0:00
I appreciate I'm all that standing between you guys and a coffee break.
0:04
So I'll do my best to keep the time.
0:06
So thanks for coming to this talk.
0:09
So we're a CRO and one of the issues that we have, like many CROs is that the guys in the labs do amazing work every day, but we can't tell anybody about it because we don't own the intellectual property.
0:24
So we look to bring in university students and industrial placement students so we can do, you know, projects and actually show people some of the things that we do.
0:37
So I do want to tell you about one such story today.
0:43
But before I get into that, for those of you who don't know us, we're CRO based in the UK concept and our heritage companies have been supporting our clients drug discovery projects for over 25 years at this point.
0:59
We do GMP synthesis and we acquired one of our competitors last year, which allowed us to do more integrated type projects.
1:10
And our chemistry labs are, you know, fortunately we're very lucky to be based in Chapel-en-le-Frith in Derbyshire, which is in the heart of the Peak District where we have 14 laboratories, close to 100 chemists.
1:25
So we're possibly the biggest chemistry department under 1 roof in the UK.
1:30
Possibly most of our scientists have PhDs and we've got our ADME department is also co-located with our research chemistry.
1:41
Some of you might recognise this building.
1:43
So we've also got our manufacture facilities in Sandwich in Kent, which many of you might know is the old Pfizer R&D laboratories.
1:55
So moving on to the project then.
1:57
So we started this project by having a little bit of an interest in liver cancer.
2:02
It's the fifth most common cancer worldwide.
2:04
It has the third highest mortality rate and there's a growing patient population estimated to reach one and a half million in 2050.
2:15
The vast majority of liver cancers are hepatocellular carcinoma and the leading causes of this disease include viral infection, HPV, HCV, liver cirrhosis and alcohol.
2:29
The disease is often diagnosed late post metastasis.
2:32
And hopefully you can see from this map that there's also high regional disparities, right?
2:38
So there's large instances of this disease in East Asia and Central Africa.
2:46
And you know, like many cancers, you know, hepatocellular carcinoma is driven by the accumulation of both genetic and epigenetic diseases.
2:58
Now, fortunately, there are, you know, some treatment options for hepatocellular carcinoma.
3:04
We've got a few kinase inhibitors here.
3:06
So sorafenib was approved in 2007.
3:10
It does increase survival time but doesn't necessarily increase the quality of life of the patients.
3:18
And it can also cause, you know, severe hepatic events.
3:23
Later in 2018 and 2019, Lenvatinib and cabozantinib both approved.
3:31
These compounds hit, you know, multiple kinase targets.
3:36
They offer similar clinical benefits to sorafenib, you know, but these compounds also have issues.
3:44
The lenvatinib is not suitable for patients with impaired liver function, which is a bit of a dichotomy if you think about it, it's used to treat liver cancer, and it can also cause severe cardiovascular toxicity issues.
3:59
Cabozantinib has black bot warnings for gut and bleeding.
4:07
More recently, some monoclonal antibodies have been approved for the treatment of hepatocellular carcinoma.
4:14
It does increase survival times, you know, But you know, monoclonal antibodies are quite expensive and don't necessarily fit the growing patient population in each stage or central Africa where the people might not be able to afford these types of treatments.
4:35
So we were interested in exploring, you know, a new target to treat hepatocellular carcinoma and we came across BRPF1b.
4:43
So BRPF1b stands for Bromodomain and pH finger containing 1B.
4:49
It's an epigenetic reader and we wanted to target the Bromo domain as a potential target for hepatocellular carcinoma.
5:01
So why do we think that BRPF 1b is a good target for this disease?
5:08
Well, firstly, BRPF 1b is the most highly up regulated Bromodomain in hepatocellular carcinoma.
5:15
You can see this little array right at the top here.
5:20
Secondly, the over expression RB of BRPF 1B has been related to poor survival rates in hepatocellular carcinoma patients.
5:32
And thirdly and lastly, the inhibition of BRPF 1B has been shown to slow the progress of hepatocellular carcinoma tumours in rodents.
5:44
So we've got 3 bits of data and they all suggest that BRPF1b is an interesting target for this disease.
5:53
And of course, you know, we're not the first ones to have thought of that.
5:58
And in fact there's quite a few tool compounds available that are known to buying the BRPF1b target.
6:06
And these small molecules were identified using a whole range of technologies such as ligand based virtual screening, structure guided design and fragment docking and screening.
6:21
So the two most well-known tool compounds to target this protein are NI-57 and GSK6853.
6:31
I think these were both discovered using fragment-based approaches.
6:36
I should also say that there are actually as far as we're aware there's no small molecules in the clinic to bind the BRPF1B protein.
6:48
There are at least 15 ligand BRFP1b structures published in the protein databank.
6:55
The first one was entered over a decade ago.
6:58
At this point.
6:59
We've got the protein in green and then we've got an N-acetyl lysine here in magenta.
7:07
So we wanted to do a virtual screen to try and find some new chemical matter to buying the BRPF1b protein.
7:16
So we, you know, we thought we had pretty much 2 two types of strategies that we could employ.
7:23
We've got ligand based techniques which are, you know, quite simple, very high throughput and you don't need a protein structure, but you know, they don't take into account binding to the protein and that's possibly an issue.
7:38
On the other hand, you know, there's structure-based techniques we could employ such as molecular docking.
7:44
These are possibly more relevant because they actually take into account the interaction between the small molecule and the protein, so the models are maybe more relevant and they possibly give you better hits.
7:58
But there are some drawbacks.
8:00
You know, the technique is lower throughput and you need to have a protein structure, and you might need to have more computing power as well.
8:09
So what we wanted to do is that we were thinking that because there's so many crystal structures of this protein published, maybe employing a structure based technique is the way to go.
8:22
So we wanted to screen a large virtual library and there's a very simple thought exercise if you wanted to screen or dock 30 million compounds into a protein, if each of those docking calculations took one second, that process would take you 347 days.
8:41
Most people don't want to wait a year, right, to do docking and get some hits to push on with their projects.
8:47
So we need to bring that process down from, you know, months or a year down to days.
8:54
So we were wondering if we could employ AI to come up with the workflow to run through this docking process much quicker.
9:07
So and you can do that.
9:08
So there's an algorithm and software called MolPAL.
9:11
So a molecular pooled active learning and this was developed by scientists at MIT several years ago.
9:20
So what we wanted to do was to take the eMolecule commercial available library, which in this case contained 30 million compounds and dock docking those into the BRPF1b protein.
9:34
So in terms of the how the algorithm works, then you first take your 30 million compounds and you generate molecular fingerprints for all of those.
9:43
That's very quick.
9:45
You then randomly select 0.1% of the compound library and you dock that nought 0.1% of the compound library into autodock VINA.
9:56
So that generates docking scores.
9:58
And you know, these docking scores are essentially surrogates for activity and binding at this point.
10:06
We train the model which allows us to predict docking scores based on molecular fingerprints and the docking scores that we've achieved here.
10:19
Once we train the model, we get those docking scores and then we repeat the process, right?
10:25
So we take another 0.1% of compounds and we continue going through this process until the model converges.
10:33
So that essentially means that the predicted docking scores match the real-life docking scores.
10:40
So we typically can run through this process up to 10 times, but for this particular project we went through the loop 4 times.
10:51
So at the end of this process, the software essentially ranks the compounds in order of docking score.
10:59
And then we did some basic chemical informatics and some clustering in order to allow us to sample the chemical space efficiently.
11:08
We then did some trials in based on some of the physical chem properties such as predicted solubility, but also estimated delivery time, right.
11:18
So we want to be able to get all of the compounds and screen them and move on with our project.
11:25
And following this triage in, we ended up with 51 virtual hits.
11:31
So at this point, right, we've done some docking, we've got some virtual hits.
11:37
And on this slide here we've got an image that shows the docking of lots of other HITs bind to the protein.
11:43
And on the right, we've got a little plot which shows some basic, some properties, molecular weight, hydrogen bonds, acceptors, polar surface area, percentage of SP3 hybridization, etcetera.
11:57
The blue line shows the average value of the FDA orally approved drugs.
12:03
The orange line is our virtual screening kit.
12:06
So, I mean, you can broadly see right, that the compounds that we've stumbled upon as our virtual hits, they're broadly drug like.
12:18
So we have virtual hits, but what we actually want is genuine in vitro hits, right?
12:23
So we start our screening cascade using GCI.
12:29
So I'm not a biophysicist, but I'm told that GCI sort of like SPR GCIs, it's very quick, it's very high throughput.
12:39
It's especially good for weak binders and crude samples.
12:43
So it's a good screening tool.
12:46
There's two ways of doing GCI easily using multi cycle kinetics which is closest to SPR.
12:53
Or you can use a wave rapid format, which means that you inject the same concentration of each of the compound each time that you vary the flush.
13:03
And once you've run that experiment, this is the typical plot that you end up with.
13:09
So we've got binding kinetics so K on K off and these diagonals with dissociation constants KD.
13:19
So we screened our 51 hits at 100 micromolar and we ended up with 35 BRPF1b binders.
13:30
At the same time we did a counter screen against the BRPF1a protein which is homologous and we didn't see any binding.
13:38
So, you know, this suggested to us that actually our virtual hits you know seem to be engaging in on target binding.
13:48
And this is just an example of the raw data that we used to construct this type of plot.
13:56
We then did our sort of initial confirmation binding using the more traditional multi cycle kinetics.
14:03
And we found that, you know, pretty much all of our hits were confirmed binders using this secondary technique.
14:10
And we then selected the 20 best binders.
14:13
So these had KDs of less than 250 micromolar.
14:19
So of course, you know like any good screening cascade, you need an orthogonal assay to confirm activity, but also to help you start triaging the best hits.
14:30
So for this particular workflow, we employed differential scanning telemetry where we're looking for an increase in the protein melting temperature relative to the native protein in the absence of ligand.
14:46
We selected 4°C as our cut off and this allow and this essentially gave us, you know, 9 compounds which we were happy to take forward into our next experiments.
15:00
So after DSF, we then took our molecules into some ligand observed NMR experiments.
15:07
So in these types of experiments we're looking for changes in the peaks of the ligand relative to the ligand in the absence of any protein.
15:19
So changes in peaks suggests that the protons are interacting in some way with the protein.
15:27
And following, you know these types of experiments, we had five compounds which we regarded as hits.
15:36
So at this point these hits, we've got docking data, we've got binding kinetics affinities and we've confirmed binding by DSF and NMR and this is what some of the compounds look like.
15:57
I mean, hope hopefully you can see that, you know, these broadly look like, you know small drug like molecules.
16:05
And for the eagle eyed amongst you might notice that compound 4 looks a little bit like GSK 6853.
16:14
And you know, we see that as being a good result, right, because it's sort of gives you some confidence that your workflows actually producing small molecules that bind the target protein.
16:28
On the right here, we've got you know the docking of compound one to the BRPF1b protein.
16:36
So we've got lots of data here that we got from the biophysics and then we also put these compounds in into some of our T1 ADME assays.
16:44
So we've got microsomes, Heps and SIP inhibition as well.
16:54
So just to summarise this small project then, we've essentially used AI to screen the 30 million compound library against the BRPF1b protein.
17:07
We've done some triage in, we bought in 51 compounds and using GCI, DSF and NMR, we've ended up with five hits.
17:19
We have docking information for those compounds.
17:23
We have binding kinetics and affinities and we have also some confirmation that we have on target binding by the BRPF1a screen as well.
17:37
And we completed all of this work in less than six weeks.
17:43
We've actually published this story in this Med Chem paper.
17:48
So for those that are interested, you know, if you scan the QR code, you'll go direct to the paper.
17:58
Before I finish up, I did also want to say that, you know, I'm a synthetic organic chemist by trade.
18:03
So my expertise is more how to make the molecules.
18:06
So when we've got hits developed those into leads, we've also got quite a lot of expertise of taking those leads which have been, you know synthesised the milligram scale and either inventing new chemical roots or do in process research and development to optimise those routes suitable for kilo manufacturing.
18:29
Now at GMP Labs and you know, we can deliver up to 10 kilogrammes of API for phase one clinical screening and we do these types of work flows for you know, American biotechs.
18:46
So what I've spoke about today is, you know, at the bottom of this workflow.
18:51
So we've got a little bit of virtual screening card and then a little bit of chemistry at the bottom here.
18:58
But of course, you know, we offer a number of services across the drug discovery and development pipeline to drug hunters.
19:08
So I did want to say lastly that, you know, we're also exhibiting in the central hall, we're at B37.
19:16
So you're interested in what we do or just want to come and chat science with us.
19:20
You can find us there.
19:22
So with that, I'll stop and I'll be happy to take any questions.