0:00
I will now introduce our next speaker.
0:03
Our next speaker will be Michael Bellucci and he's a Senior Director of R&D at XtalPi.
0:10
Doctor Michael Bellucci is a Senior Director at XtalPi Inc where he leads a research team that develops algorithm technologies for solid form development and risk assessment, as well as drug discovery.
0:23
Prior to his current role, he was research Scientist at MIT in the Novartis MIT Centre for Continuous Manufacturing, where he led many research projects in collaboration with pharmaceutical companies.
0:35
His solid-state research has focused on combining computational chemistry and AI machine learning methods to study processes and properties associated with molecular crystals such as solubility, morphology, surface adsorption, polymorphic transitions, and crystallisation.
0:54
Michael is a theoretical and computational chemist by education and holds a PhD in chemistry from Boston University.
1:02
I'll invite Michael now.
1:09
Thank you, everyone.
1:12
It's good to be here.
1:17
It's sunnier here than in Boston.
1:23
So today I'll talk to you about XtalPi's efforts over the past few years to transform solid form screening and selection from an art to a predictable science using our XtalGazer platform.
1:40
So in the first part of our talk, I'll provide an overview of our solid-state platform that we have built in a sort of abbreviated timeline.
1:52
So our company was founded around 2014, 2015 where we developed our first pharmaceutical solid-state R&D platform.
2:04
The original service that the company was founded on was Crystal structure prediction for polymorph de risking of pharmaceutical compounds.
2:15
But we always sort of had this parallel efforts both in drug discovery but also in experimental sort of work, experimental screening.
2:25
So over the past, I would say roughly 10 years, we've developed different capabilities in both experimental screenings.
2:36
So we can do polymorph salt screening.
2:38
We also have various different structure determination methods such as CryoEM and MicroED.
2:46
We also have added a number of different computational services, which I'll talk more about today.
2:53
And more recently since about 2020, we've started to develop our own automated experimental workstations, which is part of our XtalGazer platform.
3:07
So this sort of core issue here of XtalGazer is to try to address the issue of polymorphism and essentially deliver high quality polymorph screening while trying to mitigate the risks associated with polymorph transformations and what is the most thermodynamically stable polymorph for development.
3:31
And we hope to demonstrate that this essentially represents a paradigm shift in solid-state research moving away from more traditional based screening by trial and error through a sort of data-driven design-led methodology.
3:48
So sorry, shown here is an overview of our XtalGazer workflow which combines automated experimental screening with AI and physics based models.
4:08
So to begin, we require only the powder sample to do experimental screening of the compound.
4:16
And for computational screening we essentially only need the molecular structure of the molecule.
4:23
So going along from the top to the of this workflow, we first perform a sort of initial solubility assessment using either experimental methods or our AI generated methods.
4:38
And from this we determine whether or not there's good solubility in any of the solvents that we would initially screen.
4:46
If not, then we would sort of try to develop some sort of co-crystal or salt.
4:52
And to do this we can use our sort of virtual co-crystal or virtual counter ion screening.
4:59
This is AI based method that recommends different kinds of conformers or counter ions that would allow you to be able to crystallise a salt or co-crystal.
5:09
The experimental screening is then carried out using our Xtal Squared platform which uses our automated experimental workstations with recommendations of different crystallisation strategies from an AI and physics based models to perform the screening.
5:28
And then after we do this screening, we can characterise the experimental forms that we obtained through either single crystal X-ray diffraction or MicroED depending on the size of the crystals that we obtain.
5:42
But also in parallel, we can simultaneously predict all of the crystal structures or polymorphs using crystal structure prediction to derisk the forms that we get simultaneously.
5:58
And this forms a feedback loop.
6:01
So we can go from end to end in terms of development, but also we can de risk to see and make sure that we obtain the most thermodynamically stable form.
6:14
So now that we have a general overview of XtalGazer, I'll go into a bit more of the components of the different algorithms and how they can have an impact on real projects.
6:29
So shown here is our various different computational services and platforms.
6:36
Our oldest service is our crystal structure prediction that we use very regularly to do fully de risking of the polymorphic forms either for clients or internally for developmental projects.
6:52
We also have a CSP-Lite service where which is essentially just a faster version of CSP and the goal here is intended for a more early stage solid form screening to aid in lead optimization and selection of a molecule based upon whether or not has good properties, good physical properties.
7:16
And additionally, we also offer these virtual co-former virtual counter ion screenings which recommend different co-formers or counter ions for co-crystal or salt formation development.
7:29
And we also finally have now a morphology prediction service where we can essentially predict the macroscopic crystal shape of the polymorph in a variety of different solvents, also with or without additives.
7:48
And the goal here is to try to optimise the morphology to avoid unwanted morphological shapes such as needles or flakes, because they pose challenges for downstream processing.
8:01
Essentially, they require more expensive milling.
8:04
And the goal here is to try to design crystallisation experiments up front that allow you to avoid this.
8:12
So some benefits of the XtalCSP platform is shown here.
8:19
So from the CSP, we can predict the most stable crystal structures for free forms, hydrates, salts or co-crystals and then evaluate their solid form stability across a range of different temperatures to find the most thermodynamically stable form.
8:38
And this is a very important point because so I'm sure that many of you know that different polymorphs can change stability as a function of temperature.
8:48
They are essentially enantiotropic and almost all polymorphs that you find in pharmaceutically relevant applications are enantiotropic.
8:57
So this is a major aspect of the algorithm itself.
9:01
But I'll talk about this in a second.
9:03
So today we have de-risked over 300 different APIs for various different clients using our platform.
9:13
So to make CSP reliable for larger and more flexible molecules that are commonly developed in the pharmaceutical industry, we essentially had to improve upon the standard CSP algorithm when developing the XtalCSP platform.
9:30
So to do this, we had to essentially incorporate a collection of different algorithms that perform three major tasks, crystal structure searching or crystal structure generation.
9:42
And then this is followed by ranking and clustering of these crystal structures with highly accurate quantum mechanical methods.
9:50
And then finally, this is followed by stability evaluations across as the crystals as a function of temperature.
9:59
So the algorithm itself is dynamic and that it's governed by a decision tree that is learned from over 200 different cases.
10:08
And the decision tree essentially features, uses features of the API to select different algorithms and also force fields to perform the CSP.
10:18
So going from left to right in this figure, the algorithm essentially takes its input, the molecular structure and then it performs an extensive conformational analysis of the molecule.
10:31
And from these conformers we can generate a large pool of crystal structures.
10:37
The crystal structures are essentially this algorithm is all cloud based and one of the initial innovations of XtalPi was to be able to essentially form large scale high performance computer clusters on the family.
10:51
And doing this we can assemble up to 1,000 core HPC computer and we can screen up to a billion different crystal structures for a different molecule.
11:01
So it's essentially a truly exhaustive search.
11:05
But after we generate these crystal structures, we then rank each of these crystal structures through multiple energy rankings, first with our force field, then with progressively more accurate quantum mechanical methods.
11:17
And our final ranking is done with high precision DFT with dispersion corrections.
11:25
After we do this, we remove redundant configurations through clustering and this whole process essentially repeats until we get a converged energy landscape, which is the figure just before the right here.
11:42
So the energy landscape is essentially just a plot of the relative lattice energies versus their densities.
11:48
And this is the stability of the crystal structures at 0 Kelvin.
11:53
But to make predictions that are relevant, we basically then extract these low energy polymorphs from this energy landscape and then we perform free energy calculations using molecular dynamics over a range of different temperatures.
12:10
And so this is done with a more, it's different than the traditional way it's done with the Einstein crystal thermodynamic cycle.
12:19
And I should also mention that the force field that we generate is tailor made for every single molecule.
12:24
It learns from all of these crystal structures throughout this process.
12:27
So it's tailor made, and it's trained on essentially millions, if not billions of different crystal structures in the process.
12:38
So shown here is an example of a case study of how CSP and MicroED can be used to fully de risk remdesivir.
12:48
So we performed this study on an accelerated time frame during the pandemic when remdesivir was thought to be effective against COVID-19.
12:58
We published the results, however, because we wanted to demonstrate how this combination of tools can be used to fully de risk a compound that is complex in just 33 days, which at the time was kind of imperative as everyone was trying to develop something that was effective against COVID-19.
13:17
So on the left, we show the energy landscape, which is just a plot of the relative energies of the crystal structures versus their densities at 0 Kelvin.
13:29
And so each one of these points is a crystal structure that has a corresponding lattice.
13:36
And if we take this and we overlay it with the structure that we solve from MicroED, we can essentially verify and validate which crystal structure and its energy the experimental structure is and the energy landscape.
13:49
So this inset here is actually an overlay of a solved MicroED crystal structure versus the predicted crystal structure.
14:01
It's difficult to see, but you can kind of see if you that there's an overlay, you can see that there's a very good agreement.
14:09
But from this you can essentially identify what is the stability of the polymorph in the energy landscape.
14:16
And from this, we find that the form 2 and form 4 crystal were the rank 2 and rank 1 crystal structures that we predicted.
14:25
And then from this, we took essentially the lowest 5 energy crystal polymorphs, and we performed free energy calculations over a range of different temperatures as shown here.
14:36
These are relative free energy calculations relative to the X1 or Form 4 experimental crystal structure.
14:44
And so, what we see from this is essentially that temperatures roughly below 250 Kelvin, the Form 4 crystal is the most stable, but above that temperature, the Form 2 crystal structure is the most stable.
14:58
This is an example of an enantiotropic relationship between polymorphs and why it's important, for example, to incorporate free energy calculations into your CSP.
15:06
Otherwise, if you just based your stability rankings based upon the 0 Kelvin landscape, you would erroneously conclude that Form 4 should be the developmental form.
15:17
And this, these kinds of relationships can be very complicated.
15:22
We observe them all of the time.
15:24
So it's just sort of highlights the point of why these components of the CSP is important.
15:31
So in this slide, we show an example of how our CSP-lite can be used to support lead optimization.
15:39
So this study was performed in collaboration with AbbVie.
15:43
So Abby had eight similar molecules as a series, and they were interested in understanding the polymorph landscape for each of these molecules as well as specific physical properties such as solubility.
15:56
So in the figure on the right, we show a plot of the kinetic solubility which is shown in these white diamonds in the thermodynamic solubility which is shown in black dots.
16:07
So the thermodynamic solubility uses information of the crystal structure that we generate from the CSP light to provide more accurate calculations of the solubility, whereas kinetic solubility only uses information from the molecular structure itself.
16:23
Kinetic solubility is often used like in drug discovery when you don't really know the crystal structure and it overestimates essentially the solubility.
16:32
And this was AbbVie’s point.
16:33
That's why they were interested in it.
16:35
They wanted to know what is the truth, thermodynamic solubility when making a decision early on in formulation to make better decisions about what they should develop.
16:44
They're also interested in like how many polymorphs a different molecule had because you've changed certain parts of the molecule and then all of a sudden you don't have that many polymorphs.
16:54
So they combine this with various different properties that they provided, such as ADME toxicity, properties, toxicity and permeability.
17:04
And so using these properties combined, they would make this as a decision about what to develop for solid formulations.
17:12
And so this is sort of the points of the CSP-lite.
17:16
It's not as for like thermodynamic de-risking, it's more about trying to figure out what are good properties that you should use early on for development.
17:28
So in this slide, we show an example of our morphology prediction platform.
17:33
And so this work was carried out with Merck Germany.
17:38
And the objective here was to determine if we could predict the effect of polymer additives on the morphological shape.
17:46
So from this study, we used metformin hydrochloride was chosen as the API and experiments were carried out in two different solvents with five different polymer additives at low concentrations, which is shown in this table on the top right.
18:02
So from this study, really the goal was to figure out if we could predict the overall morphological shape.
18:08
And so Merck Germany provided these polarised like microscopy images to sort of validate the project here.
18:17
But from this we can see that we can predict the overall morphological shape of these crystals under these various different conditions using different polymer additives.
18:27
And what we found from this was that essentially metformin hydrochloride in pure solvents is a very like highly anisotropic needle, as you can see in these figures here with just pure methanol or IPA water mixtures.
18:42
But if you change and you add just a bit of polymer additives, you can, especially HPMC, you can get these drastic morphological changes from a needle to a prism.
18:55
And so this for example, is an example of how you can redesign your crystallisation or just add very small quantities of polymers that would change your morphological shape and prevent you from having to do expensive things like milling.
19:09
But this is essentially the goal of that project work.
19:14
Another component of our XtalGazer platform is the Xtal Squared Polymorph screening platform, which combines physical models and AI models to recommend crystallisation strategies.
19:28
And so these experimental strategies are then carried out by robotic automated workstations and they results in crystal structures that are solved by either single crystal XRD or MicroED.
19:40
And we can also do the de-risking with the CSP at the same time.
19:46
So these workstations are essentially, they look like this as sort of like a cabinet like workstations and they're essentially operated by a robotic arm and supervised by our scientists.
20:02
So they're highly consistent.
20:04
They're also standardising crystallisation and removing human error.
20:08
But we can also from this essentially capture data in real time and record it in electronic notebooks, which can then be fed back into an AI algorithm to refine the crystallisation prediction strategies.
20:22
We should mention though that we don't capture this for commercial cases, just for public data or cases that we do internally to build these data sets.
20:32
But the crystallisation in these workstations can be done in eight millilitre vials and then it can use a variety of different techniques such as evaporation crystallisation, slurry crystallizations, rapid precipitation such as anti-solvent addition or reverse anti-solvent addition, cooling crystallisation or vapour diffusion.
20:54
So it covers a lot of sort of various techniques but also can make these recommendations.
21:04
So one of the issues when we were developing the AI for this was that generally there isn't a lot of data for you to use to train AI to make crystallisation strategies efficient.
21:19
So we took a different approach in this case, and it's proven to be useful in a number of different areas.
21:26
So essentially what we do here is we predict the crystallisation outcomes from different strategies and conditions using physics based models such as classical nucleation theory or two step nucleation theory.
21:39
And these physical models themselves aren't perfect, but what they do get right generally are the trends.
21:46
And it's AI that picks up essentially on trends.
21:50
And so, what we do initially is we build a large database of virtual results and conditions, We train an initial AI and then it's refined on actual experimental data using 10,000 data points.
22:03
So you start off with the sort of in the known parameter space for the AI and then you refine it so that it gives you good predictions.
22:11
And so here we're showing a case study where we show a comparison of the timelines and materials used for an internal double-blind case study that we did.
22:23
So here crystallisation experts, essentially humans were performing screaming of different compounds using 200 different experiments.
22:34
It took about 3.5 weeks and required a four gram sample and from this screening there was a crystallisation hit rate of about 87% and three forms were discovered.
22:47
But in comparison, the same compound was screened using our Xtal squared crystallisation platform and supervised by a single scientist.
22:57
And only 70 experiments were needed and recommended by the AI itself and the then they were carried out by these automated workstations.
23:08
But from this screening we can see that we can achieve a better crystallisation hit rate.
23:13
We also discovered all of the known forms.
23:16
It was done in less time, with less sample and with less human resources.
23:22
Another case study is shown here where we show the corresponding molecules.
23:29
So in each of these cases, the Xtal Squared platform outperformed the corresponding human scientists, which is sad to say, but it's so.
23:37
This traditional mode is the column that represents the human results.
23:43
So essentially in this we achieve a better hit rates.
23:48
We also get less sample consumption, but also significantly less time.
23:53
If you compare the time consumption from each of the experiments carried out, this is one example of crystallisation, but it becomes actually more apparent in various different cases, like particularly when you're trying to develop a salt or a co-crystal and there's more parameters.
24:12
So this is my final example.
24:14
So here we show our results from virtual salt screening.
24:18
And so this is in comparison also to traditional mode done with humans.
24:23
And so the API under consideration was Meloxicam and the specific counter ions and solvents that were used in the screening is also shown here.
24:32
So it essentially in this column where it says results.
24:35
If it's a plus, it means that it led to a salt.
24:37
If it's a negative, then it means that it didn't lead to any sort of crystal.
24:44
So and basically in comparison to the traditional mode, we achieved 22% better hit rate and in certain cases we were able to achieve a salt using a counter ion that the traditional mode was not able to.
25:00
And this was a result of the solvent selection, since it's not obvious which solvents to use for a given counter ion in experiments.
25:09
Usually in the traditional screening approach the solvent is more or less fixed and you change the counter ion.
25:16
And this is often solvents chosen based upon solubility.
25:20
But in this case, the AI itself can recommend not only the counter ion, but the solvents that would go along with those crystallizations.
25:28
So as you can see, as sort of you add more parameters that make it more difficult that the number of experiments you have to do traditionally explodes, the AI becomes even more efficient at predicting what would be a good crystallisation route.
25:43
And so in conclusion, I hope I was able to give you an overview of our solid-state capabilities.
25:48
Thank you all for coming to the talk.
