It is difficult to read through any news source without being inundated with the phrases artificial intelligence (AI), deep learning, and machine learning. There’s no doubt that in some fields, these computational concepts will be transformative.

In the drug discovery and design literature, however, searches for those same phrases often link to articles discussing more familiar terms like neural networks, linear regression, quantitative structure-activity relationships (QSARs), and computer-aided and structure-based drug design. So, how does AI fit into these other computational approaches, and is the buzz more than hype?

During their graduate work, AtomWise founders Izhar Wallach and Abe Heifets realized that the same types of architectures and approaches being applied to image recognition could be modified to perform molecular recognition.

credit: atomwise

“It's getting ridiculous in terms of so many people connecting with this buzzword-compliant phrase, AI, which just almost has become meaningless,” opined Andrew Radin, co-founder and CEO of Aria Pharmaceuticals (formerly twoXAR Pharmaceuticals) .

Echoing the sentiments of several people in the industry, Radin views AI more as an evolution of existing technologies rather than a revolution of innovation.

“Because artificial intelligence has gotten a lot of news lately, [people] think it's new,” he said. “These algorithms, some have been around for decades; this class of technology started back in the 50s.”

He pointed to logistic regression, which is built into Microsoft Excel and yet is an important part of the AI toolbox.

“I don't know where, but certainly somebody has used it somewhere along the progression of doing drug discovery and development,” he mused. “But no one tied that to a claim of being an AI-discovered drug.”

Melissa Landon, Chief Strategic Officer at Cyclica, offered similar insights.

“A lot of computational scientists who worked in pharma their whole careers sort of laugh at this whole craze over AI and drug discovery,” she said. “They're like, ‘People, we've been doing this for decades. A QSAR model is AI. It’s a machine learning model.’”

Whereas AI and machine learning might not be new, the scale and complexity of what can be accomplished has certainly changed in recent years.

“Up until the advent of AI systems, computational approaches in drug design, docking, and screening have merely been an adjunct to human decision-making,” offered Andrew Hopkins, founder and CEO of Exscientia. “Interpreting complex multi-dimensional data comprehensively is beyond human-led endeavour, but well-designed AI systems are able to accomplish this.”

“AI generative methods are providing a step change by automatically designing new compounds and optimizing them against a complex set of drug discovery parameters,” he pressed.

The ultimate goal is to minimize human intervention to the point where systems become completely autonomous. However, there is still work to be done, Hopkins said.

Although Exscientia can design project-aligned molecules efficiently, for example, knowledge gaps remain in areas such as synthetic pathways for molecules. In the meantime, the company does its best to combine the strategic insights of drug hunters with the tactical efficacy of AI design and optimization.

What led to this expanded capacity for discovery? Data has become more plentiful and much more data is available in the public domain.

"We spend a lot of time making sure the models aren't cheating. We explicitly devise tests that say, if you are good at this test, you're probably cheating."
– Jon Sorenson, Vice President of Technology Development, Atomwise

Credit: AtomWise

“Pharma has always been sitting on a ton of data, and in my view, this is also one of the impediments to AI or any kind of computational model really transforming industry,” Landon noted. “You can build a local QSAR model for certain things with a relatively small amount of data, but if you want to do true deep learning, you need a lot of data.” Equally important has been the growth of academic expertise in AI and machine learning, largely the result of efforts by companies such as Google and Amazon.

“In terms of methods and more sophisticated deep learning techniques, ways of training on sparser data, all these things are now expanding the domain of applicability of machine learning,” Landon continued. “Thinking about new ways of representing data and new computational methods for building models, we're borrowing a lot from other industries, and that's really helped us.”

Jon Sorenson, Vice President of Technology Development at Atomwise, pointed to advancements in other fields like structural biology, suggesting that they too facilitate the growth of AI.

“The PDB [Protein Data Bank] has finally gotten to a much larger comprehensive set of structures, as well as just the ability of the community to aggregate data and put it in places where it’s more uniform and more accessible through efforts like ChEMBL or PubChem or various other curated databases,” he noted.

Digging into data

Beyond simply getting access to data, however, a second challenge comes with the reality that biology is messy, and typically, the data on which machine learning models are developed can be quite heterogeneous. One way to deal with that heterogeneity is data standardization.

Landon highlighted participation by Cyclica co-founder and CEO Naheed Kurji and others in a group known as the Alliance for Artificial Intelligence in Healthcare (AAIH). Among other initiatives, the advocacy group establishes standards for data used to construct models and develops mechanisms to evaluate the quality of the models based on that data.

The other alternative, Landon continued, is for researchers to simply accept that the data is messy and build from there. That’s never going to go away, she added.

“Even though there's a lot of data out there for building machine learning models, the reality is that only a fraction of that data is used to build models, because only a fraction of it is considered good enough,” she pressed.

To that end, Cyclica has been working on ways to improve the quality or utility of messy data. In fact, said Landon, models are often better when they’ve incorporated the messier data as opposed to being constructed using only the data defined as high-quality.

According to Radin, Aria has a similar mindset, suggesting that maybe it’s not about creating valuable data so much as extracting value from the data that exists.

"If you only train [your model] on positive data, you can answer maybe the question of whether or not something binds, but if you add in the negative data, you can answer the more fundamental question that's going to drive decision making of what's important chemically for that interaction."
– Melissa Landon, Chief Strategic Officer, Cyclica

Credit: Cyclica

“Our approach has been to build a proprietary AI system that can deal with heterogeneous data, which is not what AI systems are built to do,” he explained. “What that allows us to do is pull in these massive data landscapes where now we can inspect biomedical data across the spectrum and then use that to find the key measures of biology that are going to make an impact on diseases that other people would overlook or would miss.”

Sorenson went a step further, suggesting that there was also tension between data quality and what he termed data diversity or data generalizability. Pharma companies may be sitting on a wealth of data, he said, but the data often only represent the programs being developed in-house and therefore are a small subset of what is available.

“You have a high-quality endpoint, a consistent endpoint, but it's going to predict best within the direction that you are already going in the series,” he explained.

Thus, Atomwise and others are combining data from several programs into a single model.

“While the endpoints themselves might have more noise, there's more hope of generalizability,” Sorenson noted. “You need to be able to combine all these different compound series, all these different functional groups, and put that together into a unified picture.”

This generalizable model approach allowed the company, for example, to identify potential therapies for a rare disease despite the absence of a molecular structure for the target enzyme. They constructed a model for this project based on principles derived from previous models they had generated. Although the details may be specific to a complex, the biophysics of ligand-protein interactions are more generalizable.

Radin similarly suggested that Aria’s focus is less on quality than on quantity of data, and their key measure of quantity is statistical independence, meaning that they cover different information.

“If we're going to incorporate yet another data set into our platform — we’ve got more than 60 of them, now — that new data set has to be orthogonal and has to have some signal in it that the other 59 don't; otherwise it doesn't bring us a lot of value,” he explained.

Because Aria’s system is blinded to known treatments and experimental candidates, researchers can watch for those molecules to pop up as highly ranked or highly probable, taking that ranking as a sign that the scientists have generated a solid, predictive model. In such cases, they may not know which highly ranked candidate will be a true hit, but they are confident one or more of them will be.

Where the focus of drug design has always been on-target, Cyclica is focused on the polypharmacology of leads, believing that it is also important to fully understand the off-target risks.

Credit: Cyclica

Hopkins went so far as to describe Exscientia’s systems as data-agnostic, able to use data from a variety of sources, including experimental assays. This approach allowed the company to design compounds against molecular targets in one scenario and phenotypic endpoints in another.

“We see each drug discovery project as a learning problem where there will never be enough data available initially, especially for first-in-class projects,” Hopkins added. “The challenge therefore is how to learn quickly to improve the models with the smallest amount of data.”

Null not void

One type of data that has historically been unexplored but is seeing increasing relevance in AI-based modeling is negative data, where molecular interactions or phenotypes have been experimentally shown not to occur.

Building machine learning models exclusively with positive data limits what researchers can do with them. Cyclica has found that adding true negative data to a model can significantly improve it, perhaps not in terms of its performance, but in terms of the interpretation of its output.

Landon recounted a talk given in May by Astex Pharmaceuticals’ senior director of computational chemistry and informatics, Marcel Verdonk, at the Artificial Intelligence for Early Drug Discovery meeting that highlighted the importance of both positive and negative data.

“They're doing fragment screening,” she recalled. “What they were trying to do is predict not just whether or not a fragment would bind to a protein, but what about the fragment was important for binding the protein.”

“If you only train on positive data, you can answer maybe the question of whether or not something binds,” Landon recollected. “If you add in the negative data, you can answer the more fundamental question that's going to drive decision making of what's important chemically for that interaction.”

Sorenson echoed the importance of models built on both positive and negative data, describing one such approach as “actives as decoys.”

Rather than simply define a molecule as active because it binds to a given target, Atomwise’s system also tests that molecule against all other targets, which act as presumptive negative controls. This prevents the model from describing the molecule as always active.

“That really helps condition and balance the model to not get so excited just about the ligand itself,” Sorenson explained. “It really has to understand the full story.”

Another way Atomwise prevents their models from getting caught up in the hype takes the opposite approach. Rather than swap in other targets for a given ligand-target model, researchers take what appears to be the best docking pose, which is presumably the active configuration, and compare that to the worst pose, which it defines as a bad example of active binding.

“Instead of just memorizing the fingerprints of the protein and the ligand, you really need to understand that this interaction is what we're looking for,” Sorenson added. “When the pose is bad, you should recognize it as bad.”

“What we're really trying to do is generalize beyond the space that we trained in,” he said. That also includes efforts to make sure the models don’t cheat.

When you're training the model, he continued, the models will see patterns that you didn't want it to see or didn't expect it to see, such as dataset imbalances in molecular weight or functional groups. In such cases, the model may determine that heavy compounds are active, whereas lighter compounds are inactive, simply because the literature bears out this pattern.

“We spend a lot of time making sure the models aren't cheating,” Sorenson explained. “We explicitly devise tests that say, if you are good at this test, you're probably cheating.”

Once the models are generated and trained, however, the real test begins as the researchers work to explore real-world questions. Given that many AI companies have expanded beyond a software-as-a-service model to a pipeline-based model, the search is on for therapeutic candidates.

For Cyclica, that search begins with a consideration of the worst outcome for a putative drug candidate: off-target effects.

Learning into practice

From the outset, Cyclica’s core fascination has been with the challenge of polypharmacology.

“The reason all drugs have side effects is because they're engaging multiple proteins in our body, and this is what is referred to as polypharmacology,” she continued. Offering statistics, Landon suggested that the average drug interacts with anywhere from 30 to 300 proteins in the human body.

“We know that when a human takes a pill, it does a lot of things in your body,” she said. “It hopefully does the thing that it was designed to do — lower your blood pressure or whatever — but it also does many other things, and that is evidenced typically by the long list of side effects.”

“Much of the reason why drug discovery programs fail, at least at the early stage, is because essentially we've failed to address the polypharmacology problem,” she said, adding that this was the indirect result of a drug discovery approach focused on defining a specific target associated with a disease outcome.

“For example, I know if I inhibit Protein X, I will cure Condition Y,” she explained. “Then you develop a drug molecule that modulates Protein X’s activity.”

“You focus on Protein X,” she continued. “You only assay for Protein X. You optimize for the molecule’s function against Protein X.”

This process alone, she noted, can require several years in a discovery process, before moving into disease and animal models, when new challenges may arise.

“You see that when you dose an animal, for example, you have some toxic side effect that you weren't anticipating, and oftentimes it's the failure to address those challenges that result in a program essentially dying,” Landon added.

Understanding what those potential liabilities are and what the possible downstream challenges might be allows scientists to start addressing them earlier in the discovery phase.

To probe polypharmacology, Cyclica developed a deep learning engine called MatchMaker, which analyzes protein-drug interaction data at a human proteome-wide scale, approximately 9000 proteins.

“From that data, we are trying to learn what features of a molecule or what features of a protein dictate the interaction of the two,” Landon explained. “So, given a set of molecules, we are predicting essentially the polypharmacological profile, and this is housed in our Ligand Express platform.”

She compared this approach to the six years she worked at Schrödinger and its portfolio of computer-aided drug design tools.

“There, the focus was on the single-target mentality,” she recounted. “Schrödinger has developed some exceptional methods for understanding, at a finely detailed molecular level, how a molecule interacts with a protein, but really, there was no thought given to what else this molecule might be doing.”

Predicting off-target effects of drug candidates and mechanisms of action was the initial focus of Ligand Express, she added. Cyclica has since evolved their approach to include the design and optimization of novel molecules with improved polypharmacological profiles and reduced liabilities, a process encapsulated in their Ligand Design platform.

Finding a (re)purpose

According to Landon, the concept of polypharmacology is highly linked to repurposing. “Because we have the ability to screen on the [whole] proteome level, we can fairly rapidly identify potential repurposing opportunities,” she explained.

“If you look at most of the successful repurposing endeavors, that's mostly been done from a biological standpoint, meaning that we know that this pathway is also implicated in this disease area,” she added. Repurposing strictly on the basis of molecular interactions — that is, between ligand and protein — is a bit more challenging because you have to understand what the chemistry is doing.

The arrival of SARS-CoV-2 in late 2019 provided the perfect impetus to leverage Cyclica’s polypharmacology expertise. Like many other research groups, in the early months of 2020, the company initiated a study of potential repurposing opportunities versus COVID-19.

"Up until the advent of AI systems, computational approaches in drug design, docking, and screening have merely been an adjunct to human decision-making. Interpreting complex multi-dimensional data comprehensively is beyond human-led endeavour, but well-designed AI systems are able to accomplish this."
– Andrew Hopkins, founder/CEO, Exscientia

credit: Exscientia

The project resulted in the development of the PolypharmDB, a database of molecules that have proceeded through at least Phase 1 clinical trial development and which they profiled against both the human proteome and the coronavirus proteome.

The work led to two studies published last year by Dar’ya Redka and colleagues at Cyclica and earlier this year by Ryerson University’s Michael Sugiyama and colleagues.

Using Ligand Design, Redka and colleagues screened a library of 10,224 drugs from DrugBank against 8700 human and viral proteins (1). The resulting PolypharmDB included the predicted binding profiles of 2118 approved drugs, 2242 compounds in clinical trials, and 5547 molecules with preclinical data or defined as nutraceuticals.

The researchers then scanned PolypharmDB for molecules that might bind the viral protease 3CLpro or disrupt the ACE2-spike interaction by binding either protein alone or the interface between the proteins. This produced a list of candidates that included drugs known to be involved in the mTOR-signalling pathway, which targeted 3CLpro, and antibiotics in the cephalosporin and penicillin families, which targeted the ACE2-spike complex.

They also scanned for compounds that might interact with host proteins critical to viral infection and life cycle, focusing on TMPRSS2 and cathepsin B. The analysis resulted in several candidates including one predicted to bind both targets: the anti-arrhythmic agent disopyramide.

In contrast, Landon explained, the work with Ryerson University involved the marriage of two machine-learning approaches (2). The first used graph convolutional networks (GCNs) to deeply probe the viral-host interactome and identify targets that might have therapeutic potential for COVID-19. The second was a repurposing analysis as performed earlier.

Screening with either machine-learning approach or by cross-referencing GCN-prioritized targets with PolypharmDB resulted in the identification of 26 possible repurposing leads, which they then tested in cell-based immunofluorescent infection assays. Four candidates showed strong potency and the orally available MET inhibitor capmatinib offered low micromolar potency across multiple coronavirus species.

“While we have not tested the action of capmatinib against SARS-CoV-2 variants that have emerged in 2021, the broad antiviral activity that capmatinib exhibits against 5 genetically different human coronaviruses (229E, OC43 and NL63 live virus infection and SARS-CoV-1 and SARS-CoV-2 pseudotyped virus assays) suggests that capmatinib may hold promise in broadly treating SARS-CoV-2 variants of concern, or other variants that may arise from further antigenic drift as well as other emerging viruses,” the authors wrote.

Although never questioning the importance of such repurposing efforts, Atomwise’s Sorenson took a moment to muse about the irony of the current COVID research.

“The going gospel a year-and-a-half ago was that you shouldn't go into infectious disease because no one would pay for it,” he laughed. “Obviously, a year-and-a-half later, everyone is paying for development in infectious disease. But classically, it looked like that was an orphan area.”

In many ways, however, the rapid impact of such COVID-19 efforts helps make the case for AI-empowered efforts in other orphan areas, such as rare diseases, where companies have historically been reluctant to invest resources into developing drugs that might only serve a market of several hundred individuals.

Rare but life-changing

Last year, a study conducted by Sorenson’s Atomwise colleague Adrian Stecula and collaborators at the University of Toledo highlighted the ability of AI-based screening to identify therapies for rare diseases (3).

Specifically, the researchers used homology modeling and AtomNet to address Canavan disease, an inherited condition effecting hundreds of people in the United States and prevalent among the Ashkenazi Jewish population. In Canavan, a mutation leads to an accumulation of the metabolite, N-aceytl-L-aspartate (NAA), in the brain. The condition can lead to blindness, diminished motor skills, and death.

Rather than focus their efforts on the triggering mutation, the researchers searched for compounds that might inhibit the enzyme directly responsible for NAA synthesis, aspartate N-acetyltransferase (ANAT). The first hurdle to overcome was the lack of a protein structure for ANAT. A scan of protein databases only identified enzymes with low sequence identity. Faced with limited opportunities, the researchers constructed a comparative model using an enzyme with catalytic activity similar to ANAT.

The researchers used the model and AtomNet to virtually screen a library of approximately 10 million commercially available compounds from Mcule, repeatedly narrowing the field based on various pharmacological parameters until they identified 60 top-scoring compounds.

The researchers identified five chemically diverse molecules that offered low-micromolar potency against ANAT. Further studies to characterize mechanisms of action are ongoing, and the researchers suggested the chemical scaffolds of the candidates offered opportunities for further pharmacological optimization.

“We show that AtomNet is capable of discovering novel active scaffolds even when on-target data are scarce or completely unavailable,” the authors wrote in conclusion. “Second, we demonstrate that discovery is possible even without the availability of a 3D crystallographic structure or a high sequence identity homologous template.”

Speaking more broadly, Sorenson suggested that such efforts are just the first step. The next goal is to discover compounds with low-nanomolar efficacy, at which point, he jokingly added, the real work starts.

“As soon as you solve that problem, tack on a couple more off-target liabilities and then all of your ADME suites,” he said. The next challenges could include issues with Caco-2 permeability, hepatic clearance, fraction-unbound studies, and solubility, for example.

It is not enough, he pressed, to find the potent needle in the haystack. Rather, you need to find the needle that is potent, selective, and exhibits few or no ADME liabilities.

“To me, if we execute on that, that just changes the whole landscape of drug discovery and drug development,” Sorenson said. “It really makes it a lot more tractable for a lot of diseases that aren't necessarily getting the attention that they should get.”

Us and them

Although AI is seen as revolutionary in many circles, that hype has also come with a price tag: fear that humans will be replaced by machines.

“Certainly, a big fear for adoption of AI and drug discovery is this fear of taking the human out of the decision-making process,” Landon said. “At best, these methods can learn from complex high-quantity data in ways that our brains simply can't.”

This, however, was a way in which she believed AI might actually empower researchers to do what they do best: think critically. Rather than figuring out how to generate the model to perform the experiments, researchers can try to figure out what to do with it. It reminded her of her post-doctoral fellowship as an enzymologist.

Jon Sorenson, Vice President of Technology Development, Atomwise

Credit: AtomWise

“What I really hated was I spent all of my time preparing reagents and experiments,” she recounted. “The part of it that was fun in terms of taking the results and doing thought work with that data, that was actually only a small amount of the work.”

It’s important to remember that the software itself is the tool used by an artisan, suggested Radin.

“You don't walk into the custom cabinet maker shop, look at the dining room furniture, and say, ‘I'd really love to see your saw,’” he laughed. “‘Can you explain to me how your miter saw works?’”

He suggested that the industry’s focus on the tool rather than its output has divorced the conversation from the reality of bringing drugs to market.

Sorenson echoed this sentiment. “We're at a stage where AI is really augmenting what a medicinal chemist and maybe what a traditional drug discovery pipeline is doing,” He continued, “We're really not at a stage that we can claim that these AIs have taken over.”

You’re not going to go from virtual screening to IND filing with a purely AI-driven approach. In Sorenson’s opinion, the endpoints simply aren’t sufficiently tied together for it all to work in a machine-learning model.

“Can an AI come along and add in extra information about the patterns and help see through the noise and the multiple parameters that you're trying to optimize for?” Sorenson asked. “I think the answer has to be yes.”

Radin added, “Now that it's actually economically viable to use this tool, how am I going to use it in a creative and interesting way? That's the underlying thrust behind these trends.”

The challenges of AI-based drug discovery and design haven’t necessarily diminished over time, and Radin acknowledged that our ability to measure biological systems is still very limited. He compared it to throwing a piano down a flight of stairs, finding a piece of wire, and determining that pianos are made of wire.

“You have this little piece of information here, and you've got other little pieces of information,” he said, “and you're trying to reconstruct the piano, which you can never see, touch, or feel.”

That said, he was confident that the field is getting better at evaluating all these disparate signals and generating better models and hypotheses of what might provide disease-modifying activity.

The interplay of intelligences — both human and artificial — will be key to that success.

References

1. Redka, D.S. et al. PolypharmDB, a deep learning-based resource, quickly identifies repurposed drug candidates for COVID-19. chemRxiv DOI: 10.26434/chemrxiv.12071271 (2020).
2. Sugiyama, M.G. et al. Multiscale interactome analysis couple with off-target drug predictions reveal drug repurposing candidates for human coronavirus disease. bioRxiv DOI: 10.1101/2021.04.13.439274 (2021).
3. Stecula, A.; Hussain, M.S.; Viola, R.E. Discovery of novel inhibitors of a critical brain enzyme using a homology model and a deep convolutional neural network. Med Chem 62: 8867-8875 (2020).