0:00
Thank you everyone for coming.
0:02
So Yep, I'm Sarah and I am a scientific specialist at Causaly.
0:07
And today I'm going to be talking to you about how you can enhance drug discovery using human centric generative AI.
0:20
So in today's talk, I want to 1st begin by explaining some of the challenges that we're facing with finding knowledge and overall efficiency in drug discovery.
0:30
Then I want to introduce Causaly, and we'll explore together how human centric generative AI can support and accelerate biomedical discoveries.
0:40
And then lastly, we will see how this works, of course, in the real world by stepping into the shoes of a discovery scientist that has a target ID and validation use case.
0:53
So I'm sure most people in this room are very familiar with these stats, but I want to reiterate to you some of the really big challenges we're facing in drug development.
1:04
So a new drug takes around sort of 10 to 15 years or even more to develop with an average cost of 1 to $2 billion spent during its development.
1:14
Now typically less than one in 10,000 potential drug compounds that enter into the drug discovery journey get successfully through the clinic.
1:23
As the earliest phase in the pipeline, pre-discovery plays a crucial role in drug development.
1:28
So I wanted to spend some time focusing on this phase today.
1:33
So this pre-discovery phase typically takes around 5.5 years, which is 1/2 to 1/3 of the time of total drug development, and it costs 1/3 of the total cost.
1:47
So importantly, when reviewing the causes of drug failure in clinical trials, 52% of drugs fail due to lack of efficacy.
1:56
This is often attributed to the poor selection of the drug target in the beginning.
2:01
For example, selecting a target without a significant disease linkage or a target that is not easily accessible to drug compounds, i.e., is undruggable.
2:12
Taken together, identifying an efficacious and safe target during the pre-discovery phase plays a pivotal role in drug development and of course increases the likelihood of success.
2:27
So the target selection and validation process requires analysing a large amount of existing biomedical publications.
2:35
Yet the conventional literature searching method PubMed was an innovation from 1997.
2:41
Now from 1997 to 2020, 17 million more scientific documents were published.
2:48
And without any further enhancements of our existing literature search tools, researchers are naturally missing important information from these scientific documents.
2:59
So in the next slide, I want to show you how this problem is slowing down the entire drug discovery process.
3:06
Before I do that, could you please raise your hand if you use PubMed in your research, expect nearly everyone.
3:15
OK, good to know.
3:18
OK, so let's say that you have performed a search looking for targets for a non-small cell lung cancer.
3:25
Here are the four issues that you'll likely be encountering.
3:29
Now the first, which is probably the most familiar, data overload.
3:33
You always have an overwhelming number of papers to read, and there's no systematic view on the scientific information contained within these papers.
3:42
The second issue is bias.
3:44
You will likely only scan the first few pages of PubMed.
3:48
Even though there are usually thousands of pages of results.
3:51
Nobody goes to page 4000 of PubMed, and there's also going to be a subconscious bias in selecting evidence.
3:58
We tend to gravitate towards papers that confirm our hypotheses.
4:03
The third problem is there is a loss of hidden evidence.
4:06
It's impossible to compare all the literature to identify indirect links in between papers, in between evidence.
4:14
And the final problem is one of poor traceability.
4:17
There's no collective memory of research performed, so the team tends to spend extra time rereading the same evidence over and over, or reevaluating the same targets and of course taking even more time.
4:30
There we go.
4:36
So in the ideal world we would want to read all publications related to your topic and highlight every single relevant sentence without bias, to give you full transparency and to maximise and optimise your decision making process.
4:52
So what if I told you have a friend able to do that?
4:56
What types of questions would you ask?
5:01
Indeed you can.
5:03
You can ask this question to Causaly, which is a powerful AI solution that reads literature like a human.
5:09
So Causaly integrates multiple data sources, including the ones up here, and extract evidence points from scientific documents to create a high precision knowledge graph which contains billions of sentences.
5:23
With Causaly, you can directly receive answers to your questions instead of manually trying to read papers and extracting insights yourself.
5:32
Now, since Causaly has already read the literature, you can very quickly view all related evidence for a scientific topic.
5:41
Causaly itself is an adoptable AI solution.
5:44
It's currently supporting thousands of scientists and enabling them to make better decisions and increase productivity.
5:51
Of course, you don't have to believe me.
5:53
Our platform is trusted by top pharma companies, including the ones up here, just to name a few.
5:59
And here are a couple of testimonials from our users about how Causaly has changed the way that they perform research.
6:10
So to help you navigate and understand scientific data even more readily, we've now integrated Causaly's knowledge Graph with a large language model and this gives you a personalised research companion, which we call Causaly Copilot.
6:25
Combined now with the capabilities of generative AI, you can ask your question to Causaly in a conversational manner and the spectrum of questions you can ask also broadens.
6:36
Causaly Copilot will give you accurate and reliable results with complete transparency in terms of the data source used.
6:45
Lastly as well, Causaly Co Pilot will give you an AI generated summary of key findings for your specific topic with linked inline citations for easy understanding of and of course access to the scientific data.
7:02
So the exciting part?
7:04
Now let's see how Causaly works in action by putting ourselves in the shoes of a drug discovery scientist in pharma.
7:14
So let's take a look at the research routine of Jo.
7:18
Jo is a discovery scientist at a pharma company.
7:21
So Jo's task is to identify a promising drug target for Huntington's disease.
7:27
As a first step, she wishes to identify a list of all potential Huntington's targets.
7:33
She would then like to find targets with human evidence as she views these as more validated.
7:39
With all this information, she then wants to prioritise the targets in two different ways.
7:44
The first is those that have genetic mutations in humans, and the second is looking for targets that are novel.
7:51
Now, in the next slide, I'll show you exactly how these questions can be identified and addressed with Causaly.
7:59
Hopefully this video will work.
8:01
Yeah, OK, so this is what's happening.
8:04
Jo will log into her Causaly account and ask her first question using Causaly Copilot.
8:10
She describes it in natural language, what are the targets for Huntington's disease in humans and initiates the search, so Causaly understands the question.
8:21
And in just a few seconds, we'll return to Jo a list of 1300 targets for Huntington's disease in this interactive dendrogram format.
8:31
So as she scrolls down, she sees molecules that have been linked to Huntington's disease in publications ranked by article count.
8:40
Now the dendrogram itself is fully interactive, meaning that Jo can choose to explore any of the targets of interest, for example Huntington.
8:49
She clicks on it to see all the supporting information, including the number of evidence points and the supporting articles.
8:56
Importantly, Causaly has already extracted the relevant evidence points for Jo from the papers, so she saves extra time doing this manually.
9:05
At the same time, Causaly Copilot has generated a short summary based on the key findings of this question, containing inline linked citations for transparency.
9:16
Jo can also choose to see an in depth analysis for a more thorough review of the topic which includes introduction, results and conclusion, allowing her to very quickly understand the scientific evidence for this search.
9:29
So Jo can now copy this report for sharing and also save it to a workspace so she can continue working on this topic in a Causaly folder.
9:38
So her new workspace has now included this AI generated report where she can add further summaries.
9:45
The articles referenced can also be found in this folder and Jo can click on any one of these to view the abstract or ultimately get to the original data source.
9:55
The targets summarised can also be found in the Relationships tab, which can be visualised as well in this network format and it's fully interactive and customisable.
10:05
Last but not least, Jo can share the entire contents of this workspace with another colleague, so they can then share and collaborate using Causaly.
10:20
So as we just saw, Causaly answers scientific questions instantly using information from our knowledge graph and all relevant evidence is available.
10:29
So there really is no black box.
10:32
Jo can make use of Causaly Copilot for easy understanding of the topic, which is reproducible.
10:37
It's reliable and due to the fact that we only generate the summary based on the data sources we integrate, hallucinations are very rare.
10:46
As a last step, Jo can explore the direction of effect in target disease relationships and therefore explore also if contradictory evidence exists for her research question.
11:01
As a next step, if we continue the workflow, she wants to prioritise genetic mutations in humans that are linked to Huntington's disease Without Causaly, Jo would have to review a different set of documents to identify these mutations and gather evidence.
11:16
In Causaly, she simply needs to add the keyword mutation to specify her search further.
11:22
By doing so, she then identifies 569 mutations and adds JPH3 as a candidate target as it has clear evidence to be genetically linked to Huntington's.
11:39
So based on Jo's preference, she also wants to prioritise the targets based on novelty.
11:45
To do this, she continues the dialogue in Co Pilot and asks Causaly to find emerging targets for Huntington's disease.
11:53
Causaly understands the question, retrieves the relevant data, and this time takes Jo to a unique timeline view.
12:02
So here the X axis represents the timeline, each bubble represents one of the targets for Huntington's disease, and each target is positioned based on the date of the very first publication that linked it to the disease.
12:16
So the very first disease linkage.
12:19
This gives Jo the option to easily view these emerging targets that have been very recently described.
12:27
In this workflow she explores the micro RNA miR-615 as it was linked to Huntington's disease in very recent publications.
12:36
She then takes this relationship and of course the articles of interest and adds them to her workspace created earlier.
12:48
So having identified targets for Huntington's disease, Jo decides to evaluate and validate the two candidate targets, the emerging noncoding RNA micro RNA 615 and the human mutated JPH3.
13:03
So here her goals now are to identify the target that has a more significant link to neural dysfunctional processes.
13:10
She also wants to discover any hidden links between the target and the disease in order to further strengthen the validation.
13:17
Then, as a last step, she wants to understand what assays she should use to validate this target in the lab and subsequently share all these findings with her team.
13:30
As a first step here, Jo compares JPH3 and Micro RNA 615 to discover the pathological functions and the cellular dysfunction processes they affect.
13:41
Using a conventional literature search method, we can use PubMed in this example, it's impossible to extract and compare all the supporting evidence for the targets in a feasible amount of time.
13:53
In Causaly, she can run a comparative analysis to see which processes are associated with the targets individually and those that are associated with both targets.
14:04
In this network view here, I can see that both targets are linked to, for example, oxidative damage and cell adhesion and invasion processes.
14:13
In addition, each target is also uniquely associated with certain processes.
14:19
For example, only micro RNA 615 is linked to nerve degeneration and demyelination.
14:25
From this analysis, and of course by going into the papers to look at the data, Jo concludes that micro RNA 615 has been linked more heavily to neural dysfunctional processes according to the literature.
14:38
With this analysis, she prioritises the micro RNA target.
14:44
Now given that this micro RNA target is a relatively new target for Huntington's disease, let's find out if there are any hidden links between micro RNA 615 and Huntington's in the literature.
14:56
Using the unique multi hop functionality of Causaly, Jo breaks this relationship down to find any genes and proteins that could indirectly link this target with Huntington's disease.
15:08
In less than one minute, the system returns 35 mediators.
15:12
One of these mediators is HIP1, which encodes for a protein known to interact with Huntington.
15:18
And these findings of course further strengthen the validation and increased confidence in this target.
15:28
As a last step to ultimately validate this target in the lab, Jo wanted to understand the key assays used to measure micro RNAs.
15:36
She simply asked Causaly Copilot and gained a summary of methods with linked inline citations.
15:44
As a last step, Jo wanted to share her findings.
15:47
She has a centralised location to do so in Causaly workspaces.
15:54
To summarise briefly as I know we don't have so much time, Causaly allowed Jo to complete her To Do List much faster.
16:00
She identified 1300 targets and very rapidly was able to jump down and get to 2 targets and ultimately one that she decided to proceed into the lab.
16:12
As a very last step, I just want to show you this quote.
16:16
It's clear how Causaly will save time and reduce bias.
16:19
And it is clear to me now that scientists of course, will not be replaced by AI, but they may be replaced by scientists that use AI.
16:28
Just something to think about.
16:31
Thank you everyone for the time and attention.
16:33
If you have additional questions and would like to see more, please come by the booth and I'd be happy to talk more about it.