Medicines address a broad range of human diseases, but discovering and manufacturing them at scale is a complex process that costs pharmaceutical companies time and money. The $2.5 billion (USD) estimate in 2014 has been growing at 8.5 percent annually, according to the Tufts Center for Drug Development. Worse, the "Ten Years on Measuring Return on Pharmaceutical Innovation Report 2019" from Deloitte Centre for Health Solutions states return on investment has decreased tenfold in just a decade.
Pharmaceutical companies need to decrease cost and time to market if they are to remain competitive. The focus on speed has become even more pressing in the wake of the COVID-19 pandemic. Artificial intelligence is providing a salve, delivering efficiencies in the early stages of the drug discovery process.
An ideal drug must meet a number of criteria. Not only must it achieve the desired effect of killing off the target toxin, it must do so without creating worrisome side effects. The ideal medicine must be of the right size and composition to attach itself to the toxic cell and penetrate its wall to deliver the medicine. Scientists must also ensure that the medicine does not transform into something else once in the human body and that its dosage does not increase or decrease as a result of other interactions in the bloodstream.
The first four stages include narrowing down an ideal candidate from millions of potentials. Companies screen dizzying numbers of molecules—three million by some estimates—to narrow a few candidates that are then evaluated for further lead selection and optimization. The entire process can take five to six years, but artificial intelligence has been making impressive dents in this timeline, compressing it at least tenfold if not more. Here’s how progress has been made.
Advances in medicine are fast-moving even in the best of times, so staying updated and building on existing research requires carefully culling published literature. The COVID-19 pandemic brought an onslaught of trials, studies, and insights, thousands of which have to be studied to figure out what the vaccine or medical treatments would look like.
Scientists start by corralling as much of the available literature that they can and making it publicly available so others in the research community can work efficiently off a centralized database. This has been noticeable in the efforts to collate research into COVID-19. Google’s Kaggle, a machine-learning and data science repository, hosts the related research challenge that acts as a clearinghouse for a variety of information about the pandemic. Larger drug companies with decades of research also maintain their own databases that they can repurpose. In some cases, companies have partnerships with clinical establishments such as hospitals, where they can obtain anonymized results of the use of drugs on individual patients, including the use of multiple drugs for different indications.
To search a repository of research papers effectively, the most basic approach would be “text mining,” where subject matter experts help to define relevant terms, and search engines extract relevant text from the research. This approach helps researchers identify drug interactions and observations regarding trials with similar drugs or combinations of drugs, which they would otherwise have missed. This could help cut down on duplicate research, speeding up the search.
AI techniques can help improve text mining by automatically identifying patterns in similar drugs or similar patient characteristics. These patterns can help propose or test a hypothesis such as: Will a certain drug model work? Or are there specific circumstances when there are side effects or interactions with other drugs?
To understand which medicine will work, scientists need to study the structure of the molecules under consideration—to see which will attach properly and how. Will the molecule stay attached long enough for the medicine to do its work? Evaluating such models for millions upon millions of molecules is tedious—if not humanly impossible—work.
Artificial intelligence helps by quickly matching potential molecules against the desired properties. It’s able to play the game of compare-and-discard much faster than humans can.
One approach is to feed off of existing datasets, learning which molecules have traditionally worked as candidates for a certain set of conditions, and which ones have not. By screening entire libraries of small molecules and pattern-matching three-dimensional structures of potential candidates, AI can sift through compounds and only serve up the most promising ones for further analysis.
Pattern recognition at scale is feasible with AI, particularly if experts in the field are able to set up the parameters of the search. Researchers can factor in other cellular conditions and chemical mechanisms for drug transport, for example. Given these parameters, an AI search engine can deliver much more nuanced candidates for further exploration.
With the COVID-19 pandemic, researchers have been able to plug in previous iterations of the coronavirus, most notably severe acute respiratory syndrome (SARS), to see whether they can repurpose existing drugs as a vaccine solution. Gilead Sciences, for example, is researching an antiviral drug used to treat Ebola, to see what mechanisms might be useful to treat the novel coronavirus.
In some cases, drug candidates that might not make a perfect fit for one disease might work for another. If the AI learning set has the information about ideal candidates for all the different kinds of drugs needed, then rejecting one need not necessarily mean that the candidate is a total dud. It could be flagged as meeting the criteria for another.
AI is also valuable in synthesizing new drugs. Instead of sifting through past candidates, AI can help model a new molecule from scratch—de novo. If we know that the best candidates have to qualify a certain set of criteria, algorithms can be developed that can plug-and-play these conditions to construct a hypothetical model. Scientists can use this model to create a new vaccine or drug from scratch.
In the case of the novel coronavirus, for example, Chinese scientists decoded the genetic sequence of the virus and shared the results in a public database. Inovio Pharmaceuticals, a California biotech company, had a vaccine candidate in just a few short days using its proprietary machine learning algorithm. The drug is now in preclinical trials.
Every small molecule that’s ingested as a medicine interacts not just with the target it is designed for, but also with other proteins in the neighborhood. In the best-case scenario, such unwanted interactions are harmless. When they’re not, they need to be cataloged, flagged, and learned for future analysis cycles. AI and machine learning are especially good at learning from candidates that have failed, storing that information and flagging which potential candidates run the risk of running into similar trouble.
Protein-destroying molecules are actually how certain cancers are treated, so cataloging and screening candidates using AI can quickly float more qualified leads to the top.
AI is only a tool. Human researchers with expertise in the field must configure this tool in such a way that it can be used. For AI to output valuable answers, experts in the field need to ask the right questions. Similarly, its output must still be validated by human researchers. AI’s intelligent computing power, however, delivers faster and more efficient ways of doing research in the early phases. When time is of the essence, as in the COVID-19 pandemic it can make an invaluable difference.
Poornima Apte is an engineer turned writer with B2B specialties in robotics, AI, cybersecurity, smart technologies and digital transformation. Find her on Twitter @booksnfreshair.
版权所有©2023 Mouser Electronics, Inc.
Mouser® 和 Mouser Electronics® 是 Mouser Electronics, Inc 的注册商标。