Machine Learning in Drug Discovery Symposium: Lightning Talks

Who speaks in this episode?

• Imran Hawk (Lightning talk speaker (industry presenter))
• Jouer Jang (Lightning talk speaker (invited academic presenter))

What is discussed, and when

1. 0:00 – 5:00: Jouer Jang, Georgia Tech
2. 5:00 – 10:00: Prostate Net, Prostate DB
3. 10:00 – 15:00: Imran Hawk, Kimia Therapeutics, PRMT5 inhibitors

Notable quotes from "Machine Learning in Drug Discovery Symposium: Lightning Talks"

• So first I invite Jouer Jang from Georgia Tech and it will come and her talk will be F A fully agentic drug engine a conversational artificial intelligence.

  about Jouer Jang · at 0:03

• All of them and send them to both to predict the 3D structures and also the binding affinities. So we have tested our workflow on two case studies but due to the time limit…

  about EGFR · at 3:28

• So hi everyone today I will to introduce about our project the prostate net a deep learning based protein thermal stabilitative predators. The prostate net will require the 3D.

  about Prostate Net · at 5:04

• The protein thermal stability change on the point mutations data from all the established database and to combine the datas together to form the first comprehensive protein…

  about Prostate DB · at 6:01

• Question that's specific and the scope of question has to match the complexity of your data. [snorts] for example, here we have data from RNA sequencing experiments on cell…

  about PRMT5 inhibitors · at 10:01

• There we go. All right. hey. I'm Hawk. I'm CTO at Kimia Therapeutics. we are a chemistry platform startup based in Berkeley focused on radically transforming the scale of.

  about Imran Hawk · at 14:04

Who and what is mentioned in this episode?

People

• Imran Hawk — Person · Lightning talk speaker (industry presenter) · 1 mention · first at 13:54
  Introduced as CTO of Kimia Therapeutics, describing a nanogram/nanoliter-scale chemistry + screening platform integrated with ML to accelerate hit-to-lead and improve selectivity.
• Jouer Jang — Person · Lightning talk speaker (invited academic presenter) · 1 mention · first at 0:03
  Introduced as a Georgia Tech speaker presenting an LLM-powered agentic workflow for early-stage computational drug discovery.

Companies & orgs

• Kimia Therapeutics — Company · Business focus · 1 mention · first at 13:54
  A chemistry platform startup focused on transforming the scale of what can be achieved in HIT to lead and chemistry model development.
  There we go. All right. hey. I'm Hawk. I'm CTO at Kimia Therapeutics. we are a chemistry platform startup based in Berkeley focused on radically transforming the scale of what…
• Georgia Tech — Company · Affiliation · 1 mention · first at 0:03
  The institution where Jouer Jang, one of the speakers, is affiliated with.
  So first I invite Jouer Jang from Georgia Tech and it will come and her talk will be F A fully agentic drug engine a conversational artificial intelligence.

Products & tech

• Prostate Net — Product · Case study · 4 mentions · first at 5:07
  A deep learning based protein thermal stability predictor discussed in the presentation.
• Prostate DB — Product · Case study · 1 mention · first at 6:15
  A comprehensive protein thermodynamic database mentioned in the presentation.

Concepts & topics

• EGFR — Concept · Core topic · 4 mentions
  A target protein discussed in the context of drug discovery and binding affinity prediction.
• PRMT5 inhibitors — Concept · Key argument · 1 mention · first at 10:10
  Promising cancer drugs in clinical trials discussed in the context of RNA sequencing experiments.

Full transcript with timestamps

0:03 So first I um invite Jouer Jang from Georgia Tech and it will come and her talk will be F A fully agentic drug engine a conversational artificial intelligence platform for drug discovery.

0:24 Hi everyone, I'm SA. It is a great pleasure for me to be here today representing my team and talk about our project on fate which is a conversational AI workflow for accelerating drug discovery. So drug discovery is always known to be a lengthy and resource inensive process. It typically requires 10 to 15 years and cost more than $2 billion for successful drug candidate to enter the job uh to enter the market. So uh to address these challenges, computational approaches have been very critical to for the early stage drug development. However, most of the methods we have today are uh are very task specific and require human uh ex uh execution by domain experts. So since the drug development is inherently a multi-step and interconnected process, we think this fragmented implementation uh limits both the scalability and the efficiency of the current pipeline. So to address this limitation here, we

1:21 propose an LLM powered multi- aent workflow which can automate and streamline the key tasks across the early stage computational drug discovery uh drug discovery pipeline. So by combining the LLM reasoning with the domain specific tools, our workflow can perform data retrieval, molecule generation, property prediction and also the binding affinity prediction. Here we summarize the our overall workflow at the top of the slide. So first the user describes their target of interest in natural languages. For example, you can ask our agent, can you suggest me some small molecule candidates that could target the EGFR target? With this input query, our agent will first uh extract the target information and the user requirements from this message. Then it will start searching the databases in a hierarchical order. If a co crystalized protein liant structure can be found from the RCSB database, our agent will

2:18 directly download the structure with the highest resolution and separate the bound liant from the protein structure. If only the apo protein structure is found, the agent will send it to f pocket which can be used to detect all potential liant binding pockets. If none of the protein structure is available, our agent will find the sequence from the uniro database. Send you use both two to predict the 3D structure of the protein and then send it to f pocket for bending pocket detection. And after this step we will send our input structures to our molecule generation agent which will execute the deep sbdd package for the novel molecule generation. So in this step uh uh we will evaluate the drug likeness of our generated molecules by our defined criteria such as the empirical rule of five the QED scores and the synthetic accessibility scores.

3:10 The users can also customize their own success criteria and ask the agent to keep generating molecules until it meets the criteria. And after this for the final step, we will send all the generated molecules to our agent so it can rank them based on the scores and it will select the top 10 molecules out of all of them and uh send them to both to predict the 3D structures and also the binding affinities. So we have tested our workflow on two case studies but due to the time limit today I will only show the results for the EGFR case at the bottom. So in this case our agent successfully retrieved the protein uh liant structure coar solar structure from the RCSB and separate the bond liant as the reference liant. So at the bottom uh left is the simple 2D visualization of our generated chemical space. The point the red point on the right is the reference liant. So when we in our molecule generation agent we also

4:06 set up a tiny motoral similarity threshold. So we ask the agent to generate molecules that are different from the reference liant if there's a no liant available. And after that we uh the two structures highlighted are the top lians we find and we use the both to generate that to predict their structures and also predict their binding affinities. So given the binding affinities we predicted and the comparison to the reference liant we think these two lians are potentially moderate to strong banders that worth for uh further exploring. So uh in summary, our workflow is one step further to the is already is one step further in applying the AI agents for accelerating drug the drug discovery process and it can shorten the turnar around time to generate some new ideas to be tested out and the modular architecture also allows the integration of new tools and models as it comes up.

4:59 So uh insuring oh okay sorry. So uh hi everyone today I will to introduce about our project the prostate net a deep learning based protein thermal stabilitative predators. The prostate net will require the 3D structure of the proteins as the input and it will to predict the protein thermal stability change on the point mutations which is to input by the users. The model will predict if the point mutations increase or to decrease the protein thermal stabilities. Here we have to use the geometry vector perceptions the graph neuron network as the core of our framework. The GVP and the GN have to can learn the scalar and the vector values over the 3D Ukrainian space which have to shows the potentials in predicting the protein properties.

5:58 So for the database we have to collect the protein thermal stability change on the point mutations data from all the established database and to combine the datas together to form the first comprehensive protein thermodynamic database called the prostate DB. And this database contained the both single point and the multiple points mutations because uh previously the multiple point mutations data didn't have the testing set. So we introduced the novel clustering methods which to split the multiple point mutation data into the training and the testing set. And at the same time the method has to ensure the generalization of the models. To solve the problems of the data limitations we have to develop the novel uh data augmentations methods called the thermodynamic loop and to combine with the anti-ymmetry augmentations methods to enlarge the data size. With the enlarged data sets we train our models.

7:06 We have to use the testing set to evaluate how's the performance of the prostate net because originally the testing set contains the large amounts of the destabilizing mutations. In order to ensure our prostate net can to predict accurately for both the destabilizing and the stabilizing mutations, we have to introduce the reverse uh mutations for the testing set. Here we call the protein A mutate to the protein B as the direct mutations and for if the protein B mutate to the protein A we will to call it as the reverse mutation. The reverse mutations thermal stability change is the opposite of the direct mutations. So for the originally of the testing set it it is called the direct mutations. So we generate the reverse mutations for the testing set. So as the performance we can see the prostay net have to show the high and accurate uh performance in both the direct and the reverse mutations

8:11 which have to shows the less bias and to best com better compon performance compared to the state of our methods to further to validate the performance of the prostate net. We use the prostate net to predict the thermal stability change of the servo single point and the multiple points mutations for the humanized nano body J3. And after the experimental validations, four out of the five single point mutations and all the multiple point mutations were predicted correctly by the prostain and the prostain also predicted correctly for the two positive mutations and both of them have to show it can to improve the thermal stability of the HUJ3. On the summary, the prostate net have to show the potentials to accurately to predict the protein thermal stability change. Thank you.

9:10 Um good afternoon everyone. Um my name is Suang. I don't have fancy models but I wanted to talk about the new design for biological discovery. One that bridges the gap between computation and the wet lab. Um in my opinion there are with the introduction of many great foundation models a powerful engine is already built for us. So this changes our focus from model to a more holistic experimental design. This is our lab's version of human in the loop cycle where we first frame the question precisely and then we fine-tune a pre-trained model and interpret the model and finally validate those uh discovery back in the lab. So now let me give a more concrete example. It all begins with framing the question. In another word, this means defining a biological question that's specific and the scope of question has to match the complexity of your data. [snorts] Uh for example, here we have data from RNA sequencing

10:10 experiments on cell treated with PRMT5 inhibitors. Where is it? PRM FIF high inhibitors which are pro promising cancer drugs that's in clinical trials and using splicing analysis of the cell line we find um we we identified about 35 hundreds of detangron and half of them were upregulated in red here and half were not. This give naturally two classes and allow us to frame a precise question within 150 nucleotide of the frame fi are there any regulator sequence that determine if the intron will be upregulated by the drug.

10:50 So next we move to the model um since we have labeled DNA sequence data we choose to find a genomic model DNA bird and the next critical step is interpretation. So um we use two methods back propagation based method and incilico mutagenesis. Um so to basically find what the model actually learned let's first look at the known fact before the model. Uh this is a simple overlay of the data which is nucleide frequency plot. You see the signature GT at the splice site. uh yeah which is the signature of this we which is we what we expected and you also see in class zero there's over representation of the G's at minus one and plus five position so now let's look at what model learned this logo this logo shows the feature importance from the model for individual instance and for a whole class so uh basically what's the first thing you see the GT is gone Right. Uh this is perfect because GT is in both most of the

11:59 instances in both classes. It's not a different differentiating uh feature. So the model learns to ignore that. And what did they learned? Uh we do still see the 2G uh minus one and plus five, which is great. Uh what's more, there are more new discoveries. Let's look at this example here. uh although the sequence has 2G it's it's it's assigning to class one uh why the interpretation shows showed us that 2G did contribute negatively but the other sequences near in this intron have a stronger positive influence uh that push it towards class one and this is the type of nuances we can only get from an interpretable model okay hopefully I've convinced you how powerful this model interpretation can And what's even more important is that these are not just computational patterns. They're easily testable by a splicing reporter. And last, I want to emphasize that this is this discovery is not the end. This is

13:06 actually a launchpad for many new exciting ideas and questions. For example, we can go through all of the sequences features to find new regulatory motifs and ask is it a common pattern? what protein binds here. And perhaps more uh interestingly, we can also cross reference features with known human snips to uh potentially infer patients drug discovery in clinical trial. Um in conclusion, foundation models are changing how we can do biology here now. And in many cases, bottleneck is not computation, it's efficient design. And this framework isn't meant to replace rigorous experiments but to expedite them. Thank you very much for having me.

13:52 Thank you so much. And finally, uh, Imran Hawk from Kimia Therapeutics. This is my reward for bragging last night at the speaker dinner that I could give PowerPoint roulette. There we go. All right. Um, hey. Uh, I'm Hawk. I'm CTO at Kimia Therapeutics. Uh we are a chemistry platform startup based in Berkeley uh focused on radically transforming uh the scale of what can be achieved in HIT to lead uh as well as chemistry model development. Uh the way that we do that is by integrating synthesis uh of molecules at nanogram nanolit scale with screening on those same molecules in an integrated platform uh with machine learning driven integration of thousands of data points that we deliver per hit to lead cycle uh to deliver uh a dynamic exploration of chemical diversity. Uh and each of those words really matters here. This is informed by by my past experience uh you know for the last uh five and a half

14:47 years I was at recursion pharmaceuticals where we generated an enormous amount of data. Um but one of the challenges that we ran into is that although biology operates at a at a huge scale, chemistry operates a lot more slowly. And so if you look at the progress of a standard chemistry program, uh folks will identify hits. You may be able to do one maybe two rounds of catalog expansion of those hits from commercially available molecules for call it $100 a molecule in a few weeks, but very rapidly you end up having to do custom synthesis where each molecule costs you $1,000 and you know you'll get your results next quarter.

15:24 That's really challenging not just for chemists and hit to lead but I think really challenging for the folks in this room trying to do machine learning because now you're trying to build models off like 10 molecules per cycle. So how do we address this? We address this by dealing with this on the data side. Um this starts at the top left. Uh we're doing uh chemical synthesis at the nanocale where we're making a thousand to 100,000 molecules per cycle uh at nanog scale. So already you're putting two to four orders of magnitude in the chemical breath that you can explore per cycle. Uh this is then screened in direct to biology and direct to adme format. Uh where we run these in functional biochemical and cellular screens. Uh so rather than looking at purely biohysical data where you might see that there's an interaction but you don't know whether or not it's relevant or functional in the system you care

16:13 about you're actually getting information about the function of the molecules that you make. Uh we then take interesting molecules uh because we're operating at such scale uh we can afford to say yeah you know maybe 90% of the molecules aren't functional but the 10% that are we can then zoom in synthesize and purify them at the micro scale at hundreds of molecules per week. uh and then integrate this into synthesis aware machine learning capabilities to integrate reactivity potency and selectivity data for for the next round.

16:43 Um on my poster you'll see examples of this uh where we apply this for everything from denovo hitf finding uh diversification of molecules and molecular scaffolds guided by potency multiparameter optimization as well as using this data to drive foundation model development. Um so what I'll do in the next couple of minutes is just walk you through one demonstration of this uh in how we dealt with a very challenging selectivity challenge. Uh I think Tom's uh presentation in the morning like presented a really great example where he talked about you know needing to drive the the selectivity uh for their their molecule IM1363 versus EGFR in order to prevent dose limiting toxicities. This is a really broad problem across a lot of uh a lot of targets uh in in oncology and and beyond where there's a wealth of targets that have lots of genetic evidence that have lots of validation that they would be

17:35 great. The problem is that there are very closely related family members uh that you have to be selective against uh in order to to to actually be an effective therapeutic. Uh and these present really really difficult chemistry challenges. So we took one of these challenges on. This was an example where there's a target uh that has greater than 88% identity uh residue identity in the binding pocket against two other offtargets. Uh we started with a non- selective uh tool molecule up here up here in the middle. Uh you can see that it's actually stronger on these anti-targets uh than it was on the target. Uh after two cycles of optimization on the on the Atlas platform. So again this is synthesis screening purification machine learning guided design. In this case, using uh models that are optimizing both select uh both potency on the target as well as selectivity, anti- potency on these on

18:29 these other two targets, uh we're able to increase the selectivity by a factor of 250fold. Uh and we got a molecule that was 34 nanomar um against a target uh and up in the up in the microar range against the anti-targets. So I think this is a really exciting platform to generate exploration diversity as well as data for model building. Come see me at poster 7 during the break. Thank you.

18:55 All right. I heard the applaud from outside. So I think we already gave a round of applause to all our speakers. Um but I would um like to welcome all our four lightning talk speakers for one or two questions. If anyone has please come here and we'll see if there is any question before we take the lunch break.

19:33 I have a question for Zu. Yes. Um, so you you've now predicted binding affinities or ranking molecules. How do you plan to validate your predictions?

19:46 That's a very good question. Uh, we don't have the collaboration with the experimental chemists to validate that ideas, but we hope this could work in a more collaborative setting. So people can take our ideas to see if they could be synthetized really experimentally or we can do a similarity comparison to any known uh like known uh drug candidates from the literature to see if that's already being explored in a similar way.

20:14 Yeah, I think that's going to be really critical to validate your method. Second question is so you showed an example EGFR where there's a lot of structural information available for that target. Yes. Yes. And you mentioned your pipeline that if there is no structural information, you would use bolts to to predict a structure. Have you done any kind of comparison to show like kind of what Woody was pointing to earlier that that the correlation or your ability to predict novel higher or tightly binding molecules if there isn't a lot of structural information? Is there that sort of linear correlation that

20:52 so we haven't done that studies by ourself but the other case study we have tested is when we have the apo structure how can we def like identify the multiple binding pockets from that and screen different lians generate different lians for different pockets but that's a really good direction we should take into consideration

21:09 thank you thank you all right so um we will give one more round of applause to all our speakers this Amen.

Questions about this episode

Who speaks in "Machine Learning in Drug Discovery Symposium: Lightning Talks"?

Imran Hawk (Lightning talk speaker (industry presenter)), Jouer Jang (Lightning talk speaker (invited academic presenter)).

What is discussed in "Machine Learning in Drug Discovery Symposium: Lightning Talks"?

In order: 0:00 Jouer Jang, Georgia Tech; 5:00 Prostate Net, Prostate DB; 10:00 Imran Hawk, Kimia Therapeutics, PRMT5 inhibitors.

What is mentioned in this Broad Institute episode?

8 people, works and ideas are mentioned, including Imran Hawk, Kimia Therapeutics, EGFR, Jouer Jang, Prostate Net, PRMT5 inhibitors. Each one on this page links to its own page with every episode that mentions it.

How long is "Machine Learning in Drug Discovery Symposium: Lightning Talks"?

The episode runs 22 min; the transcript is about 3,366 words in 28 timestamped sections.

More from Broad Institute

• All analyzed episodes
• People mentioned on Broad Institute
• Topics mentioned on Broad Institute