Apoha emerges from stealth with $36M to teach AI how matter behaves

0:00Hello and welcome to Path Founders with me Mike Butcher. We like to look at the code entrepreneurs entrepreneurs produce, the capital that backs them and the consequences. Right now, science can sequence molecules and model molecular structures, but a new start-up thinks that the missing layer is behaviour, how molecules, formulations and materials respond under stress, in liquids, in conditions closer to reality. That matters for everything from drug discovery and antibody development to food, materials, science and the next wave of physical world AI. That new startup is Apohar and I'm joined now by the two founders, Shamit and Anishka.

0:47And Anishka, you are the co-founders of Apohar, the frontier tech company looking at this space. What I'd like to do first of all is just let's answer the basic question. What on earth are you doing? Thank you for it, Shaman. If you ever have thrown a pebble in a pond and you look how the waves come out of it, those waves exist anywhere, everywhere and anywhere. And basically we have built a new platform that essentially studies those waves to understand materials. and that's the fundamental real breakthrough here because ultimately that is the kind of data that really does not exist when we want to understand materials that surround us and yeah and it's it solves all kind of problems particularly to teach machines to touch taste and smell yeah there are startups producing simulations especially in industry say modeling how an engine part might behave inside an engine.

1:44We've seen a few of those startups around such as Cusp AI, etc. What do you think you're doing that's different to what they're doing? I think fundamentally we always ask this question, why simulate when you can measure? And if you think about the holy grail of material science, it actually always gets to the function and behavior of materials, how they feel, like how sticky it is, how stretchy it is, how it is going to be injected in the body. Will actually it get injected in the body? Or if you think about something like tangy, what is tanginess? So everything that we think about from material science, how these materials behave, that is a class that contributes a lot to the function in different conditions and different environments.

2:30That is something that has not been possible to measure in the past. And so at APOHA, we are measuring this behavior, this complexity of materials in different conditions, in different environments, and that is what we are unlocking. So simulations are dependent on the classes and the algorithms, the way they are structured. We believe for behavior, you finally have a measurement and grounding in physical reality. I think just really just to basically sharpen on that gap effectively, Like any company that is building models, they are dependent on the sources of truth that they have around them.

3:08So as you introduced in the production, sequence and structure when it comes to protein is that ground truth for them. And then you build a model to predict behavior. What we are essentially doing is kind of giving behavior itself a structure in a data class. So essentially you can now triangulate from composition and behavior and model can learn from both and then be more accurate and faster and more efficient. So you're really looking at actual reality as opposed to simulation. That's absolutely fascinating. You've just come out of Stealth with, I think,$36 million round funding. Was that a Series A, a seed round?

3:46It was a Series A. What was so important for you to remain in Stealth for probably a while? And how long were you in Stealth, Paul? Yes, we started the company in 2021. and while the science that Shamit and his peers were building is much older, I mean, Shamit himself was working on it for 10 years before we started at OHA. I had a front row seat and we were watching these amazing discoveries happen in real time. Every time there was a moment, it was almost like, oh, these are the things I used to read in textbook, you know, what Newton and Einstein and likes were done. So this group was doing something marvelous.

4:22And every time they were finding something, it just felt this has to be something that gets out of the lab. So when we started the company in 2021, we very much focused on let's put our heads down and take this amazing science and build it into the technology platform that is not just a scientific concept, but something that is accessible, something that is available to other scientists. They have worked on it, leveraged it, and then it has moved the needle for them. So that is what we've been doing. And while we're coming out of stealth for the world we've been working with many partners so we've actually done more than 40 projects with multiple customers multiple research institutes always testing our technology and how that is unlocking things that was not possible to do before was the strategy to stay in stealth for reasons of intellectual property or competition from other companies who do you feel you're competing with absolutely so i mean the intellectual property is a big part right because this essentially our IP layer goes as close as you can get to physical laws before you can patent them so building a huge portfolio of patent was part of this because we wanted to cover every layer it's a whole stack that we have protection on now and when it comes to competition essentially anyone in the space who essentially are interested in modeling behavior ultimately made modeling materials in a way are working in the same space with us.

5:55So how do we essentially work with them is something we are still figuring out but that's where we are. So your first product out of Apohar will be something called Vibe which is I gather a readout that uses tiny samples and applies controlled stresses and captures wave patterns. This is very deep science but for the non-scientific watcher and listener who is this product aimed at what kind of companies will be using it etc maybe I can just start with like kind of drawing the analogies from my background so I used to work in finance and one of the things that we really closely always looked at was the risk profile and if you think about from a material design perspective the whole problem is structured around how can I get confidence on the early designs and make sure these are my final materials so if you're working on a drug discovery then you have a funnel very risky molecules you reduce the risk over time so what why does is give you that risk score at the earliest stages of your discovery as possible which is something that is impossible to do today so why tells you hey great candidate but how likely it is to have a good behavior that will allow for later stages of successes and a lot of times molecules fail at later stages so you've spent a lot of time effort money love them and then years later they would fail whether in clinical trials or earlier in the process so so why is that risk score it gives you a very high amount of confidence and if you're flagging that your molecule is risky then there's a very high likelihood that that's a bad molecule that you can then improve upon tweak their properties but at least you know that early on and will you be able to expand beyond say insurance companies and finance yeah i mean uh fundamentally this is exactly like the analogy was from insurance and finance but it it's applicable everywhere right like people who are designing materials they want to identify risk early on and it's kind of the insight that helps them to decide whether you you know you want to invest more and get exactly that molecule to the to the patient or to the customers but yeah pharma is a is a key focus area but I mean we have worked with companies for example that are making alternate chicken and that is really like a very beautiful example in my mind which kind of really anchors what is exactly the new thing here so typically if you are like a like someone who is designing a new alternate chicken you are essentially your problem statement is find me a plant protein that behaves like chicken protein, right?

8:35And you can, if you try to do it today, you kind of do a lot of hit and trial because essentially you can find plant protein and you can do all the analysis. The analysis can tell you what that protein is but not how it behaves. And one of the companies that we helped very early on was essentially to be able to show them that the data that we generate, it actually, if that data is similar between two proteins, it tells you if their behavior is same, not this structure or composition. And that's how essentially, you know, we basically help them get the product into the market. There's been a space of companies coming out of the gate recently that crossed this chasm between biotech, pharma, and just, I guess, the traditional tech industry or tech startups.

9:16Many of them powered, obviously, by generative AI on the new ability of AI to be applied in certain ways. Isomorphic, for instance, as well, coming out of DeepMind as a spin out. what's your positioning in that sort of universe of these new kinds of companies where you're applying AI to these much more deep science problems? No, that's a great question. I mean, it's a fundamentally new technology layer or science layer that we bring into this, I think, the new space and the new area where people are excited about it's physical AI, right? So generative AI, which we are now basically all familiar with, generates text, images, etc.

9:59But in order to generate material, you need to have that kind of, like, when you generate material, what do you, what is your specs for the material to be generated? They are about behavior. Oh, I want this to be viscous, I want to be sticky, etc. And I think, like, what that leads to is kind of word model. Like, essentially, you don't want to just generate, you want to test the hypothesis in the real world, and then come back, and then improve your models, and do it in cycle. And I think this wave kind of started already a few years ago, when people started putting labs in loop with AI. And what we really bring to is like, you know, automation of labs can only make things faster and more organized, but they don't generate new fundamental information that is needed for doing this iteration properly.

10:42So that effectively, you know, like labs are working in a way with two or extension of two of our senses. Most of our scientific instruments either look at the material or they, you know, they kind of test it in very sparse ways. It's like adding three new sensors to these automated labs. And so essentially that is the layer. It's much earlier in the stack where we fit and then it's basically inevitable essential data that would be needed by everyone. And I was reading that you identified high risk antibody candidates with more than 90 percent precision. How much further is this taking us? How much of a leap forward is this using your technology?

11:24That's a great question, and this is something that we are going to be publishing again further, enhancing what our findings are. So that specific study where we showed with the pharma company Borenger-Engelheim how exactly our performance matched with the benchmarks. We also released our own independent case study where we showed that we have outperformed the 12 standards that exist in the industry. So essentially we are outperforming by 1000x in terms of the amount of material that today you need to capture this kind of information, the speed at which you are able to capture information, and the amount of information you can capture.

12:09So we are surpassing it by that. Now in terms of the ecosystem that we have, where there are multiple model-based companies, what we provide is that unique layer of data that can improve the performance of these models that as they exist today now there's another aspect to this which I think everyone is going to be finding quite fascinating is that some of the technology you're producing is going to be applicable to world models and a lot of people thinking about that in terms of the next version of AI after LLM's Could you expand on that? Yeah, I mean, word model is essentially about not just predicting what happens inside the silos of, or like the in silico or inside computers, but computers and machines being able to predict things that are happening outside.

12:59So, for example, if a robot has to find its path in a room, it has to understand, you know, what we call equation of motion. Like, if I throw a ball from here, where would it land? And the exact same kind of problem actually appears when you're trying to design materials, because instead of moving materials around, you want to understand if I increase the temperature, if I increase the pressure, would this material, would this ice remain still ice or would it become water? So there are these kind of fundamental behavioral aspects of material data that are missing, and world model needs the what we call a ground truth like okay when I take an action what do I measure in the real world like as humans we use our senses to do exactly that if you taste a you take a sip of sauce or a wine you know exactly what where I might tweak something and then you can make that change and understand and make a new cocktail for example how do how would robots or how would AI do that and that's fundamental real-world data that this basically provides in a loop and that's that's that's what behavior is ultimately I mean biology's answer to understand behavior was touch taste and smell and that's what we bring to machines so where do you go next you've raised this 36 million dollars what's the next six months a year look like for you one of the biggest thing that we're working on is scaling the platform and we wanted to prove the technology we wanted to prove it the early partners we've done that and now the scaling of the platform also includes like the amount of data we can generate the different kind of data that we can generate not just antibodies but like more more kind of molecules and not just being able to do that in London but expanding across the world so that's what the series they really unlock for us from a scaling perspective and the rest the next axis is going deeper so the kind of data that we are generating how much more richness can we have and the more number of features we can generate.

15:00So Vive is the first product and we have some exciting products in the pipeline that will be coming out soon. Well, it's absolutely fascinating to talk to you. I'm sure we'll see you again and I'm so excited to see this very, very frontier technology being produced out of the UK and out of London as well, I gather. But thank you so much for appearing on Path Founders, Shamit and Shika from Pohar. Music, very nice to meet you. Thank you.