0:00
Yeah, thanks everybody for sticking around.
0:04
So while I'm waiting for the slides, I have to say that somebody asked me this morning, how do you add time to my drug discovery projects?
0:11
And I thought, well, oh shoot, that wasn't the idea with the title.
0:15
So what it actually should be titled is how do we add the dimension of time to structural drug design?
0:23
Anyways, my name is Dennis Nenno.
0:24
I'm running Examol with a team of small team of great computational physicists and computational chemists and we're very excited about kinetics and dynamics and small molecule drug design.
0:38
Now, obviously a lot of drug projects, if they have access to a crystal structure, they look at the static crystal structure and that can give you things like the binding site geometry tells you about the key interactions of the protein often times, and it can be sufficient to design inhibitors and design better binders.
0:56
But you know, the physicist Richard Feynman said a long time ago that biology is all about the jiggling and the wiggling of the atoms, and it can only be understood that way.
1:06
And so when you design your drugs, often times their most important effects only reveal themselves when you look at the dynamics of those drugs.
1:14
So how long do they stay bound?
1:17
What's the residence time?
1:19
How do they get in and out of the binding pocket?
1:22
Where do they bind in the first place?
1:24
Is there a cryptic pocket?
1:25
Is an allosteric side?
1:27
Is there a transient state involved?
1:29
And then on top of that, Robert Copeland's work showed that static properties like affinity actually don't translate well to clinical efficacy.
1:39
And if you look at kinetics or dynamic properties, all of those translate much better to the clinic.
1:45
And there's a number of examples that show that, just to name a few, Tiotropium, for example, targets M3 muscarinic receptor.
1:54
You can take it once a day just because it was optimised to have a very long residence time.
1:59
There's lapatinib for example, that actually trades in affinity for a slow dissociation rate.
2:05
And there's way more examples that have been optimised, sometimes by serendipitously, to have favourable kinetic effects in the clinic.
2:16
And there are a number of experimental tools that you can use.
2:19
Surface plasmon resonance, BLI, jump dilution, HDX-MS, they're all great tools for screening.
2:24
They can give you some insights.
2:26
None of them give you structural insights into how you design drugs that are better at kinetic properties.
2:32
There's machine learning approaches and if you've been to the panel before, I think most of those approaches are basically nice interpolation machines, but they don't really give you more insights.
2:43
And so you can guess what I'm advocating for here.
2:46
It's molecular dynamics which is a physics-based tool that gives you atomistic insight into how your drugs are interacting with your target.
2:56
Now unfortunately molecular dynamics has a huge time scale problem.
3:01
The longest simulation that has ever been run was a second and it took 100 days on a special purpose supercomputer called Anton by DE Shaw Research.
3:12
And even if you were to run a simulation that takes a second, a lot of biological effects like the residence time you've seen, they're beyond the second.
3:19
So you get one simulation for one second, and you might not even see what you've been looking for.
3:25
And that's a big problem.
3:26
If you were to simulate it on a GPU, it would take decades.
3:31
Thankfully, people thought about that problem and there's an approximate way at least to solve it.
3:36
And it's weighted ensemble.
3:37
It's the technology that we use internally.
3:40
Now, what is weighted ensemble?
3:43
Now imagine you were to look for your drug and trying to figure out how it unbinds from the pocket.
3:50
Now you run one simulation and it's maybe a simulation of a microsecond, but the process happens on the order of some minutes.
3:57
So you might run a simulation, and you will never see the drug unbind.
4:00
How do you cope with that?
4:03
Well, what we do is we actually start a big number of simulations in parallel.
4:08
We stop them after a while.
4:10
We look which of those simulations show the desired behaviour.
4:14
That can be a conformational change in your protein.
4:16
That can be the drug unbinding.
4:17
That can be a PROTAC forming, for example.
4:21
And then we stop the simulations, we replicate the ones we like and we merge the ones that don't show the desired behaviour.
4:30
And we do that as often as we need to get some statistical significance in the results that we want to see.
4:38
And then you might see something like that.
4:39
In the end, you see the drug unbinding and it's great and it gives you already some biophysics insight and it gives you some structural details that then the med chemists can play with.
4:52
Now this creates a whole lot of simulations and one of the best ways to look at is we found is in terms of these binding networks.
5:00
So each of those dots is its own conformation starting from bound states to unbound states.
5:06
And the distance between these states is how different the confirmations are in that space.
5:13
And the colour coding is basically how far the ligand is from the binding site.
5:17
And you see that for this ligand, it can look fairly complicated.
5:20
For this ligand, it looks fairly simple.
5:23
And your medicinal chemists or your drug designers can click on any of those dots, figure out what the transient state is and design the drug to avoid that transient state, for example, if that's what you want.
5:35
Now a lot has been talked about the limitations of AI.
5:40
This method obviously also has its limitations.
5:42
We can simulate hours or days, processes that take so long or we cannot simulate very large systems.
5:50
Certain membranes are out of reach even for a technology like this.
5:54
And then there's cases that molecular dynamics as a technology is a poor technology to use in actual programmes, maybe not so much in academia, but if you want to use them in programmes, I would strongly suggest it.
6:09
You don't, which is if your system has too much binding, if it's too charged or too disordered.
6:14
But there are some cases that I would advocate looking into that have been traditionally not been approached with molecular dynamics and that are cases like kinase specific selectivity.
6:29
You're targeting as particularly particular kinase family, particular kinase, you want it to bind longer to one than the others.
6:37
You can do that with this approach.
6:39
There are protocols by us and our competitors too that help you identify cryptic binding sites and something that we work on a lot is resistant mutations where you try to find dynamic differences in the protein motion that you then exploit that's between the protein and the mutated protein and then you then exploit to design a drug that only targets the mutation.
7:03
Of course, just two case studies before I wrap it up.
7:08
This was the original study that one of my Co-founders published.
7:12
It was the first study that had a correct prediction of a 12-minute residence time that was previously unreachable.
7:20
Time scales for molecular dynamics.
7:22
It's about targeting soluble epoxide hydrolase, which is involved in diabetic pain.
7:30
There were a lot of known inhibitors.
7:31
They all had similar affinity.
7:33
They had vastly different residence time and what the medicinal chemists were able to do given such a network is they actually found a dihedral rotation in one of the transition states, which are the states somewhere in between.
7:47
And then they designed binders that are confirmation constrained.
7:51
They avoid this dihedral rotation and it improved residence times by three times and it was later confirmed in experiment.
7:59
Something I'm personally more excited about is for a newer modality like PROTACs.
8:05
You can use this methodology even if you don't have a structure to predict the entire assembly.
8:11
E3 ligase, the target and the binder, of course.
8:15
And you can extrapolate from the simulations that we do, for example, the different degradation efficiencies.
8:23
And one cool thing that we now do relatively often is to involve HDX-MS or cryo EM data to inform these simulations, because HDX-MS itself only gives you something of a surface, not the entire structure.
8:39
But it can guide the simulations to arrive at a converged structure much faster than when you use vanilla MD or other tools.
8:47
All right, so to wrap it up, we think kinetics often matters more than affinity.
8:51
It's been proven many times.
8:53
And a method that we use with an ensemble or sometimes called enhanced sampling.
8:57
It makes the dynamics at biological time scales accessible today, gives you design insights in a matter of a couple of days instead of waiting weeks for one or two experiments.
9:10
And we showed utility of that method and it has been published as well by incorporating HDX-MS data.
9:17
For example, we do projects on a partnership basis.
9:21
I'm here today and tomorrow.
9:24
My Co founder Sam might be here tomorrow.
9:27
So talk to me, talk to Sam or shoot me an e-mail.
9:30
I'm happy to chat.