The failure of many potential candidates during clinical trials significantly impacts the cost of drug development. Computational drug design has the means to lower these costs by allowing not only more structural and functional simulations prior to synthesis but also more virtual absorption, distribution, metabolism, excretion, and toxicity (ADMET) and pharmacokinetics work before investments are committed to animal tests and clinical trials.
Computer simulations will never replace experimental benchwork. The vision is an iterative process—virtual design then experimental validation—initiated at the beginning of therapeutic discovery. These simulations require massive computing power. Companies address this hurdle by either building up in-house computing environments or by utilizing cloud-based platforms. The latter option is available to companies of any size, leveling the drug development playing field.
Although computational drug design is still an emerging field with only a handful of players, the transformative nature of these platforms has inspired many facilitators, including regulatory agencies, that recognize the field’s potential.
Search and property calculation capabilities are key concepts in computational drug discovery. “Drug discovery tries to find molecules with an efficacious effect,” says Anthony Nicholls, PhD, corporate vice president at OpenEye, Cadence Molecular Sciences. (The company was formerly known as OpenEye Scientific Software, and it is now a business unit of Cadence Design Systems.) “It would be easy if there were only a few hundred molecules, but that number is close to infinity—and many may have an effect.”
With a history that reaches back more than 25 years, OpenEye and its initial downloadable software applications and toolkits have long employed physics-based approaches to give compounds real-world physical properties for 3D viewing, movement, and fit-to-target attributes. Initially, OpenEye focused on increasing the speed and scale of its search capability to screen millions of compounds instead of tens of thousands.
“If you have a good assessment metric and expand the search, eventually you will find more good candidates,” Nicholls explains. Having decided that cloud computing would be able to add many more zeros to the number of compounds that could be screened, OpenEye introduced the Orion software cloud platform for molecular design and simulation in 2019.
After compound identification, the next step for computational and medicinal chemists is to calculate the properties that potentially position the compound to become a drug. This set of calculations often involves quantum mechanics and molecular dynamics.
“You have ‘big’ on one side because you are screening a lot of molecules, and you have ‘big’ on the other side because you are doing a lot more physics calculations on a few molecules,” Nicholls points out. “Both are very cloud friendly. The scale at which we operate differentiates us.” Cloud access also opens up opportunities for companies that desire minimal in-house footprints. (An example of a company that uses OpenEye’s Orion for drug discovery is Mirati Therapeutics.)
Nicholls has a long bucket list. For example, he hopes OpenEye will improve dynamics simulations to identify new places that molecules might bind, and to address crystal polymorphs to predict dangerous polymorphic shifts during compounds’ shelf lives. “Science has come a long way. It is almost possible to supply answers in advance,” Nicholls relates. “We want to make drug discovery more predictive, as well as help lower the cost of medicines and speed the drug development timeframe. We see computations and our relationship with Cadence as the way to make that happen.”
Faster and more accurate
Determining the structure of a compound complexed with a target—a protein, DNA segment, or RNA transcript—and establishing its function are critical in structure-based drug design. X-ray crystallography or cryogenic electron microscopy (cryo-EM) approaches are often used to experimentally determine structures, and artificial intelligence (AI) approaches, such as AlphaFold, are used to predict structures when targets of interest lack experimental structures.
“Our mixed linear-scaling quantum mechanics/molecular mechanics (QM/MM) tools in combination with experimental data can more accurately determine experimental structures or refine predicted structures,” says Lance Westerhoff, PhD, president and general manager, QuantumBio. “Then our patented XModeScore technology is used to more accurately determine protonation, rotamer, and stereoisomer states.”
Besides determining structures, the company’s software can predict functions and assess binding characteristics such as affinity, free energy of binding, and pairwise energy decomposition. “Using a structure for one drug candidate and its target, we can pair a novel candidate with the same target and predict the binding characteristics, often without going through the expense of structure determination,” Westerhoff asserts. QuantumBio’s patented MovableType method is designed to use free energy simulation to improve speed and robustness.
Software can be accessed through standard licensing as a standalone or as a plug-in to the Molecular Operating Environment (MOE), which is a drug discovery software platform developed by Chemical Computing Group. Also, the free energy MovableType calculations are available through GridMarkets Pharma, with the crystallography and cryo-EM tools coming soon.
“AI and ML [machine learning] are very important for both structure determination and function prediction,” Westerhoff notes. “For example, we can take predicted structures and join them with experimental data to make them more accurate. For function predictions, we can provide data to be used within ML or AI models.”
According to Westerhoff, the modeling field is highly synergistic. Software packages—even if they are from different vendors—act as cogs in a machine that can solve different sets of problems or the same problems with different methodologies. Each tool has strengths and weaknesses.
“Some tools provide very good starting points, which we can then process using more calculations to determine a more accurate picture, while others can be used after we analyze the data,” Westerhoff explains. “Our approach comes at the problem a different way and provides similar results in a drastically reduced timeframe.”
Safety and therapeutic efficacy
ADMET machine learning and simulation offerings from Simulations Plus allows prediction of key drug dynamics for safety and therapeutic efficacy assessments. In a discovery setting, the company’s ADMET Predictor and GastroPlus platforms combine AI (or, more specifically, ML) with physiologically based pharmacokinetics to create a virtual human or animal model. The technology integration enables chemists to determine if a potential compound will be active with respect to the target, and to predict concentrations systemically in the desired tissues or organs.
“Our platforms are used before in vivo animal testing to help evaluate and optimize the combination of events for the molecule to have the correct concentration profile and be absorbed, metabolized, distributed, and excreted like predicted to elicit the therapeutic outcome with minimal safety risks,” relates John DiBella, president of the Lancaster division of Simulations Plus.
Optimizations performed first in virtual small animal models and followed by in vivo validation build confidence for the extrapolation of findings to humans. Subscription-based licenses are available, or the platforms can be hosted on cloud environments. Simulations Plus also has a team of experts for project-based activities. “We also embrace contract research organizations and invite them to license our software to create their own modeling and simulation offerings,” DiBella notes.
Modeling approaches such as quantitative systems toxicology and quantitative systems pharmacology can be used to create detailed bottom-up models of tissues and organs for predicting changes in safety or efficacy biomarkers. “We are only scratching the surface of where computational technologies can go,” DiBella says. “The FDA and other regulatory agencies are facilitating adoption, and modeling technologies will become the standard at the beginning of every therapeutic program.” Indeed, Simulations Plus has six FDA grants in six different areas in partnership with the agency and industry.
“Designing drugs within a virtual animal or human setting to target preclinical or clinical pharmacokinetic endpoints is a game changer,” DiBella declares. “Virtual human twin environments to enable precision drug dosing are bringing us to the next frontier. They are now being used in academic clinical trials.”
Combining tools and expertise
Molecular Forecaster is a research-as-a-service company that provides computer-aided drug discovery (CADD) technology and expertise. The company’s offerings include proprietary software tools and expertise in drug discovery, quantum mechanics, molecular dynamics, chemoinformatics, and AI. In addition, the company is home to a cross-functional team of chemists, biologists, and computer scientists.
“We are democratizing CADD,” says Josh Pottel, PhD, the CEO of Molecular Forecaster. “We offer the intimate collaboration of a partner with the logistical integration of a contract research organization.”
Molecular Forecaster recognizes that niche problems require more custom solutions. “We can generate computational modeling hit rates of 15–20% versus the 1% of traditional high-throughput screening,” Pottel asserts. “We decided we would essentially ‘teach chemistry’ to our software. [This is part of the reason we believe] that when the platform can understand the rules, it will be able to apply them more broadly.”
For example, chemical reaction principles were encoded into the covalent docking modeling, which predicts how a molecule covalently binds to a specific residue of a protein, determines whether a covalent bond could be formed, and then virtually demonstrates the outcome. Several new features broaden its applicability to a wider range of covalent enzyme inhibitors and metalloenzymes, where metal coordination is essential for binding.
“Chemical research is complex, dynamic, and challenging,” Pottel stresses. “CADD can bolster the medicinal chemist’s toolbox, but software alone will not save lives. You need to combine sophisticated, practical tools with a level of research knowledge and expertise that is often difficult to source.
“The impact of technology in biotechnology is increasing, with a lot of talk around AI and, more specifically, ML. While this is exciting, it is critical we do not stop valuing the decades of knowledge, indispensable learnings, and knowhow of domain experts.”
Pottel believes that Molecular Forecaster’s technical multilingualism—its ability to communicate in chemistry, biology, physics, math, and computer science—is one of the company’s greatest strengths. Different threads of expertise often intertwine to enable a synergistic whole greater than the sum of the parts.
Expanding cloud access
GridMarkets is applying its expertise in setting up cloud platforms to simplify mathematical processes in animation and visual effects and computational drug discovery. “A computational chemist told me that in order to gain a higher degree of confidence before actually synthesizing a molecule and going to clinical trials, the more virtual testing, the better,” says Mark Ross, director and co-founder, GridMarkets. “But virtually testing is complex and computationally heavy, and setting up an in-house computing environment is laborious and expensive.”
The cloud-based GridMarkets Pharma platform was built specifically to address these issues and to provide impactful efficiency increases. To gain access to an unlimited number of software licenses and computational power, GridMarkets partners with secure platform and cloud providers such as Oracle, Amazon, and Google. Platforms currently supported include NAMD, GROMACS, Amber, and QuantumBio. The company is actively expanding its offerings, including reference databases.
“We deliver a solution that bursts to the number of machines needed,” Ross asserts. “If you can apply many machines to a problem and run that problem in parts and in parallel, you get it done much faster. Our service helps to design and test more molecules in a shorter period of time.”
After an account is established and the software plug-in is installed, a menu option allows submissions to GridMarkets. The user specifies the machine number, decides on configurations, and hits “go.” Local files are securely transported to the GridMarkets cloud for processing, and then results are consolidated and downloaded to a specified folder.
“To run molecular simulations faster, our customers just need to specify more concurrently running machines. We do the rest,” Ross says. “For an hourly fee, you get access to the software licenses, the machines, and data storage. The pay-for-use pricing approach makes the platform affordable. We want to simplify the workflow, eliminate set-up complexities, and provide cost transparency.”