#423 Teaching AI to Predict the Future of Your Brain: Dr Hannah Madan from Prima Mente

29:18Yeah, or it's too hard, you need too much data, too expensive. So it wasn't accessible for academic study super easily. You need to do high depth sequencing to find this information. so we kind of were a very we still are like hope to be a well funded kind of academic research lab in some senses because this is science that should be happening it just needs a lot of capital so it has to be a start up

30:00that was I guess our biological hypothesis the technical side was that transformer architecture foundation models were just becoming to be very very exciting and used a lot in the space we knew that they could study sequences of information super well so words and books and text the whole web so chat gbt was just kind of coming out then or like the the actual good models were on the verge of becoming around the world. So we thought the cell-free DNA information is a series of ACTGs with a methylation token in it as well. How do we, yeah, let's test if transformers can predict the next token in this kind of data and therefore by like proxy basically diagnose disease better do they have an understanding of this kind of data can we create novel architecture that would suit this problem and keep it really locked into the clinical application side we were seeing that lots of people were just throwing this data into these kinds of models but not very cleverly or not really understanding the biology or the application and it was a bit of a waste of time so we wanted to make sure that was all tied to the patient outcome and that was very the very beginning so one modality one kind of model let's see what we can find um and we could offer diagnostics for uh patients as that was the very kind of beginning business model it's now super evolved into novel biology drug targets kind of care pathways but that was the the initial yeah the best b2c or direct to hospital systems um following the kind of cancer liquid biopsy um kind of world maybe a tempus model collecting a lot of data building distribution out across lots of hospitals or maybe yeah just licensing that to Quest Diagnostics and something like that.

32:15But we, yeah, we knew that we had to, that scale was a real thing. So compute was necessary. We needed to make bigger models to see if there was any scaling laws with this kind of data. We actually won two UK Innovate grants in the beginning and got some initial VC money just to do that, those parts.

32:52Yeah. Yeah.

33:11Thank you.

33:44Yeah, so we do have a patent on this now. Like a year ago, we managed to file that. Recently, we published our model Pleiades, which is a foundation model of epigenetic disease biology and neuro. And that shows that we can diagnose Alzheimer's from a specific cohort, but we still need to validate on much more cohorts of patients. but it's more sensitive than existing protein biomarkers alone. So when we combined it with P-TAL217, we increased the sensitivity of the detection of the diagnosis, I guess. Or maybe I'm not saying that right, but the diagnosis accuracy was improved.

34:37I think that's what we're working on next. Earlier is an interesting word, I think. The samples we had now were from patients already diagnosed, so we needed to know those labels to train the models. So what we're doing now is we're running a study called Sandbox, where patients who turn up with dementia symptoms and have not yet been diagnosed, we sample from them and run those samples through our models and feed that information back to patients. So we're trying to eke earlier and earlier. I'd love to get to, yeah, age like 35 earlier, maybe. Yeah.

35:34Yes. «una I gente condensate.«

36:03Tom, lecil est buscar v� buscar v�

36:09Thank you.

36:51Mm-hmm. Mm-hmm.

37:21yeah this is really hard um it's been really hard for lots of people lots of companies so we do take this quite seriously it's the thing that me and rabbi spend most of our days thinking about to be honest no yeah yeah yeah yeah we're still really bullish that novel biology is what's really important for the for the patients and for the field to evolve um yeah new new targets for drugs we think so let me break this down might take some time but you can unravel this with me um we think in the future five to ten years maybe sooner but once you get a target that's interesting there are loads of models now that can build predict what drugs or molecules might be a great drug for that target that will speed yeah so like receptor piece of biology whether you're interested in modulating we may be able to do that in-house but much quicker it may be we think it's going to be much cheaper to develop drugs once you know what you're supposed to target actually so we we kind of obviously pharma exists and they they can do that very well or very poorly depending on what like drug we're looking at but there's lots of startups working in that space so we think that problem will be solved actually in the future we think in especially in diseases like neuro this there's going to be it's going to be cheaper to generate lots of data compute will become slightly cheaper once we've got out of this bottleneck of compute there's going to be more of these models innovating they're going to be models and gpus will be better at using complex ai architecture um so that this will become easier as well but actually finding the most clinically actionable target is still going to be super difficult because not everyone has that clinical data that longitudinal data from patients so what we do is we collect a lot of a lot of that data we develop our own studies purposefully I've learned that you don't wait around to to create your studies that you want to use to validate the thing that you're making to sell it so we just start that super early we have it all up to ISO standard it's all ready to go at accreditation level so we can we basically build a lot of interesting data we build models we could sell the data we could sell the models we could sell the care pathway that we generate in the end so we're basically trying to create as many opportunities as we can along this way to develop this novel biology and then hopefully we arrive in this wonderful world where care is actually way more decentralized maybe you're having it at home maybe there's a robot delivering it maybe you're running models for your own healthcare on your phone on your own gpus or your own mini data center maybe you have your own nuclear energy storage in your own house i don't know that's very crazy yeah exactly so yeah yeah so we're trying to imagine that world in the future and how can we I guess sell the certainty of knowing what's going to happen to your brain in the future so I think that's a big conversation that I'd love anybody's help with we would love to talk more about that we have some things in in in the pipeline that we're experimenting with right Basically, this phase of our company is the next three to six months, we need to figure out where that revenue is going to come from.

41:17Is it important? Will we get distracted with it?

41:23Or are we more likely the type of company to raise? We need to raise a lot of money, almost like those self-driving car companies that just need to collect a lot of data. we're collecting a lot of multi-omic, multi-modal data because we know that models are much better those models that we're using, the foundation models that we're building the reinforcement learning that we can include in there will just benefit from extra context, extra clinical information and we just trust that something will come out at the end we're not exactly sure what that looks like so yeah there's a lot of things to break down in there maybe James I don't know if I make any sense let me know

42:15yes yes

42:29Thank you.

42:59I mean, we've done incredibly well so far. We have access to some really exciting longitudinal data sets through connections, through Ravi, having been in this industry for 10 years, studying and also thinking about this in huge detail. And generally, people are very excited about what our models can do so that we haven't found any resistance to partnering in that way, especially when we love to be super collaborative. if we don't win in this space somebody else should so we actually share a lot of the data that we generate back with people because we you're right if we can't execute fast enough we shouldn't have the right to hold on to the data basically and create that mode so we do have no i don't think so i think that's a slight twist on yeah us increasing the pressure on ourselves to execute quickly um it is i mean yeah well i think it's valuable for everyone to see what this data looks like and be brought along in this journey actually um we'd love yeah for we love those collaborations and people like working with us for those kinds of reasons because we're thinking really about the future we're not going in incremental steps we want to delight everybody involved

44:40yeah very interesting i think we're processing our longitudinal data now So hopefully sometime in the next year we might be able to uncover some interesting things. The data is super rich across lots of multi-EMIC information. But again, as science goes, it's super hard to know for sure. But we'll have all of our lab set up ready to validate those things as well.

45:36so we were 10 people at the beginning of the year we're now 20 people likely will be 25 30 by the end of the year or early next year so me I'm in charge of like the culture the growth of everyone individually we want everyone to be on the trajectory to being the 0.1 percent or 0.01 percent of their field that's the kind of insane growth like pace that we want everyone individually to have but also as a company so we're just recently kind of remodeling our structures our meeting touch points because all of these individual areas clinical operations partnerships wet lab ai have doubled in size in the last three months so sharing information is now a big a big thing for me to make sure that we yeah democratize information decentralized decision making and really allow everyone to do the job they were hired to do right and to have that intellectual freedom to run so that's very top of mind and then yeah just continue to build the the awesome rich data sets that we have and collaborations around the world where people have already created the data that we want and then yeah thinking how do we imagine the future of care and make sure we have this flywheel in in place for when our models are ready to be deployed to the clinic essentially so um talking to more government things like that Yes.

47:15We've noticed that there is on the kind of AI side where compute will become, I guess, like, sovereign AI plays, I guess, like the UK has its own resources for compute. India does the UAE it does so how does how if we have in the future if our models are running on compute every patient every time a patient has seen inference will run compute will be used so we need to have super super alignment with that and then generally just dementia Alzheimer's neurodegenerative diseases are usually a big priority for the government so our broad vision should be exciting enough for them to strategically align with us basically

48:14yeah we hope so the first grant we got was from a like a the dementia goals mission which was run by the the uk government so um we know that they like what we're doing at that form and obviously the nhs study that we're running now hopefully um will prove that the money they gave us in the beginning has has led to some exciting changes so yeah just continuing that conversation

48:48Mm-hmm.

49:14yeah

49:27yeah

49:31we take growth very seriously I think if we we do that well um yeah there will be people being poached being like leave leaving to do their own thing um but for me it really comes back to this why I show up every day is I want to spend my time doing something super meaningful and I know that chain has changed for me personally over time of what environment I want to be in um what kind of things I need to be learning so I think it's just realistic to expect that that will happen at some point and what a wonderful thing to happen rather than someone saying I'm leaving because you're a manager it's more i've left because i've now got the skills to do something on my own and we really really need more people working in this problem heavy space that have the right mindset so we're looking for a wine this is an ecosystem play i guess in some ways that we want to create more people who think like us um and to yeah improve the healthcare space use these tools that we have to do that and really change the world but also I think it just makes it a really exciting place for people to come to work and be super motivated it helps us select for the right kind of individuals that have the right mindset to do the the really tough hard things that we want to do we're not just we're not okay with just intent to do something we want people to go and do it So I think that's where the founder kind of mentality comes in.

51:06It doesn't have to be everybody in the team at all, by no means, but Ravi and I have great networks now in this space that we'd love to leverage to support everybody in the team. Yeah, we see it as a really positive thing and a great outcome.

51:40Yes.

51:46Oh yes, of course. Yeah, yeah, yeah.

51:58definitely I think you are shaped by your experiences I think I'm very reflective I journal a lot I will try to improve iteratively so yeah I think it's like I said I have a lot of check-ins with everyone individually so I do get to know everybody well one of the things I test myself on is can I like I like to make draw birthday cards for everybody on the team do I know them well enough to make a very personalized kind of card for everyone individually um so I am trying to align the company and the individual together in a nice way so yeah we don't necessarily expect everyone to to yeah give away and take on the risk that founders do but um uh yeah it's definitely basically like I think my journey through startups I'm very biased because that's most of what I experienced but you experience such extreme growth and such learnings that it's very it brings a lot of proud feelings and joy and I love seeing that in every one of our team so I just wanted to nurture that and I think time has shown that it does really motivate everyone individually because they get they're going to get that hard feedback that they need they're going to get judged what we judge people at their next level, right?

53:19So they're going to just not waste their time personally, career development wise and for us.

53:47Yeah.

54:39Yeah.

54:52yeah I think one it comes very naturally to me so like I don't see it necessarily it's like a huge amount of work um yeah yeah um exactly so um it's just something that comes with me I guess I want to care about the individual I think the feeling you get when somebody trusts you and rewards you and values you it can unlock massive creativity and this world we're in needs massive creativity the space is also moving incredibly fast I cannot keep up with what's going on in my own company let alone the rest of the world so um with that context in mind we all need to grow in the same kind of speed because competition is increasing every day.

55:43Execution, like we called out, is the only thing really that will save us. So we need people to grow for the business to succeed ultimately. It's how I, yeah.

56:06Thank you.

56:50Yeah, I don't think it's... Yeah. Yes. Yeah. Yeah. recently I have made space um I was always inspired I met a few people last year who were saying I'd spend like 10 % of my week optimizing what tools I use and things like that so we've kind of made it made it a collective we have a tools and applications channel in Slack where people share the AI tools they're trying so people can learn collectively I am trialing this week now chat gpt can do tasks for you in the background i've said it to survey the ecosystem in various ways every week i've said it to prompt me for certain reflection questions so it's yeah using it's really hard but you have yeah i'm trying to think about i use a the tool called co-founder um that connects my just came out a few months ago that connects like slack gmail notion everything together so i can be like let make me a standing meeting for this at this time and it will just do it for me um so trying i think i have some newsletters as well that i've refined i think it's ben's bites or something is one of my the top one that gives me tools plus news and so I just cancelled all the rest of them because I was feeling very overwhelmed like you I think that's very easy to do so yeah just tried to reduce the number of newsletters I take in we already have a lot of people posting things in our slack and so just learning from the team um I would love to do like yeah some if anybody knows there probably is an integration but all of the things I say for later in my Slack I want notebook LM or something like that to make me a podcast for my walk home so I can figure that out on the way that would be great and that's my next thing to think about how to optimise but yeah, I would love please reach out if you have any other tips it's something I've just decided I'm going to have to come to terms with I'm not going to know everything

59:23a slither of like our the one thing i haven't mentioned is our team is now across san francisco as well and marie our chief of staff that's part of her role to ai fi everything um in our team

59:49Thank you.

1:00:25it's a really interesting question we spend a few we have like a few i guess like truths that we want to be true i think um that basically healthcare will be super disintermediated or of decentralized um so lots more care where wherever the patients are um uh with the i guess the doctor will become more of the system plus the data plus everything um and that will look yeah like a mix of ai a mix of more frequent contact points because hopefully what we'll find at specific points when your biology changes all the time but healthcare is a one very discrete points of time so how do we build this monitoring or longitudinal aspect into it more realistically we also think that monitoring people with one or two biomarkers is not enough and it has always been a weird, it's not connected to the data-driven biology that we can now do.

1:01:40There are way more things we can monitor all at once. So how do we push for that to change? I know it's hard to validate those lots of markers at once and that's why we usually do between one and seven biomarkers for any kind of test or monitoring. So it's very difficult. We would love to see more data points being used to make decisions more rapidly and at the point of care which actually makes sense for the patient so for example with the brain you lose your neurons and they do not necessarily come back unless somebody's going to invent something magic for that which maybe they will but we need to act way earlier than we do now as we alluded to earlier so how do we yeah use diagnostics and prognostics more fully i guess as a way to to really help patients yeah get the right treatment at the right time so shunting everything earlier and earlier i guess to more monitoring and management and prediction as i think we've all been trying to do for decades but i think the this collecting the right data and feeding that back and creating that that circle of feedback around the patient and with the doctor in tandem will really help.

1:03:24Thank you.

1:03:58Yeah. Yeah.

1:04:24Thank you.

1:04:55Thank you.

1:05:31Exactly, that's where the novel biology angle from us comes in. It's absolutely pointless to get a diagnosis without actually understanding the biology and what you can do and a drug you could take or something like that.

1:05:52I am excited about our clinical study that's just launched in the NHS because again like you're just saying we want to be closer to patients get that feedback see how they're they're like interacting with the stuff that we're building for me that's yeah critical to bring back into the team that feedback and that real love working with people that can in the team that can see the impact so clearly so very excited about that and the growing team the people that are joining in November December and also in the next years coming up I'm already excited and very proud of what they're going to do so

1:06:37yes we'd love that

1:06:53I know

1:06:59yeah exactly yeah I'm making like paper sharks birthday cards just a full-on craft store right now

1:07:17yeah that'd be wonderful

1:07:26yeah both of those yeah just primamente.com there's a roles or open positions um it has my email as well as my yeah all of the roles that we have so that's a good place but LinkedIn also works super well so yeah looking forward to meeting everybody thanks James hey everyone thanks for listening and making it all the way to the end of this episode remember to subscribe rate us and leave a review and you can head to the description of this episode to follow me on all of my social media so you don't miss out on any of the latest health tech content.