Unlocking the Power of Generative AI for Scientific Literature Review

I'd like to thank everybody who's joining us for our webinar today. I will be starting in about a moment, giving an opportunity for everyone who wanted to understand and unlock the power of generative AI for science lit review. We'll be starting in about a minute or so. I'm going to go ahead and share my screen. Okay. Good morning, everyone. I want to welcome you to our webinar, Unlocking the Power of Generative AI for Scientific

Literature Review. We have two wonderful speakers here with us today. First, I'll introduce myself. My name is Anne Licata, and I'm the VP of Customer Success at Huma AI. I'm joined by Heather Desmarais, President of HJD Consulting, and Lana Fang, Dr. Lana Fang, CEO and co-founder of Huma AI. Heather, would you like to introduce yourself? Sure. Thank you. Yes, it's actually Heather Desmarais, but nobody pronounces that correctly. That's okay. I have been in the life science industry for about 15 years. I started out at Cubist Pharmaceuticals, where I managed the knowledge services group.

And when we were acquired by Merck, I set out on my own. And in the last eight years as a knowledge and information scientist, I've been working with over 40 companies, helping them to get the information that they need to do their jobs. Great. Thank you, Heather. And welcome, Lana. Hi, Anne. Thank you so much for moderating this exciting discussion. And thank you, Heather, for joining us. My name is Lana Fang. I'm the co-founder and CEO of Huma AI. Spent pretty much my career over 20 years in pharma, particularly precision medicine, and kind of founding Huma AI. So I've been working with Huma AI in 2018.

So we've been around for several years, really kind of see can we use AI and particularly natural language processing to automate the massive amount of unstructured data, such as publications, where experts like myself had to manually curate. So, and of course, generative AI, we've been working with open AI for over a year, actually, really kind of transitioning into the next level, which is generative AI, applying it to some really impactful life science use cases, including publications. So look forward to this, my favorite topic. Great. Thank you, Lana. To our audience, I want to encourage you that if you have any questions,

feel free to utilize the Q&A section. You see the little button on your panel there. Feel free to type in any questions. We'll be getting having a wonderful panel discussion and question and answer opportunity for our audience coming very soon. Let me start with the exponential growth of medical knowledge poses major challenges for life sciences professionals. This is coupled with the swift evolution of the medical field with emerging research, treatments, guidelines, and technologies constantly reshaping the knowledge landscape. Assessing the quality and relevance of such an abundant information is crucial, as studies

may differ in validity due to limitations and biases. Moreover, professionals must understand interdisciplinary developments, manage complex data sets, and disseminate relevant information effectively. To overcome these hurdles, life science professionals are encouraged to continuously hone their skills, leverage AI and machine learning to sift through and analyze data, and collaborate across professions to maximize the utilization of accurate and relevant information. Medical knowledge is growing at an unprecedented rate. We are now in the age of the digital age. We are now in the age of the digital age.

We are now in the age of the digital age. We are now in the age of the digital age. We are now in the age of the digital age. We are now in the age of the digital age. We are now in the age of the digital age. For instance, PubMed, a database of life sciences and biomedical literature, is adding an average of over 1 million articles per year. The sheer volume of this information can be overwhelming for informational professionals who need to keep up to date with the latest findings and development. Heather, I'd like to turn it over to you, as you have an interesting presentation to share about this topic.

Great, thank you. Next slide here. Yeah, so as a knowledge and information management consultant, I've seen the issues and the pain points that life science companies run into, and one such issue is dealing with literature. So just as Lynn said, in the biomedical field alone, more than 1 million papers pour into PubMed database each year. That's about two papers per minute, and that citation is actually from 2016. So next slide here. There's a vast amount of literature produced each year.

Worldwide, STEM publication output continues to grow on average of nearly 4% each year. From 2008 to 2018, output grew from 1.8 million to 2.6 million articles. So we're probably close to about 3 million articles a year. So the next slide, please. So you can see on this graph, just citations added to MEDLINE by fiscal year. Look at those last two years. It's increased tremendously. I don't know if that's COVID related or just the nature of literature. I don't know, but yeah, next literature.

Next slide here. Yeah. So within literature. Within life science companies, there are many consumers of literature across the organization. Literature plays a vital role in shaping scientific research by providing the foundation of knowledge, guiding methodology, validating results, and fostering collaboration among scientists. So the R&D group is reading articles for background knowledge to explore previous findings and methodologies and to build a foundation for their own research. Pharmacovigilance group monitors the literature for safety concerns with their weekly literature alerts.

The medical affairs group uses literature to identify KOLs. The field team references literature when they're meeting with health. The field team references literature when they're meeting with health care providers. Medinfo, they're referencing the literature when they respond to inquiries from health care providers. The regulatory group references literature in their FDA submissions. Legal uses literature to write patents and defend them as well. The competitive intelligence group obviously is keeping an eye on their competitors through the literature. And medical writing. And medical writing is referencing the literature when they write new papers.

So it's crucial to be familiar with the most up-to-date information so that organizations can make decisions in a timely manner. But finding and digesting or making sense of the literature is a time-consuming process. I've read estimates that scientists are spending 11 hours a week just on, you know, trying to stay on top of the literature. And they're tracking literature. it from various sources. So they may have alerts set up in literature databases or directly from specific journals of interest, RSS feeds, social media platforms like Twitter, Facebook, LinkedIn, as well as from just email distribution lists as well. Streamlining access to information

and resources enhances productivity. A literature management tool can be a big help in getting your organization's literature centralized and organized. And some have incorporated various AI features such as recommendations based on your literature viewing or identifying terms or concepts within a paper. And some of my clients are even using various methods to help them with analyzing. So I'm going to talk about some of the tools that are used in the literature. So text mining and natural language processing can be used to extract and analyze textual data from research articles.

This involves processes like text parsing, entity recognition, and topic modeling. By applying these techniques to a large corpus of research articles, you can identify keywords, topics, and emerging themes within the literature. Now, these can be used to identify latent topics within a collection of documents by analyzing the co-occurrence of words and phrases. These algorithms group related documents together and provide insights into emerging research trends. Analyzing citation, nope, still back. Sorry. Sorry about that. No problem. Analyzing citation patterns can help identify influential papers and detect certain

differences in the literature. So I'm going to stop here. Thank you. emerging trends. And by examining the citation network, one can identify clusters of papers that receive a high number of citations, indicating that their impact on the research community is high. So this can help identify the most influential research and the direction that the field is moving. Research communities often have strong social connections and social network analysis can be used to understand the connections and detect emerging trends by analyzing co-authorship networks, collaboration partners and information flow within the scientific community. One can identify key players and communities driving

emerging research trends. Effective data visualization techniques can help researchers gain insight from the vast amount of research evidence that is available on the platform. data and research literature. Visualization tools like interactive graphs, heat maps, and word clouds can help identify clusters of related research, highlight key topics, and track the evolution of research trends over time. Machine learning algorithms can be used to build predictive models based on historical data, enabling the detection of emerging trends. By training models on past research data and associated metadata, one can predict future

research trends, identify potential breakthroughs, and guide research funding decisions. And continuous monitoring of the literature through automated data collection techniques allows for real-time tracking of emerging trends. By aggregating data from various sources, such as academic journals, pre-print repositories, and conference proceedings, one can analyze the latest publications and identify emerging topics as they emerge. Combining these data science techniques can help enable researchers to systematically analyze research literature, identifying emerging trends, gaining valuable insights

into the direction of scientific research. Next slide. These methods are executed through collaboration between large and small organizations. The data science techniques are used to enhance edge accuracy, help researchers, and management teams to impact your data. By working with data science experts, researchers and data scientists can help you better understand the way data aid is executed. This multi-disciplinary approach fosters a comprehensive understanding of the data, encourages innovative thinking, and maximizes the potential for breakthrough discoveries in life sciences. The goal is to enable faster research for faster discoveries and decision making.

Last slide. And with that, I'm going to turn it over to Lana. Lana, if I could, the adoption and integration of AI and language models like ChatGPT and Literature Review have gained substantial momentum. The speed at which these technologies are being embraced is astonishing, enabling researchers to enhance the efficiency and accuracy of their work like never before. I'd love to hear your thoughts on that with your presentation. Great. Thank you so much, Anne. Thank you, Heather, for a great overview on why analyzing publication actually is really important. So I'm going to kind of change gear a little bit, right, bringing this kind of AI flavor to the discussion. So ChatGPT,

right, I think it's like it's going to be really difficult to find someone here who has not had the opportunity to do that. So I'm going to turn it over to you, Heather. I've heard of ChatGPT, right? It really kind of took the world by storm, really kind of stretched our imagination, right? I think it's fair to say ChatGPT is the most rapidly adopted technology in history. I think it's probably come second or close would be the internet. But kind of just give an idea, right? How does it compare to kind of our regular life app, if you will, social media adoption? So if you look at the, you know, start from Netflix,

and some of these Airbnb, you know, Twitter, Facebook, that we kind of familiar with, right? We use them in kind of our life with our kids and what have you. And it typically takes like months to years to get to that 1 million milestone, if you will, of users. And look at ChatGPT on the bottom, five days. So they released it before the holidays last year. So it's about six, seven months now. And they already hit 1 billion users worldwide. It's really, really important. And I think it's really, really important. And I think it's really, really showing you the numbers don't lie, really is how popular and and also the sustained the

growth shows like, folks, every everyday folks, right, are using this and finding it useful, right? Because otherwise, it's going to be drop off. So. So that's why I find this topic really interesting. And also very easy to relate. Because when we're talking with our clients, and what have you, they're actually already use ChatGPT in their personal life, right? Or maybe it looking for planning vacations or trying to like remodel, and more importantly, probably more frequently because their kids are using ChatGPT, helping them with homeworks, and what have you. So they're already familiar with the ChatGPT, right? And then we're saying,

and saying, like, you know, like we're saying, can you use this leverage tech technology for enterprise? So this is kind of a very unique time in history, I think, where this consumer, a very, very, very, very good customer, right, we just saw a lot of people saying, hey, why don't consumer arena is intersecting with enterprise arena, particularly in life sciences. Next slide, please. So I want to kind of expand sort of GPT into generative AI, right? So GPT actually stands for generative pre-trained transformers. So before, so we call them collectively called large language models. And then in turn, we call them generative AI. So the difference,

we call them generative AI is because the generative AI is actually creating content using AI. So that's not originally in your data. So, and then these large language models is basically kind of the models behind that. And how is that different with the standard transformers, like BERT transformers and BioBERT? So those, and then some of the actually transformers used in life sciences, drug discovery. So those things, think about them, predate the, key on large language models. In fact, there's an interesting factoid is life sciences as an industry is actually the earliest adopter of generative AI, right? Remember I said,

generative AI is basically using AI to generate new things, right? So there's actually like molecule models and there's like self-painting, different aspects of generative AI has been used, had been used in life science and drug discovery for several years before large language models came kind of onto the scene. So this is a really, really great representation of, if you look at the generative AI landscape, right? There are many, many different models, right? Large language models primarily deal with text, with documents, but then there's actually GPT-4, for example, and some other models like Dolly can work with images.

And of course these are from open AI. And then Google is actually releasing very different, various kind of different large language models and capabilities. You have stability AI with the image. There's basically, you have the, large technology companies releasing their own large language models. You have also stability AI releasing models from open source, right? So there's kind of really complex and fast moving industry where these foundation models, as we call them, large language models are released almost on a daily basis. Next slide, please. So it's very, very exciting field. So how does that GPT model

or generative AI apply to life sciences? So, so we actually have done a survey, basically talking with 72 medical affairs leaders and themselves. So this actually happened in March. I think the, if we were to do the survey today, the results is going to be even more dramatic. So, so we asked the question, like, you know, what is, you know, we said, do you see generative AI applied to life sciences? I think we had like 65% in March saying that it is. Notice freaks? I want to know your answer to that. For more information to跑omatch.it.com. Thank you. Bye, everyone, and welcome to Scarlet Musical, nunca segunda, painful presenters.

should be applied. And we had over 60% actually of the surveyed experts said they would prefer to use a chat experience for analyzing their enterprise data. And then more importantly, which is kind of pertinent to this topic, we ask which data sources that you work with is most suitable for applying generative AI. And if you look at this data, this pie chart, 40% right off the bat said publication is the best use case. Next slide, please. Okay. So using chat GPT for your personal stuff is really, really different from trying to analyze using large language models and GPT models for your private enterprise data. So this is where

human AI kind of has been focusing our effort through the working with open AI, being able to access a large language model. And I think that's a really important part of the data. Actually, early summer of last year. So being able to really kind of take a scientific approach to I call it refine those GPT models for life sciences needs. If you think about it, there are some unique high bars for us to meet. For example, privacy and security, because we deal with a lot of patient data. We deal with healthcare data and also those data highly proprietary and the needs high levels. And then secondly, is that I need to be able to analyze my private data. And ideally, that private data should not be commingled with public data and more importantly, should not be seen by some of these large language models that are kind of in the public.

Right. And then, and thirdly, it's probably getting more and more important as you know you've probably heard the phrase hallucination from these generative AI large language models. So that's a no no for healthcare life sciences right we need that accuracy. So being able to really kind of take this expert in the loop approach, going through this. We call them reinforce learning through human feedback loop to basically increase the accuracy, there's actually a really interesting. Interesting fact that, like, you know, only kind of people in the industry knows is the reason opening ice GPT models are so much better than the other language language models because very early on.

They kind of use this reinforced learning through human feedback loop in training their models of 170 billion parameters. I mean, in fact, they hired folks from like the Philippines and Africa to basically manually going through the content. Right. And then they actually use the data from the public data that used to train the models and go through them and hand label. And in fact, and now pretty much everybody from from the big tech companies who are developing large language models are adopting this approach. Of course, for us, we don't use human experts labeling the data, but we do use human in the loop on experts QC the data and use that to refine models.

And lastly, this the ability to actually see what are the supporting document right yeah used to generate those. Those generative content and being able to cite the provide citation provide transparency is a huge must. Any questions. Great. Thanks, Lana. I'm going to stop sharing for a moment. I, it was a great both ladies, those were very compelling presentations. And I'd like to discuss the topic further before I get into some questions that I've prepared prepared for us to chat about. I'd like to encourage our audience to utilize the. Chat function, excuse me, the Q and a function on their zoom toolbar to go ahead and type in those questions.

I promise to get get to your questions as you ask them. So please feel free. Don't be shy. So if I could start out with a question. This is for Heather. Heather, how do you effectively manage the overwhelming volume of information coming in every day in the biomedical field? Are there any strategies or tools that you have found particularly effective? Yeah. So I find that it really has to start with a well honed query so that you're not dealing with just an absurd amount of information to go through. So really, you know, being specific about what you're looking for in the literature is important, which, you know, oftentimes I find that, you know, it comes back to me as a librarian to.

To assist with getting a really accurate query so that we're not pulling in all kinds of other stuff. Great. Thanks, Heather. Lana, I've got a question for you that just came in through our chat. So thank you, audience. Lana, do you have any numbers or an ROI resources, a return on that investment of the resources saved by using generative AI in literature review? No. Okay. That's a great question. I do have the data. And before I talk about that, let me kind of give an example of like, how does a generative AI solution apply to literature analysis? Right. And Heather mentioned this, right. I came from grad school. I've been in life sciences.

I've done my share of going to PubMed. You type in your search queries and return like 100 papers and you kind of read them one at a time. Right. You either kind of reduce your narrow down your search criteria so you get less papers to manually sift through or you just have to power through this. So so our solution, human AI set 1.0 using transformers prior to generative AI was to basically use transformers to actually mine the full text article. Right. And then being able to kind of surface the regions of interest, if you will. This is a kind of a pretty simple. Standard term in AI to, you know, rather than manually read 100 papers.

Now you have 100 kind of regions of interest, a relevant paragraph to your question. So that kind of transformed that experience. Right. But you still have like 100 paragraphs to read. And then experts need to kind of convert it into that higher level intelligence. So what generative AI does now and be able to take those relevant publications from 35 million articles, for example. Right. And then being able to actually synthesize it and create those higher level summaries or key takeaways. So what we do then is that how do you verify whether those key takeaways or summaries are actually accurate?

Right. So what we do is we actually then take you for that each key takeaway to the citations. What are the papers that actually support that key takeaway? So give you the ability not only to kind of validate it and also kind of see what are the key takeaways. Is. Is correct or not. Right. It's a it's that level of trust and also that level of validation and and then kind of the the able to. And then you can actually then say, OK, this key takeaway maybe it's not right. Right. And then you can actually put in the feedback. So incorporating end user into this reinforced learning human feedback loop as well.

So it's very much that generative AI create that higher level intelligence experts now just need to verify. So it's very, very different. If you don't have that level of value proposition. That is why it's transformative. Because a lot of these topics could be hundreds of articles. Right. That's why the teams typically we're working with would work with someone like Heather or with a team of experts to kind of manually go through this and also do an update on a monthly basis and create a PDF report. Right. So because that's how you narrow down that data. But using generative AI, you can actually kind of analyze out of thirty five million dollars,

thirty five million dollars and I'm going to get to the bottom of it. million publications, hundreds of these relevant publications create that higher level summary. Yeah, it's quite transformative. So going to the ROI question, yes, we actually have done thorough validation side by side between generative AI and AI in general, and then expert manual curation. So we've seen really astounding. So the time to value, we know how long it takes, right? Particularly in some areas, for example, post-market surveillance and what have you, where you have to do the systematic literature review or PV, right? So to satisfy

regulatory requirements. And then for any product, typically it takes months to go through that analysis. And then for us, it's minutes. That's the first time to value. And secondly, is scalability. You can actually put, one of our clients put a hundred products on the platform for regulatory submission. So if you think about it, if you do a manual curation, it's basically like, you know, months times a hundred, right? You either increase resources or you increase the duration, the time required to create that intelligence. That's the second thing. And the last thing is that it really, really allows your team to do more with less, right? And then we always like to say we're AI,

we're not artificial intelligence, we're augmented intelligence. We're not here to replace AI. We're not here to replace AI. We're not here to replace AI. We're not here to replace AI. We're not here to replace AI. We're not here to replace AI. We're not here to replace experts. We're here to augment experts, right? In fact, if you look at chat GPT, you probably ever run into this experience that the better questions you ask, the better answers you get. We see very much this generative AI is an expert-centric AI solution, right? It's centered around experts. That's how we built our generative AI platform. This is how generative AI should be used in sort

of real serious applications. So here we basically say we have the data showing that's equivalent to 15 to 20 experts manually curating that data. I hope that answers your question. It's a long-winded answer. Lana, thank you for that very comprehensive question. And because another question came in, you're getting a thumbs up now on that, on your response. And because another question came in that would be a really great follow-on to that question, I'm going to go ahead to our audience question and ask you, Lana, do you know of any existing AI software? With the ability to generate a literature search protocol and report per MDR and IVDR requirements?

That is a great question. So the way we see this is that report creation is a highly expert kind of a driven endeavor. And it's really complex. And we haven't seen a generative AI being able to create a report for submission. In fact, you probably don't want generative AI to create that final report. So what we've done is actually using generative AI to create that first pass, if you will, particularly doing the kind of initial, either the systematic literature review. And in fact, we're also doing really outstanding work, shout out to our team, being able to do state-of-the-art appraisal using AI,

NLP, and generative AI, right? You know, like give a product code, you can actually traverse, multiple data sources, find out who your competitors are. And then based on the 10 criteria, going into the literature review, going into the publications, going into maybe other great literature review, like OnBase, to be able to find automatically who your competitors are, and how do you and your competitor measure against those criteria. But then no, we don't, I haven't seen an AI solution be able to generate that report. But the industry now is very much moving into particularly, you know, the data science, the data science, the data science, the data science,

the data science, the data science, the data science, the data science, the data science, the data science, in life sciences. You see generative AI is applied to maybe creating marketing materials on the commercial side, but it's still, you need that expert, kind of that final sign-off approval. So I don't recommend using generative AI to create that final report. You could do a draft, but you need expert to do that final review and sign-off. I hope that makes sense. Great. Thanks, Lana. I'm going to switch it up a little bit and come over to Heather. And Heather, how do you effectively manage the overwhelming volume,

the volume of information coming in every day? Excuse me. That was the repeat of my question. My apologies. With PubMed adding over a million articles each year, how can we ensure crucial, actionable insights are not lost in the flood of new information? Yeah. So I think that it comes back to that really honed query. Once you have the proper search strategy, you can then set up alerts in your literature databases so that you're notified anytime new content is added to the database that's relevant to your search. You can then review them in the emails generated by the literature database or pull them into

SharePoint via RSS feeds or pull them into a literature management tool where you can then capture the review process. Yeah. Great. Thank you, Heather. And on that, what implications does this information overload on decision? What are the implications, rather, on the information overload on decision-making processes and life sciences? How can professionals maintain the quality of their decisions in such an information-dense environment? Is that to me as well? Yes. To you, Heather. Yeah. So I think, that you know just having having uh you know the tools that will help them to uh review

the literature more efficiently um will cut down on you know the time that it takes um and yeah just it sounds like um ai can really help with just decreasing the amount of of information that they would need to go through to to come up with the answers that they're looking for great thank you lana you mentioned that you know in your role uh in life sciences you had a flood of information coming at you how you know how do you see ai helping to um to keep control of the flood so that you're not getting drowned i think this is actually a great

use case for ai particularly generative ai right and then and the ai actually can understand my interests like heather said once i have a really awesome query i just need to be notified right so this is my area of interest so ai can actually know your area's interest and then rather than sending you all the relevant information for you to curate and synthesize it actually can do a first step and then send you all the relevant information for you to curate and synthesize it actually can do a first step first pass of sifting and summarization and synthesis and then give you that kind of and then provide like a you know a citation or reference for you to go back and trace and

uncheck so that's what what i see is and is is is kind of the the angle if you will across data sources by the way great thank you for that you know um this question is actually for both of you how can we use data science to detect emerging trends in research literature what methods are being developed to um to allow for predictive analysis in this area heather do you want to start yeah sure so i think that one of my slides actually addressed um you know some of the different approaches that you know there's text mining and natural language processing topic modeling algorithms um analyzing the citations

and the literature and the literature and the literature and the literature and the literature um so that you can find what's been um cited more so that you know those are the more heavily weighted let's say um articles um you know data visualization obviously trying to pull everything in and presented in a format that's um that's uh more of a visual representation of the literature um yeah great bonnet would you do you have anything to add to that um i think that's a great question you have anything you'd like to add yes um this is why i'm so excited about generative ai it's you know heather presented that slide with all these different things that we can do for

literature analysis providing insights right to help us with our research and what have you and um you know you could do you know knowledge graph right you could do like some kind of sentiment prediction you can do like you know um sentiment analysis and what have you all these different just on the nlp side right and then we mentioned about manually tagging and annotation and indexing right so um the reason i'm so excited about generative ai is that there are so many functionalities that generative ai can do the large language models can do we haven't probably we haven't even touched yet so if you look at the hype cycle from

gardner that was published a year ago they had this hype cycle for natural language technology very clearly mapped out natural language understanding right knowledge graph and sentiment analysis yada yada for kind of and you know analyzing text data right could be publications what have you so they're all in distinct spots kind of within that hype cycle but then now if you look at large language models actually they can do everything right so the fundamentally changed how we um analyze unstructured data so for example in our hands we could do you know in addition to kind of that high level generative content we can do sentiment

analysis and we can analyze really really complex um data based on very complex business rules we could do knowledge graph so it's basically really i am just so excited that we have probably even only touched the tip of the iceberg great thank you lana i'm going to go to an audience question now um lana this is for you does human ai search databases to provide necessary references to download or does it directly help in extracting them um where does it sit in within the search screen extract report cycle that's a great question so we actually do all the above so if you think about search in a very general term right beyond sort of the keywords into kind of um

sentiment right sentiment not sentiment um semantic search and you can tell that i'm not a techie into semantic search using transformers right so that's kind of your first step out of the 35 million articles and you come up with the relevant publications you're going to be able to look at the top of the list of journal articles you can see the top of the list of journal articles and also the top of the list of journals that are outlined in the title of a journal publication remember i mentioned reasons of interest and then you put generative ai on top to provide that kind of that summary the key takeaways and then you provide the citations that's what we do provide

citations to each of the generated content so this is kind of the cycle that we created so we found in our hands is that you know like l m's large language model is really not the end game it's a advancements in different kind of the platform we've already built prior to generative AI combined with L&Ms is kind of that general picture is the summation of all pieces that could provide that high accuracy, be able to provide citation. We're really proud. I mean, the work we've done, we're the only one on the market who can actually provide true citation on the generated content, right? Because the summaries are not really in your original

publication. It's the summarization of all that key insights collection of the hundreds of those publications that's relevant to your favorite search, if you will, in Hedder's terms. Great. Thank you, Lana. I have a, excuse me, I have a question for both of you. How does the complexity of big data and life sciences affect the speed of information delivery? What techniques can be used to improve the speed of information delivery? And if so, how can we improve the speed of data analysis? Hedder, would you like to start? I will start. That is such a great question, but what really, really general, maybe I can try my

best, right? I came from pharma, right? I know it takes like 10 years and 2.6. I think now it's a, Deloitte released a report. It's like, now it's down to 2.3 billion. We cut 300 million of the, of the price tag. It's still, it's really, really expensive and takes a long time to bring a new therapeutic to market, right? And in the meantime, patients are waiting. So part of this really is because of the lack of data automation. And the reason we lack data automation in life sciences, because first data is highly complex, right? Big data. And secondly, data is siloed in these like different, different data sources, right? Publication, for example,

I think Hedder, you mentioned you have to source from all these different sources. And then thirdly, 80% of that data is, we call, unstructured, right? They're not nicely laid out in tables, in numbers. They're in, you know, documents, right? You have PDF, you have like Word and you have PowerPoint, God forbid. And, and also, you know, even free text fields in tabular data set, right? So it really kind of limit the type of information we can get. And I like to say like, you know, in pharma life sciences, we don't have a date. We don't, we don't, we have too much data. We have enough data, but we can't get, put that to good use because we can't get insights. So really,

I think that's where the bottleneck is. And publication is a great example of doing that. And I think we don't have a magic bullet to solve all the problems, but we start with the use cases, right? What are the impactful use cases? And publication is one of them where we apply generative AI. It's very much this, in fact, you know, cause we, we, we work with other AI companies. We talk with other AI founders. As soon as they hear like, oh, you're generally for life sciences, man, you're legit. You're, you're targeting like really the hardest stuff. And, and like, we, we actually, you know, have a second mention by Gartner about kind of where the leader in applying generative AI for

life sciences domain, which is really, really difficult. I'm glad we started the year ago. Yeah. It took a lot of hard work. Great. Thank you, Lana. Hey, Heather, did you have anything you wanted to add to Lana's comments? Yeah. Well, to her point, I mean, it's just such a, an undertaking to kind of, you know, process all of the data out there and it requires cooperation across the organization and multidisciplinary collaboration. And for a lot of the smaller companies that I work with, you know, they don't maybe have all of the staff that's required for this. So, you know, it can be a bit much for smaller companies to handle.

Yeah, I bet. I bet. Okay. So we've talked a little bit about a number of different ways of, of how AI can help and such, but I will also want to ask this question. How can we balance the need for speed and disseminating information with the need for thoroughness and accuracy? This is for both of you. So who would like to jump in first? Heather, you want to jump first? Yeah. I mean, I think it, it, the goal is just to be efficient. You know, making the process as efficient as possible and, and as accurate as possible. So whatever means to get to that point is, you know, it's, it's a large undertaking, but I think that

getting, to the point where, you know, the data can be synthesized and distilled into useful information is worth it in the end. Great. Lana, do you have any comments? Anne, would you mind repeating the question? Not at all. How can we balance the need for speed and disseminating information with the need for thoroughness and accuracy? I think thoroughness and accuracy always come first, right? It's because it's, exactly. So you can't sacrifice accuracy, right, for speed. It just, you know, it's, I don't think it's even, it's a non-starter for us, for our industry, right? Yeah, I'd agree with that statement.

Thank you, Lana. Lana, I have another question for you. You've talked a whole lot about research. I'm curious, what are some of the things that you've learned from research that have been really helpful in the field of artificial intelligence and machine learning? Regions of interest and sifting through the data and things. So this kind of like goes back to that thought here. In what ways can artificial intelligence and machine learning be leveraged to sift through all and prioritize the information for the researchers and clinicians? That's actually a great question. So if you think about what humans do really well and what machines

do really well, humans were very, very good at this causality, this linear thinking, right? So basically, drawing, kind of logic conclusions, if you will. And what we don't do well is massively parallel process, parallel process, a lot of data. And then what machines do really well is massively process a lot of data, right? So, you know, it can sift through $35 million, 35 million publications and then, create hundreds of these. And then you can do it at ease once you have that platform built. And then what an AI can really do well, these days, at least in our hands, is to ingest very complex life science business rules, right?

So once you set up the business rule, the machine knows, okay, this is the protocol recipe I should follow. And then you can just crank it through, crank it through, right? You know, Heather, you mentioned about 3 million publications a year, right? Was it two papers a day? Right? It doesn't really care. A minute. Yeah, a minute. Yes, exactly. Amazing. It doesn't care. It just cranked through. Yeah. Terrific. Lana, I want to go back and this is kind of like combining a question from the audience with another thought. And it's really around the ability that you mentioned of human AI to come up with the citations. Could you speak to the ability of human AI citing the content and

how that differs from other sources that are out there? Okay, so if you look at like Google and Bing, right? Yeah. So what they do is that they actually come up with the search results first, because they're a search engine. And then they create a generative AI and correlate it with the pre-existing search results, right? But then the generative AI that we do is actually quite different because you have the sort of the generated content, which is actually really not in your original content, right? It could be, it's a summarization, if you understand what I'm saying. So basically, it's already gone through one, two, three, four, five, six, seven, eight, nine, 10, 10, 11, 12, 13,

14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 38, 40, 41, 42, 45, 42, 43, 45, 45, 46, 46, 47, 48, 42, 45, 46, 49, 40, 42, 45, 47, 48, 50, 51, 52, 51, 52, 51, 52, 53, 52, 53, 54, 53, 55, So then what you have to do is actually take that generative content and saying what is a correlation with your underlying data? So that's another level of transformation. So in fact, I think I saw the comment there, because we came out with this Q4 of last year, right? It's been a while. I'm actually really happy to see some startups actually also working on this, because it's really a critical

must have for utilizing generative AI in life sciences. So right now we're doing it, actually being able to do like, you know, multiple levels of transformation. But in fact, when we presented this to OpenAI, that when we launched this, basically they were like silent for 10 seconds. And they were like, this is really hard to do, right? Being able to have that and also accurate. So kudos to whoever, I think there was a comment saying another company that does it, yeah, kudos to. To this other startup that's also being able to accomplish this. Great, thank you, Lana. Heather, question for you. Can you share any examples

where rapidly evolving information has significantly altered your approach or the approach of your clients' organizations? I see a good number of my clients trying to, you know, pull in APIs and to adapt. You know, we were working at home and we had to ramp up the amount of time that we needed for them. But a lot of the time, people were trying to do whatever they could without the help of anyone and the really important thing that we're trying to do, which is basically, you know, you're going to have a fast forward to, you know, create these up-to-the-minute visual representations of all the information.

As you can imagine, that's, you know, quite an undertaking and requires quite a team. But, but I think that it's really beneficial to the organization too, you know, because there's so many disparate sources of information. Great. Thank you, Heather. Lana, back to you. You mentioned you can, and this is from the audience, you mentioned you can monitor products and its competitors using the Huma AI solution. Does this mean that the search is done based on a product and not on other parameters like diseases, treatments, or symptoms? Oh, this is the beauty of an AI solution. You can put as many parameters as you can.

We can do basically a business rule that is like three pages long. It doesn't really matter. So it really depends on what you want to kind of search, quote, unquote, analyze. Terrific. Thank you for that. Heather, in light of the constant flow of new medical information, what are the biggest challenges you foresee in the future of information management and life sciences? And how might these be mitigated? Well, I think that, you know, the data science team, I think that librarians will be needing to really work with them to hone, well, to select resources, because that's where all the information is coming from,

as well as, you know, honing those queries for what you're looking for. Yeah, I think that, you know, these are the things that are going to be a big part of the data science team. These solutions will definitely help with just, you know, getting the information across the organization that they need in a more timely fashion. And that's always been a challenge for information professionals, is making sure that everybody has access to the information that they need. Great. And I actually have a question for Heather, if I may. Yes, of course.

Heather brought up a really good point that, you know, it actually needs these collaborations among, like, you know, life scientists and data scientists and statisticians, right? And then you also kind of work with multiple clients, really see at the higher level how those... I would love to kind of ask you if you can share some of the experiences, because what we see is that it's really hard to collaborate, right? It's like the data science is like, you know, feeding them into, into the models or dashboards and what have you, right? Sequels and all these different stuff. And then you have, they basically come in with different skill sets, if you will.

And sometimes it's like language is a barrier, right? I mean, like, the domain language is a barrier. How do you see kind of a success, like seeing multiple teams work together? I think it requires just a lot of communication, a lot of meetings, and, yeah, just trying to get on the same... I mean, everybody has their own expertise. And yes, trying to translate that to the people that you're collaborating with is... can be a challenge, but obviously it's, you know, if you can get that piece of it, it's really going to be a beneficial collaboration in the end. Great. Thank you. Thanks for bringing that up, Lana.

Because actually, yeah, I wanted to know more about the collaboration. I mean, I know that you're a big part of the collaboration, but I wanted to know more about the collaboration around science statisticians and data scientists. Do you facilitate that, Heather, with your clients' organizations? Do you find the need to facilitate that more? I know in my experience, I've always seen them be kind of siloed. And, you know, and many times I've found myself having to make introductions at a client. Hey, did you know your colleague is doing this? Do you find that same kind of behavior? I do, yes, with some of my clients.

Some of my clients have more of a kind of data science group that kind of works together with all of the different stakeholders, like a data innovation kind of group. And their goal is to, you know, kind of assimilate all of that information from the different groups. So it hasn't fallen. It hasn't fallen to me, but it has fallen to this, like, data innovation kind of group to collaborate between the stakeholders. So I'm wondering kind of my take, and if I may, is that maybe being a consultant, being kind of an outside person actually helps, right?

A lot of times, you know, I've been in that. I came from Novartis. It's like, oh, we'll never be able to do this internally. But somehow having an expert, like, you know, external expert and saying, oh, you got to do this. It makes it easier to digest. Yeah. And especially since, you know, I see so many different organizations and how they work. So I can pick and choose, like, you know, what works at different companies and kind of pass that along. So, yeah. Shout out for Heather Services. Terrific. We have a question from the audience. Lana, this is for you. How does Huma AI allow collaboration within an organization?

Would the results be different? And for each researcher on the same topic as can be seen in ChatGPC? Oh, that's a great question. This is a kind of a concern across the board. So what we do is we call them single tenant system, where each client has their own kind of encrypted private environment. On AWS, we actually bring public data sources into your private environment. And so all the questions you ask, even on the public data set, are private. And the answers to your questions, are also private. That's how we do this. And we actually, when we started out, we had to fight pretty hard with OpenAI

to basically, in the contract, they won't use our data for training their models. And of course, it's standard now. Because every generative AI company applying it for enterprise will require that. So they don't see our data. So they don't use it to train our models. So we really kind of silo that. And then in some companies, I think there was a similar question. And I think that's a really good point in there that, you know, if it can't leave your environment, then we can actually create some kind of architecture that basically pushes our models and all the analysis contained in your sub accounts as well.

Right. That kind of also touches on maintaining privacy and security, while managing analyzing big data and life sciences. Any comments for both of you on that? I mean, privacy is a huge concern these days. Any other additional comments, Lana? I think it is a big, big issue, right? It's that, you know, you see that, you know, recently Google barred their internal employees using barred, right? Their chat thing. But then it's, I think part of the reason is also barred is using, Google is using OpenAI GPTs, chat GPT to train their barred models. So therefore, you know, there's a little bit of cross-pollinization here.

But I think it's, you know, it's a big issue. There is a, something that we can do absolutely from the vendor perspective, right? There's also a human factor here, right? And it's like, you know, how do you prevent someone to basically, you know, feeding their private data into chat GPT, right? So I think there's, there's a policy level, there's the company policy level, there's the regulation from the, I think, from the regulatory body. I mean, they're moving so slowly and the industry is moving so quickly. They will have to act fast. They have to see in the broader term, right? How do we put these guardrails in place to secure privacy and security?

So there's like multifaceted approach and it's going to be uphill battle, I must admit. Great. Heather, did you want to add anything to that, to Lana's comments? No, I agree with her. I mean, it is, it sounds like it is going to be an uphill battle. Great. We're coming up on time. So I've got one final question. I'd like some final thoughts from you both. In order to manage these challenges, life science, information professionals need to constantly develop their skills, employ advanced technologies like AI and machine learning to help sift through and analyze information, and work collaboratively with other professionals to ensure the most relevant and accurate information is utilized effectively.

In the biomedical field alone, more than a million papers, Paul, are being poured into PubMed databases at a rate of about two papers per minute. So how can we prepare future generations of information professionals to handle this ever increasing volume of medical knowledge? What skills and competencies will they need? Maybe I can take this one on. I will let Heather closing this. So I am in kind of a unique position, right? So I came from life sciences. I'm what you call a subject matter expert. I'm a professor. I'm an expert. But then I'm a founder of an AI company. So I see kind of both sides and of course talk with clients all the time.

So I really see this like I think we have to embrace technology. We have to embrace particularly that AI. It's changing so quickly. I like to say it's really difficult. You see these like coming at you at like 100 miles an hour. It's like, you know, but generative AI is different. It's not a shiny object. So I really kind of like and like, you know, call it. I put out a call to my life science colleagues. It's like embrace, embrace AI because, you know, we worry about, right? You know, like, you know, we're all human. We worry about is this going to replace us? And I like to say, you know, AI is not going to replace you.

But people who use AI will replace you, particularly that AI. Now the new AI solution can solve really tangible problems for us. Right. And then we really need to embrace this rather than bury our head in the sand. Sometimes as a CEO of an AI company, I'm like, oh, this is moving so quickly for even for me. But but, you know, it is fact of life. It is this is where we're going. Right. And then also from a human perspective, from expert perspective, we embrace the technology. It's also our contribution to make sure AI is applied for good and like really saving lives. Right. Pushing a new drug to market and saving patient life rather than being used for some more

frivolous pursuit, if you will. Well said Lana. Well said. Heather, anything you'd like to add as your final thoughts? No, I definitely agree with her that, you know, I think that it's the future, you know, using using generative AI is going to be beneficial to organizations. And I think we need to embrace it. Terrific. Thank you, ladies. I'd like to thank Heather Desmarais. I hope I got it right that time. And Dr. Lana Fang from Humanae.ai. Thank you for taking the time from your very busy schedules to chat with us today and help us to learn more about this ever changing topic. So I would encourage everybody if they have any additional questions, feel free to reach out to us.

And I wish you all a wonderful day. Thank you. Thank you so much. Bye bye. Thank you. Bye.