is the substrate for a lot of what people think memory consolidation is. Awake replay is now thought to be a substrate for inference and planning. This is hippocampal phenomenology, which involves recordings from the hippocampus and decoding awake experience, and then looking in sleep at the replay of that experience.

I worked with Matt for five years. I took this approach of recording in behaving animals and decoding the content of neural activity into a new area. I did not do any hippocampal research; I worked in thalamocortical circuits. My early career as an independent researcher was shaped by that. It was purely basic; I did not do any clinical work.

That was really born out of my postdoc. During my postdoc, I did a lot of these recordings in behaving animals and went into this area of thalamocortical research. I discovered things that were really unexpected based on what the textbooks told you. A lot of what the textbooks say is based on very few recordings in anesthetized animals. People put wiring diagrams together and then tell you: “This is how this area of the brain works.”

Once you subject it to modern techniques of what we call “cell cracking,” meaning high-density recordings, behaving animals, and causal manipulations using techniques like optogenetics or chemogenetics, you end up with a very different picture as to how complicated these brain circuits are and the variety of things that they can do.

The biggest thing that came out of that postdoc was that we tend to think of all of our cognitive functions, sensory functions, and motor functions as things that are subsumed by the cortex, which is the outer shell or outer covering of the brain. But a lot of the brain, a lot of evolution, and a lot of the animal kingdom is subcortical. The thalamus is this area that is connected to all parts of the cortex. The dominant view is that it is a sensory relay, but it is connected to all kinds of non-sensory areas of the cortex. My postdoctoral work was my entrance into identifying the different parts of the thalamus that do non-relay functions and things that have to do more with cognition.

That started my independent career. I started my lab at NYU Medical Center in 2014, and I did not do any clinical work from 2014 until 2020.

Ben Everett: I am loving all the background because we have gotten to know each other a little bit and done a couple of projects together, but I did not know anything about the desire to do physics. You just have this love for physics, and I see a lot of physics in this.

Sometimes having a different disciplinary view of something can really change the way we think about it. I am thinking of an old paper that used a quantum mechanical model to understand photon receptors on the retina. It is brilliant work regarding the essentiality of the 22-carbon length omega-3 fatty acid, DHA, which is essential for the eye to work right. It really took quantum mechanics to understand exactly what photons were doing and how DHA worked there. Anyway, nobody wants to hear about that.

I think that is exactly why it is important sometimes to think about things differently and to turn things on their head a little bit. This is a good segue into talking about your paper that we had out. We will put links to all this in the show notes, but we will talk about algorithmic circuit psychiatry. What inspired you to write this paper with some collaborators? You went back to your Substack and said, “I think I maybe went too deep too early.” You went to the core of the onion, and maybe we should have started with the outer part of the onion.

10:37 – Returning to Inpatient Psychosis Care and Reframing Clinical Reality

Dr. Michael Halassa: When I started my career as an independent investigator in 2014, we were basically just doing basic neuroscience: training animals to do various types of things, recording, and interpreting. A lot of what we do in the lab is interpreting what the brain is saying by looking at these “spike trains.” You have a bunch of neurons, and they are basically caring about different things that you are giving the animal. Then you are trying to imagine: “What are these echoes of? What is the generative process in the brain that is giving rise to these things that you are measuring?”

Once you start getting into that kind of business, you absolutely need models. There is no way to interpret or explain things to another human being by showing them thousands or millions of spikes. You have to be able to compress that into a process or a mechanism that you can use natural language to explain. You do that enough and then you find that there is no other way other than having some process model that you can fit the data into to see what the appropriate mechanistic story is that fits the data.

We had a few of these stories from 2014 to 2020. When 2020 hit, it was COVID. I was at MIT at the time. All institutions had to go into a shutdown mode. We were not allowed to do experiments anymore; there was no way to get into the lab. We were encouraged to do other things like analyze data. I remember hanging out in my pajamas all day, having Zoom meetings, and talking to students.

At one point, my wife said, “Well, the psychiatry thing that you did, why don’t you just go back and try to see some patients and see how that works?” I said, “Oh, that is a pretty good idea.” At that point, my medical license had lapsed since I had not done psychiatry in a long time, so I said, “Fine, it is a good idea.” I did that. I picked up a few clinical shifts when I got back to see severely ill patients in the inpatient setting for psychosis, which is what I did.

My mind was completely blown. Spending close to a decade just thinking about behavior and fitting it into models in order for it to make sense, and aligning neural data and behavioral measures together using models, and then to see people telling you things that are so non-normative. The people with psychotic experiences, it is just so amazing that people just do that. I feel very privileged that another human being with an altered experience trusts me enough to tell me how they are thinking. That is a privilege, and I think every psychiatrist should feel that way. There is no reason for people to tell you when they are suffering. There is no reason for anybody to tell you anything. I feel really grateful for all the patients who have taught me this.

So I am talking to some of these folks and my mind is just, I cannot believe this. It is like I am a different person compared to residency and training. I did not have that perspective developed enough in my head to be able to start seeing patients. The closest analogy I can make is like Neo seeing the Matrix. You start seeing the generative process behind the person, like what possibly could be going on in that neural code that would give you this kind of experience or how the person is talking about it.

That was a moment that completely changed my life and put me on a different track. Before 2020, my research was purely directed at understanding what the neural code is, what the thalamus does, and what predictive processing is. These things are very abstract from the perspective of psychiatry and may inform psychiatry in the next 10,000 years. But after 2020, a lot of what I started doing is a lot more relevant and current.

14:47 – Moving Beyond “Chemical Imbalance” Thinking in Schizophrenia Treatment

Ben Everett: Yeah, translational by design instead of just the basic research. That makes perfect sense to me. That is very interesting. Another theme that we have talked about a little bit in the introduction is, and I can speak from a drug development standpoint, even the way a lot of medicine is now, and I know a lot of medical school training and residency is just diagnosing. It is thinking about what drug for this disorder. “Okay, you have got high blood pressure; we can do a beta-blocker, or an ACE inhibitor, or a diuretic.” We have all these different things you can do, but we really think very “cause and effect.”

From a traditional drug development standpoint, we even talked about the Human Genome Project, I was in grad school when that was reading out. There was all this hype and promotion. I thought it was great for the general public, but for people that really understood, it was like, we are so much more complicated than this. This might help for autosomal recessive disorders where we now have gene therapy, or things as simple as a SNP, or a gene that makes a protein that is misfolded. You think of sickle cell disease or something like that. But even with those examples, we do not necessarily have better treatments now for sickle cell crisis than we used to.

It is very reductionist. We were thinking in terms of: “There is a gene that has a problem, and once we understand what that is, we will be able to develop something that fixes the protein or can bind to the receptor that that protein would normally bind to.” But that is pretty reductionist and that really has not translated to the way drug development works. It has not necessarily translated to better medicines or cures, except for a few things like gene therapy. You have taken that concept and flipped it on its head and expanded upon it. Do you want to get into that a little bit?

Dr. Michael Halassa: It has definitely been on my mind how drug development in psychiatry works. It is sort of a “chicken and an egg” kind of problem. On one hand, let’s say we want to develop a new drug or a new class of drugs for schizophrenia. But then you look at schizophrenia and it is not a thing. It is a variety of ways people can start to reason differently.

I think you would be able to talk to many psychiatrists who would probably share this view that people are very different in how they reason. Even the delusions and the altered perceptual experiences are different in different people. If you wanted to use symptoms to stratify patients and say, “Oh, this is the person who thinks that people are out to get them, but also has a running narrative plus this altered perceptual experience,” you are going to end up with this high-dimensional space that is very sparse. People are sitting in different bits of that space. There is no way for you to find common mechanisms for anything.

I hope schizophrenia is not that. My optimistic perspective on what we currently call schizophrenia is that it is not one thing. Hopefully, it is not n-dimensional as the number of patients on Earth. That would be really problematic where everybody needs a particular type. Hopefully, it is a lower-dimensional set where there are three or four core mechanisms that can alter reasoning and perception. Those hopefully map onto particular neurobiological substrates.

But in order to identify what those are, you would have to have some biomarkers or some measurement, whether it is a task-based measurement or a brain-imaging-based measurement. Something that would tell you what those core symptoms are. If you translate them to just number lines, let’s take a five-dimensional space as an example, then you have a list of five numbers for each patient. Then, basically, you can start tailoring drugs to that five-dimensional vector of five numbers. Does that make sense? That is kind of how I hope these things are. And those would be interpretable dimensions, meaning one would be reinforcement learning. I do not know if that is the mapping, but something like that.

Ben Everett: Well, there is certainly a trend that schizophrenia is, I think we could say it is a syndrome, not a disease. In so much of medicine, the more we understand, the more we are seeing expansions, or umbrella terms, or things like that. These things are very often multifactorial. There are myriad gene-environment interactions that are at play. It is very complex, but hopefully not infinitely complex, because that would make our job just about impossible.

Certainly, if it were as easy as: “Boy, Haldol just works great for everybody and it is all we ever need,” we would not probably be talking about schizophrenia when we had Haldol decades and decades ago. The reality is now we have got all these first generations, the atypicals, the partial agonists, and we still have really not done very well for caring for patients with schizophrenia. Somewhere between 20% to 40% of patients still have residual positive symptoms, and we really have done very little to address the larger drivers of functional impairment: the negative symptoms and the cognitive impairment.

I think this sort of idea of algorithmic circuit psychiatry could hopefully shed light on some of these other areas that could lead to better solutions, whether it is a drug or something else, to help these patients that are suffering. I just used the term, you started with “algorithmic psychiatry” in your Cell paper, and you have expanded it now to “algorithmic circuit psychiatry.” Maybe that is semantics, but why did you feel that change was necessary?

20:43 – Fixing Computational Psychiatry Limits With Mechanistic, Circuit-Based Models

Dr. Michael Halassa: You know, there has been a long tradition in cognitive science of taking choice data from tasks and fitting it into models. These are very simple things. For example, you can have a simple perceptual task where you show somebody moving dots. They can move mostly to the right, mostly to the left, or somewhere in the middle. You ask them, “Is it moving right or moving left?” They give you judgment data. Based on that data, you can fit something called the psychometric function. It is like a sigmoid. For every subject, you can have a sigmoid, and some people are really, really good at detecting motion and some people are terrible. For each person, you can evaluate how good they are as a psychophysical observer. Those are very simple models because you just have one variable, pretty much.

But when you get into things like learning or making decisions on more complicated types of scenarios, then you need slightly more complicated models. Reinforcement learning is a good example of that. Temporal Difference Learning is a particular algorithm that was developed to solve reinforcement learning type problems, which is basically: how do people or machines find the best option among a set of options?

What are the ingredients that are required for that? Well, you have to be able to encode a value, and you have to be able to update that value based on feedback. That is the simple architecture that you need: a value, an update function, and feedback from the environment. You have to have some mechanism by which to increase the value if the feedback is positive, and decrease the value if the feedback is negative. Then you have a running average of how well you think your value of a particular option is.

Based on those models, you can fit a lot of choice data for people doing reinforcement-type tasks. To be honest with you, without simplifying computational psychiatry too much, a lot of it is that. It is taking different patient populations and running them through “bandit-type” tasks where they have to learn one of two or three options. Then they show different types of biases: either they are too sensitive to negative feedback, or they do not update very much, or they do not switch their mind very much when change points happen.

People have had a lot of hope that this would be a good way for us to be able to go back into the brain and see: “Okay, what correlates with this parameter that I got from my model?” Whether it is the learning rate, the positive learning rate, or the negative learning rate. Whatever parameter you have in the model, you can look into the brain and do some regression on fMRI data, and that would tell you: “Oh, this has changed in depression” or “this has changed elsewhere.”

But again, it is a chicken-and-egg kind of problem because depression is not one thing. It gets really complicated. The biggest limitation of that approach is that the form of many of these models is not meant to derive neural mechanisms. Nothing about the form of a Temporal Difference Learning algorithm is necessarily indicative of where in the brain some of these components could be perturbed. That is a challenge.

One of the things that we have been thinking about, and other people have been thinking about, is: how do you derive normative models that can fit to behavior, but that can simultaneously point to a brain mechanism? How can you show that this person is not updating their belief because of some problem in inhibition or excess excitation? How could you tease those things apart? Ultimately, when you want to develop a medication, the substrate that the medication is working on is excitation, inhibition, and things like that.

I am not saying that I have the answer. I am just trying to point out where I think it would be fruitful to dig. This idea of developing algorithms that have a mechanistic component to them, that you can still fit to choice data or more complicated data like eye tracking or movement, could be particularly fruitful.

25:08 – Creating Task-Based Biomarkers to Measure Belief Updating and Reasoning

Ben Everett: It seems to me it is almost another one of these areas where technology is allowing us to go from binary to digital, and then you can go to quantum from there once we figure that out. Certainly, I feel like the brain is one of these areas of science where the more we know, the less we know. We have all this fMRI and imaging data going back 20 or 30 years. We understand circuits a little bit better, and there are some things that we have learned, but I do not think we have learned as much as we had hoped to learn after a couple of decades of imaging. We have got a long way to go.

Getting back to your article: mental illness is not necessarily about a broken brain part. In a neurodegenerative state like Parkinson’s disease, we know exactly what happens. We have these striatal dopamine neurons; they are dying off and we are not having dopamine signaling there anymore. That works for certain neurodegenerative diseases, but it does not necessarily apply to depression, eating disorders, schizophrenia, mania, or psychosis. Why do you think this is more precise than a chemical imbalance? It seems like in the 70s, psychiatry went to this whole receptor-mediated view of everything, thinking we just have to figure out where the chemical imbalance is. I am not sure that is nuanced or precise enough.

Dr. Michael Halassa: I do not think the evidence is very good that any particular disease is simply an increase in serotonin or a decrease in dopamine. There is just no good evidence for that. Furthermore, the brain is far more complicated than a drawing of a synapse. The dopamine in the ventral striatum versus the dorsal striatum versus the cortex versus the thalamus has very different meanings.

Inhibition on an inhibitory neuron is very different than inhibition on an excitatory neuron. The meaning of that from the perspective of the circuit and the neural code is why thinking about the neural code is really important. The brain functions in packets of information sent between different areas, and these are different patterns that have to be interpreted in the presence of noise. There is a lot of circuitry that goes into trying to make that communication efficient: the balance between excitation and inhibition, and the precise role of different types of neuromodulators.

A lot of this is really hard to describe in natural language right now because we do not have really good models for it yet, but I think the models will come. Those models will never be reduced to the image of a single synapse. It is just not how things work at all. People tend to say, “Oh, GABA is inhibitory.” Well, yeah, sure, it is inhibitory if you are taking the perspective of a postsynaptic neuron, but it can have a variety of actions depending on the circuit. You just need one synapse in between to reverse the sign. You can inhibit an inhibitory neuron and get excitation. That is actually a lot of how gating happens in the cortex: you gate through disinhibition because it is a very powerful mechanism to implement things like gating.

I think that there is probably some broad mapping between neurotransmission, neuromodulation, dopamine, serotonin, and broad changes in brain state. But I think we have a lot to learn in terms of trying to understand how to think about the level of chemistry and behavior because there are so many different levels of organization in between. We absolutely need to be able to model that in order to predict how an effect at the chemical level will translate to behavior. There is not a simple transform where “this goes up and people become happy.” That idea, in my mind at least, is dead.

Ben Everett: It makes sense to me. You have specifically outlined four goals for algorithmic circuit psychiatry. Goal one is establishing objective measures. That makes perfect intuitive sense, but why don’t you expand on it a little bit?

29:10 – Operationalizing Delusions Through Counterfactual and Decision-Making Tasks

Dr. Michael Halassa: Maybe I will give you a really good example. In schizophrenia and psychosis in general, people can come in with very fixed beliefs about things that are objectively incorrect. “I have a bottle in my chest and that has replaced my heart and it has been pumping my blood.” Something like that; something that you would feel no reasonable person would be able to hold. Yet this person, who could be reasonable in other ways, just has that belief.

This probably is a reflection of an altered reasoning style. If we knew exactly the kind of changes in the underlying circuitry that could give rise to that, we would be able to help this person better, especially if we can push the circuitry in different directions by drugs or other methods like TMS or ECT.

How do we get at that reasoning style? Well, we have to be able to have some way of objectively measuring it the way we measure an A1C. In the literature, there exists a set of research on this topic called “bias against disconfirmatory evidence.” People have done cognitive research in healthy individuals, collected a bunch of normative data, and looked at patients with schizophrenia and delusions in particular. What it looks like is that in multi-step decisions, when decisions are complicated, people tend to establish some initial belief and then nudge it in different directions as they accumulate more evidence.

People with high delusional content push it into some initial state and then all the nudges do not move it. That is a particular way of thinking that we are lucky not to experience. In neuroscience, we tend to think of it as part of the brain: whatever is holding that bit of information gets into an attractor state. It is like an Ising model in physics where it is kind of stuck in that attractor state. Healthy individuals can nudge it into other states, but people with psychosis or delusions have a problem with that.

Part of what I have been trying to do recently is come up with a quick set of measures on the inpatient unit to be able to measure that process. It would give you a set of numbers for individual patients that would track this belief-updating process. I think I have something that works in that direction.

Now, part of the issue of the past and the limitation of computational psychiatry is that models to fit that data had not existed before because they use natural language. But guess what? Now we have Large Language Models. You can train Large Language Models on exactly these types of tasks and you can ask them to tell you how they are encoding these beliefs and how they are updating them over time. Then you can use the Large Language Model as your in-silico model to try to understand how something like that can come about. Then you need to take that model and somehow map it onto the brain. I haven’t figured that part out, but that is the process by which we are approaching a problem like that.

Ben Everett: Well, you have got to have data, right? It is all about finding the data that gives you a clue to whatever symptom or phenotype we want to discuss. Phenotype is probably not the right word.

Dr. Michael Halassa: I think it is close enough. There is something that cued me into doing that. I will just ask you the question, and this is what I ask a lot of the folks that I see: “Tom and John missed their train. Tom missed it by five minutes. John missed it by an hour. Who is more upset?”

Ben Everett: I am not sure. You could say Tom, because he was really close. John was like, “Yeah, I was an hour away; I was never going to make it.”

Dr. Michael Halassa: That is called counterfactual reasoning. “What if I had been five minutes early? I would have caught that train; I would be less upset.” Folks have a really, really hard time doing that if they are in a highly delusional state.

That was kind of the impetus for me to start developing tasks that could capture that. I think that is where the algorithmic circuit psychiatry idea is most useful. You know exactly what the phenotype is. Phenotype is the right word, I think. We can communicate using that. That is the phenotype: the person is having a very hard time being able to do “if-then” type of reasoning. You come up with a task that is all about “if-then” type of reasoning. That seems to correlate really well with what you see clinically and delusionally. It is a good measure. It is a reductionist approach because ultimately you are going to get a list of numbers, but it is not a single number. It is a pattern that you can then study with neuroscience.

33:59 – Translating Algorithms Into Drug Targets and Better Animal Research

Ben Everett: There is so much I want to say about that. I am just thinking back on some of my work and some of the stuff I have done previously. This is a totally new way of thinking about it, but it is something that I have argued forever regarding textbooks and just how old and dated some of them are. So much stuff is based on “frozen slices.” We took a brain, we froze it, we took these slices, we did some immunohistochemistry, and this is what we saw. It is like: “So that is what the brain looks like.” I am like: “Yeah, that is a snapshot of one point in time.” You do not even know if you caught the point that you were trying to catch.

These are incredibly dynamic processes. I think it was overly reductionist, but people were using the best technology and resources they had at the time to come up with things. I think we just have not done a very good job of going back and updating some of these textbooks in the way we are teaching younger students. But that is neither here nor there.

All right, so number two: reverse engineering deficits algorithmically. I think this is a little bit more intuitive based on what you just said.

Dr. Michael Halassa: Exactly. That is basically what I am saying: once you have this data, what kind of model form do you use? What model is actually able to compress the data in a way that is interpretable? The Large Language Model example here is a good example, but there are other models that people can use for other types of data.

Ben Everett: And now once you have this, essentially you have got a biomarker, and I am using air quotes for people, so you can go to the drug developers or the people who are actually researching and say, “Okay, so now you have got the biomarker.” Again, I am just using that term because it is the term I think is going to resonate with people. You can use this now to move forward.

So much has changed with drug development, but in-silico is just a lot of what you can do now. Now you have got a better idea of how to move forward. I would imagine it would be much more precise. You could probably do a lot of this in modeling and, hopefully, we would not have the failure rates that we have in Phase 1 when we are going from animals to first time in man.

Dr. Michael Halassa: Absolutely. I have to say one thing about this: measuring resting-state fMRI or EEG or whatever it is that you are measuring has a much, much lower rate of success than being able to align this to task events. I have learned this from working with animals for a long time.

The only way to understand what a circuit is trying to communicate is to put the circuit in its optimal regime. If you think the frontal cortex is important for “X,” give the subject, whether it is an animal or a person, a task that engages the frontal cortex. Then you are going to be able to read out whatever the neural code is in a meaningful way. If we give the animals a task that we know engages the frontal cortex, we can completely interpret what the frontal cortex is doing. If they do some random behaviors, it is a very, very hard thing to do. I think that is kind of the issue with biomarkers: there is no behavioral clamp. In psychiatry, you absolutely need a behavioral clamp.

Ben Everett: And biomarkers… while we rattled a few off, there are certainly a bunch of other ones we could mention. They have certainly come a long way, but a lot of people would think we are further behind in biomarker development than we probably should be or would hope to be. I have worked in a number of different therapeutic areas and this issue of biomarkers is always a hot topic. It is like, “Yeah, we tried this one and it did not work.”

Even with A1C, all these medicines that reduced A1C never actually reduced cardiovascular events. But when we finally have ones that reduce cardiovascular events and heart failure, whether it is the SGLT2s or the GLP-1s, they do not necessarily move A1C. So it is kind of like, “Ooh, I don’t know.” That was a fun Advisory Committee meeting to listen to way back when. So, rule number four would be defining the role for neuroscientists. So, expand on that if you would.

37:54 – Using LLMs and Machine Learning to Test Psychiatric Mechanisms In Silico

Dr. Michael Halassa: I think there is an amazing thing about discovering just the language of the brain. Just being able to say: “Okay, I really understand how the brain can keep a thought in mind,” or “how can simple addition of different values occur?” How does it take different things and put them in some common currency? How do you compare a vacation in Hawaii with three weeks in the gym or something like that? These are very, very different types of things, but the brain somehow can put them in some common currency and be able to compare them with one another and make decisions. We are making decisions like that all the time.

The language by which the brain does that is fascinating, and I think there is a lot to be said about why it is really cool and awesome to just discover that. This is not just neuroscience; it is everything. I mean, it is great to discover how galaxies work or what happened before the Big Bang. All of these things are exciting.

But if the goal is to address the burden of a particular type of disease, then you absolutely need to be motivated in a particular way to ultimately solve that disease. I think that some of it will come from things that are basic and serendipitous and had nothing to do with the disease. That is true. A lot of what we learned about action potentials, for example, came from the squid giant axon from people who did not think about mental illness or anything like that.

That is fine, but I think if we wanted to make rapid progress in a particular area of mental health, it is really important to be able to know what those things are. Those are not just an additional sentence you can put at the end of your R01 and say: “Okay, and by the way, whatever mechanism I have will solve schizophrenia, bipolar disorder, blah, blah, blah,” and a bunch of different boilerplate type of things that you have heard on a podcast.

So I think that these types of measurements that you can get from patients are hopefully, like just the one that I described, will be able to inspire enough people who do decision neuroscience to say: “Ah, now I know exactly what type of experiment I can run in whatever animal I am studying and be able to look with high resolution at how the brain does that.” And that is kind of very, very different from a forced swim test as your pre-clinical measurement for depression. That is kind of a different type of neuroscience.

Ben Everett: So we have got the four goals for algorithmic circuit psychiatry: one, establishing objective measures; two, reverse engineering deficits algorithmically; three, giving drug developers algorithmic targets; and four, defining the roles for neuroscientists. We did a lot of talking about that, which was very helpful for me, and I appreciate it. We talked a lot about this computational psychiatry stuff already. I think the role of machine learning in circuits… do you want to get into that more?

Dr. Michael Halassa: Only to say, honestly, that the development of machine learning was originally inspired by neuroscience. Now, the fact that these things… before machine learning, a lot of computational neuroscience, a lot of these models of neural circuits really were not capable of doing very interesting behaviors. They were not powerful enough; there was no way to wire them in a way that would make them very useful.

But machine learning has solved that problem. It basically developed the types of algorithms that can wire neural networks in ways that make them really solve very difficult things, like they can beat the world champion in chess and Go and do all kinds of amazing things like machine vision. I think the opportunity for psychiatry is that now you have very powerful neural models that can solve things in ways that people can solve them potentially, and you can learn a lot about how altered thinking can arise in people by studying it in machines. That is really the big upshot of all of this.

Ben Everett: If we think about a book like the DSM-5 and the way we think about diagnosing mental illness or disorders now, how do you envision this informing the way like a DSM might work, or at least moving beyond categories? I do not know exactly how to say that, but right now you have got a chapter on schizophrenia and schizoaffective disorders and it has got all this different stuff in there. It would need to be much more nuanced, perhaps much more refined. Maybe this is just semantics. We have these books, but if we have these big computer models that can just tell us what is going on, do we need DSM categories like that anymore?

Dr. Michael Halassa: I think we have to be realistic about what these models can and cannot do. Right now, they think and reason “off the shelf” very differently than how people do it. They are not wired with the same objectives, they do not interact with the real world, and they have very different, for a lack of a better word, drives and motivations than we do.

A lot of what we do is basically tell each other stories that make sense from our experience of the world. These models… the DSM does a fairly reasonable job at that. A list of numbers without being put in a form of a compressible story that we can communicate with isn’t going to be very helpful just in communication. I do not think that these models can replace people, and I do not think they can replace the systems that currently exist.

I am not sure the DSM is such a terrible thing. We use it for a lot of things. I think it is useful for many things, but at the end of the day, these are not the categories, probably, that we are going to end up using to develop mechanisms. I am not sure what the alternative is just yet.

Ben Everett: Well, I certainly hopefully did not imply or denote that I thought the DSM was terrible. I am just a neuroscientist and biochemist.

Dr. Michael Halassa: No, I know! I know, I know.

Ben Everett: I am just not trying to get in trouble with anybody.

Dr. Michael Halassa: I mean, this is now like the favorite thing for people to say.

Ben Everett: Look, I have seen that on Law & Order too, but it is not just the DSM. I think the DSM is put together as any type of book like that in medicine is. You have got a group of really intelligent people working in concert and doing their best to take evidence and put it into, and I love that you said “story form,” because I think without the story to help communicate, that is where we end up missing a lot of things.

Dr. Michael Halassa: All of science is stories. Even the equations are stories that just are in a particular form.

Ben Everett: You just became my favorite person. Even when I lecture to high school kids or whatever, I am like: “Yeah, you want to major in science, that is great, but you cannot forget the… you have got to do the humanities. You have got to be able to communicate, and you have got to be able to write.” You can have the best data in the world to put in an R01, but if you cannot communicate why it is important, it does not matter.

44:57 – Redesigning Animal Models to Validate Causal Brain Circuit Algorithms

All right, so circuits. We had not really talked that much about circuits yet, and they are really central to your framework. How do systems neuroscience methods like electrophysiology and optogenetics help us test whether an algorithm is truly implemented in a circuit?

Dr. Michael Halassa: That is basically what I was mentioning before. Let’s give the Temporal Difference Learning algorithm as an example where you update your value based on feedback. That is a really, really good example of a success story in neuroscience. When Wolfram Schultz recorded monkey dopamine neurons in the VTA, basically what he did was exactly that: do classical conditioning on the monkey and record from these neurons.

What the pattern of activity of these dopamine neurons ultimately was able to fit is the reward prediction error variable in these Temporal Difference Learning models. So basically, if you have a particular value for an object and your feedback is exactly the same as that value, you do not update. If the feedback is higher, you increase the value; if the feedback is lower, you decrease the value. That is a very simple kind of algorithm to adjust the value of things in the world.

The prediction error can be positive, meaning that the reward that I got is higher than what I expected, so the value goes up. If I did not get a reward, that is a negative prediction error. That is how dopamine neurons function, which is pretty amazing. You can fit the firing of dopamine neurons to Temporal Difference Learning algorithms. You can replace the reward prediction error with dopamine neuron firing. That is pretty cool.

Ben Everett: Absolutely.

Dr. Michael Halassa: A bunch of people now have expanded on that, and it turns out that there is a whole distribution. There is something called distributional reinforcement learning, where the reward prediction errors are biased towards positive. The neuron mostly does positive reward prediction error, while some neurons mostly do negative reward prediction error, and you have a full distribution of positive and negative.

So some neurons are more optimistic; they are like: “Yeah, let’s just update it when it is good and not worry about it when it is bad.” Some neurons are very pessimistic. One of the things that we are… I think it is really exciting in that particular subfield of neuroscience is: do these neurons project that distribution to all circuits, or are some circuits ultimately more pessimistic? Are some circuits more optimistic? Are some circuits more balanced?

It looks like the ventral striatum, for example, or the nucleus accumbens is more balanced. It looks like some circuits we have been looking at look a lot more pessimistic, and that is important for them to do particular types of operations. That is kind of the gist of it. Now once you have that kind of level of detail, you can start thinking about: “Okay, when you have some sort of change in the circuitry that makes you vulnerable to a particular type of mental illness, could you start thinking about that level of detail that would explain why certain things can be done with an optimistically biased versus a pessimistically biased distribution of reinforcement learning machinery?”

Ben Everett: It is interesting trying to kind of wrap my head around some of it. It intuitively makes sense. Again, if you could kind of figure out what is the healthy response, and then you can figure out: “Okay, this is an abnormal response,” and if we can have a way to alter that response to shift it back to the healthy way.

Dr. Michael Halassa: Exactly.

Ben Everett: The next thing that you went to in the article is highlighting artificial neural networks acting as test beds for these hypotheses. You kind of alluded to this earlier. These are sort of like a new in-silico type of experiment, except you are using Large Language Models instead of a chip. How do you think that is going to help us understand our biological brains?

Dr. Michael Halassa: Large Language Models have their particular type of machinery. They are neural, but they are wired in a very different way than the brain is. But at least they may expose a particular type of algorithm, like a set of instructions that they follow, that we can start to think about based on our knowledge of the brain. Where are those sets of instructions potentially implemented?

We can basically run these types of tasks in people. That is why you still need people; you cannot do this in a Large Language Model entirely. You still need brains to tell you where those things are being implemented. You run it in a scanner, or with some high-density recording, or something that allows you to kind of dissect the neural process. Then, on that basis, you can see what medications potentially push it in different ways and so on.

Ben Everett: That would also alter the way we move from in-silico to animal studies. You mentioned you do animal studies as well. But instead of thinking of it like: “Okay, this animal, when we knock out D2L or whatever, you end up with this phenotype,” and then we have an “ob” rat or a Parkinson’s mouse…

Dr. Michael Halassa: Exactly.

Ben Everett: They are always models, but they are not even… again, I will just come back to: we are much more complex. The first trace amine receptor trial failed, and a lot of that was just because we have a better understanding of circuits. Our circuits work in a more complex manner than they do in the animal models where they originally did the work. How would you imagine this shifting the way we are thinking about doing animal research?

Dr. Michael Halassa: I do think of animals as extremely useful and valuable. You are not going to be able to take a person and stick electrodes in their brains and measure whatever neural activity from any brain area. You are not going to understand the neural code at that level of resolution.

And you are certainly not going to be able to inactivate and activate circuits the way you want to be able to test causal hypotheses. If I basically say: “Well, I really think the way a decision process works is that information comes from posterior cortex into one area, let’s say orbitofrontal cortex, they get loaded up, and then they get sorted and go to dorsolateral prefrontal cortex, and then they get added up in some other area.” That is how I imagine the algorithm is implemented of value comparison or adding up different values from very different sources or attributes.

How on earth are you going to do that in a person? In order to really know that this is how it works, you have to be able to record neural activity at high resolution, and you have to be able to systematically inactivate these areas at different portions of the task to be able to be sure that, yes, this is how it works. And I think that is what animals are for. I do not know what the true utility is of animal models of a particular disease. Some are probably okay; some are terrible.

I tend to think of animals as a way of really dissecting something that we think is interesting in people at a very high resolution and with causal detail. You can know for sure that this area is important for this process in the algorithm because you can inactivate it at that point in the task. And that is what optogenetics is. Optogenetics was developed by a psychiatrist, by the way.

Ben Everett: We probably need to have an episode on optogenetics because it is kind of newer to me also, but it is pretty fascinating.

Dr. Michael Halassa: Karl Deisseroth is a psychiatrist at Stanford who developed that technique.

Ben Everett: I have read a little bit about it and talked to some of those people. So as we move into… let’s say we figured some of this out. It is 10 or 20 years from now. Maybe it is five years from now.

Dr. Michael Halassa: Maybe it is five years from now, hopefully.

Ben Everett: Hopefully. The way these Large Language Model models are increasing and learning and getting exponentially better and faster, who knows? So let’s say five years from now we have kind of shifted the way that we think about a lot of these things. How does that change impact the way psychiatry residents are trained?

53:03 – Training the Next Generation for Precision Psychiatry

Dr. Michael Halassa: I think that psychiatry residents have to have some hands-on experience with kind of structured behaviors, tasks, potentially collecting data, and thinking about what data forms are important. Psychiatry is a lot more complicated than the rest of medicine for the various reasons that we have been discussing.

The burden is on us to be able to think about ways to train the next generation to be able to tackle that complexity. I think that there is going to be a next generation of computational psychiatrists who will be trained in these various disciplines and can ultimately come up with quick approaches to collecting that data faster and using these measurements longitudinally for treatment, prediction, and response.

Ben Everett: And I think that kind of gets back to… you wrote once that you had a patient ask you why psychiatry seemed so imprecise. How do you see this answering that question for a patient? You know, it is one thing; we have had a very erudite conversation here, and I am not sure a lot of patients could access it at this level. So how do you take this and distill it to a level where a patient says, “Boy, why is so much of psychiatry more trial and error? We are going to try this one and we are going to bump the dose. Okay, that one did not work and now it is three, four, or five months later, and we are going to try another one.” There is a lot of imprecision that we do not have in other areas of medicine. So how do you explain this change to patients?

Dr. Michael Halassa: I talk to patients about that and I basically say… because sometimes patients do say, “Can’t you just put me in a brain scan and figure out what’s wrong?” And my answer is, “I wish that was possible.” I mean, I would do it.

There is nothing in our current technologies that can allow me to take whatever measurement I get from you and distill it into something that I can communicate to you as an explanation for what is going on. We are just not there. And therefore I am using you mostly as the arbiter of what is working and what is not. That actually is a very good way of getting buy-in from people.

When I talk to patients, I talk to them like fellow human beings. I do not see a distinction between them and me. I am just lucky enough not to be experiencing what they are experiencing, but that could change tomorrow. I could be in their situation tomorrow. You never know. None of us are immune to these things. We do not understand how they come about, and we are all vulnerable. So I think it is important to keep that in mind.

I try not to… and that sometimes can actually go both ways. Some people really appreciate that I can be very kind of collegial with them and talk to them like equals. Some people do not like that very much and want to be talked to in a different way. That is something that I have to feel out. Sometimes I start talking to people like they are just my equals, and I adjust and basically talk to them like traditional doctors and just tell them what they are used to hearing. But my default is to talk to patients like equals. That is my default.

Ben Everett: That is empathy and compassion at work.

Dr. Michael Halassa: Because I do not think that, on average, there is anything different about patients, or their level of intelligence, or their ability… this stuff can be complicated for neuroscientists too. It takes us a long time to be able to understand it, and I hope that I find at least enough stories and analogies to be able to explain it to somebody who is actually suffering.

56:44 – Defining Clinical and Scientific Milestones for the Future of Mental Health Care

Ben Everett: What do you see as the biggest barriers, scientific, clinical, cultural, whatever it may be, to bringing this vision to reality in the next five years?

Dr. Michael Halassa: I would say the biggest unknown is this issue of how many mechanistic components some of these illnesses have. We just do not know. If they end up being a small set of things, then at least the science part of it will be quick. If it is more than that, then maybe double the time or triple or whatever. If it is an intractable problem, it is an intractable problem. We just do not know. I hope it is not.

Ben Everett: I think any reasonable, objective person would just look at what is happening with HHS, the FDA, and everything else. Vaccines and “Tylenol causes autism” and all this sort of stuff. We are not grounded in very good science right now. We are seeing people leave very important posts; Francis Collins resigned.

I really want to get into this mental illness issue, especially in these young men who seem more prone to it. We have got big problems with gun violence in this country, and there are some people that I have reached out to that we are going to explore some of these things with as best we can.

This transformation… you note in your Substack article that this transformation is not a matter of feasibility, but really of time. And so what milestones should we listen for that tell us we are getting closer or on the right track? When are you going to know, “Okay, this is like we’re close” or “this was a big step forward”?

Dr. Michael Halassa: I think this is happening, and maybe this would lead us into the next segment, which is the rise of things like the muscarinics. That is pretty awesome. For 40 years, we have had this one knob that we could turn, which is the traditional antipsychotic D2 blockade.

They are different from one another and they have different affinities, but for the most part, they work very similarly and seem to nudge potentially dopaminergic action in a particular way. For some people, particular medications work; for others, they don’t. We do not really know why. But now we have a different dial. So now we can at least move people with two dials into potentially states that they can be more functional at.

So I am very happy about that, and I think that is real progress in psychiatry. There are going to be more of those types of stories in which we see success. I think that success can be made faster and us being able to get to these milestones faster once we have some of these algorithmic circuit objectives met. When we have some of these assays deployed and use them as biomarkers, we will be able to know things much faster.

So that is what I would know. Once we have these traditional emerging medications tested, and we have them validated in some of these tasks and some of these biomarkers, I think that is when we will know that we are off to the races. And I am optimistic about that.

Ben Everett: Well, this has been a really fascinating discussion for me to really think through. I have a better understanding of your writings now and sort of your vision for how this is going to move forward. I really want to thank you for your time this afternoon.

I did say in the intro that you have been an early adopter of xanomeline-trospium in the inpatient setting. I think just for the sake of time, we are going to have you back and we will talk through that. I know also a little birdie told me that you have written some of that up and it is moving towards publication, so maybe we can get this podcast to come out about the same time your paper comes out.

But I can tell you it is very exciting stuff that Mike has been doing in the inpatient setting. He is also doing really groundbreaking, exciting work in the lab. This was a very fascinating conversation. I really thank you. I want to invite everybody to follow your Substack; it is michaelhalassa.substack.com. I am really going to continue to follow the work your lab puts out in the future; I am sure we will stay close.

Please join us next time as we continue with Dr. Halassa and talk about some clinical pearls of early use of Xanomeline-Trospium in the inpatient setting. Until next time, this has been the JCP Podcast, insightful, evidence-based and human-centered. Thanks for joining us and until next time, stay curious, stay informed, and take care.

Episode Resources & Full Transcripts