It’s a bit difficult to pin down exactly what people in the biotechnology industry mean by “artificial intelligence” (AI). In general, they seem content with a working definition, one that describes AI as a computer program that can learn and predict outcomes based on the data sets it receives.

Given that the working definition of AI is vague, it follows that the status of AI in the biotechnology industry is vague, too. For example, it is unclear whether AI is to be regarded as something new and revolutionary. Are news stories in the popular press any guide? These include breathless reports of how AlphaFold, an AI system developed by Google’s DeepMind, accurately predicted the structure of hundreds of thousands of proteins.

Although AlphaFold is a groundbreaking technology, it isn’t the be-all and end-all of AI in biotechnology. AI-related spadework in biotechnology is occuring in several fields of endeavor. Indeed, biotechnology companies far and wide have been implementing AI in their pipelines.

If we are to clarify AI’s status, we should begin by recognizing that AI in biotechnology hasn’t suddenly become mainstream. In fact, it is already mainstream. Moreover, it is diverse and ready to produce results. In the biotechnology industry, AI that is accurate, predictive, and productive is within reach. Such AI will be worth a hundred times its weight in bench lab scientists.

AI is as AI does

A vague definition of AI—big algorithms learning from big data sets—may be all we need. What’s really important is what AI systems do. And what AI systems do will depend on who is developing them.

Agilent Technologies, the instrumentation and consumables company, has taken steps into the AI world, including partnering with Visiopharm in AI-driven studies of cancer pathology. The companies see the work as a combination of different approaches.

Agilent’s perspective is well described by two of the company’s executives—Anya Tsalenko, PhD, computational biology section manager, and Stephen Laderman, PhD, director of research laboratories. They confirm that the most advanced AI approaches use computationally intensive models trained with large data sets.

“These models broadly fall into the category of machine learning and include different flavors of deep convolutional neural networks,” they say. “The successes of these new approaches are based on the combination of the power of modern computers; the ability to collect, store, and access large amounts of data; and the recent advances in computer science, mathematics, and statistics.”

“I totally understand that people define [AI] differently,” adds Carl Hansen, CEO of AbCellera Biologics, a Vancouver-based company using AI to create therapeutic antibodies. “And I think it’s really dangerous to talk about AI as a category. It’s way too abstract. In my opinion, there is a lot of very powerful and interesting applications of AI in drug discovery. And there’s also a lot of fluff.”

Yet another perspective is offered by Fabrice Chouraqui, PharmD, CEO of Cellarity, a Cambridge-based startup focusing on cell-based therapeutics derived from imaging- and machine learning–based approaches. He sees AI as a complement to existing paradigms, but one that hasn’t always lived up to past promises. “I think, today, AI is mainly used to improve existing approaches,” he elaborates.” So, the vast majority of the companies in drug discovery will tell you that they can leverage AI. By making the current discovery paradigm more efficient, they’ll tell you that they can use AI to find new targets.

“I have experience working with people who had, in the past, actually made claims that they couldn’t support. And not that they didn’t have the algorithm, to be fair. I think [the idea that the algorithm is all important] is perhaps one of the most common misunderstandings about AI. AI can only be as good as the quality of the data.”

Putting AI in the picture

Data sets aren’t the only source of biologically significant patterns. Such patterns can also be found in pictures of cells, tumors, and tissues—even if these pictures are analyzed using time-consuming, eye-strain-inducing, brute-force approaches. To find patterns in images more easily, researchers may take advantage of AI. According to Tsalenko and Laderman, AI image analysis can enhance both basic research and medical pathology.

“[We have] strong positions in live-cell microscopy and histopathology,” they explain. “With appropriate selection of methods and model training, AI image analysis enables one to locate, classify, count, and identify patterns in microscopy images, including those that are challenging for a human eye to see efficiently, comprehensively, and quantitatively.

“These attributes are among the drivers for the ongoing digital transformation of microscopy-based tissue analysis in clinical pathology laboratories. It is envisioned that in the end, AI will, for example, improve the precision with which cancer can be diagnosed and treated. In the case of live-cell imaging, AI can improve tracking of the dynamics and evolution of cells under study.”

Streamlining discovery and development

Stacie Calad-Thomson, PhD, is the chief strategy officer and head of drug discovery at BioSymetrics, a Boston- and Toronto-based startup studying electronic health records and in vivo data using machine learning approaches. She believes that AI can be used to streamline drug discovery and development.

She emphasizes three possibilities: “First, AI can speed up the timeline in which we discover new therapeutics because informatics and multiparameter optimization can help us make more informed decisions about the experiments to run. Second, AI can help us better understand biology and the underlying codes, signals, and pathways that drive disease, or that may be used to treat disease. Third, AI can identify and define the patient populations for which a specific treatment works best, by linking phenotypes of disease with underlying codes such as genes and biomarkers.”

“In the near term, a lot of AI success is focused on chemistry and molecular design as an efficiency play,” she continues. “But in the longer term, biology is the harder problem to solve, and tackling that will lead to reduced attrition and increased clinical success. The ability to acquire and integrate diverse patient data at the outset of drug discovery, using AI, is going to be the rule, not the exception, for how we discover drugs.”

A virtuous cycle of data and drugs

“When you think about a disease, you think about malfunctioning organs, you think about malfunctioning tissues, you think about malfunctioning cells,” says Chouraqui. He quickly adds, however, that our ability to understand disease in such rich contexts has been limited.

Because our understanding of disease has been cramped, our conception of how to develop treatments has been correspondingly reductive. “Our only way to address a malfunction in cells was to find a target,” Chouraqui complains. “If you take a single molecular target, that’s your assumption from the get-go. Obviously, it fails to address the complexity of biology. As you move along the development process, you go to higher species, and you see that the assumption that you’ve made earlier will not be confirmed.”

cellular dysfunction
The Cellarity Map provides a comprehensive view of disease pathology and the changes that occur as a cell moves from a state of health to disease. It is not a theoretical model. It is a digital representation built from in vitro data that depicts cellular dysfunction. Using a proprietary Intervention Library, Cellarity identifies drugs that address this cellular dysfunction and engender a desired cell change. Cellarity learns from features of effective interventions to create new chemical entities with a higher probability of clinical success.

Moving beyond a target-centric approach to drug development and embracing, as Cellarity does, a cell-level approach, means collecting and analyzing large amounts of data—and quality counts. In general, high-quality data and high-quality analytical work are mutually reinforcing. This point came through when Chouraqui was asked whether experimental biology will recede in importance with the advance of algorithm-driven drug discovery and development.

“No, I don’t believe that’s true,” Chouraqui stated. “It’s going to be always a conjunction of wet lab and dry lab because it’s a loop. You absolutely need to access and generate high-quality assets. And you need to test the outcome of your AI generative process and then feed the outcome of that test into your AI process.”

The same question was posed to AbCellera’s Hansen. He responded, “I agree with that 100%. The companies that are going to get the most leverage out of AI are the ones that have the most formidable experimental capabilities as well.”

Many pieces to the AI puzzle

For a moment in time, AlphaFold seemed poised to kill crystallography and reform molecular interaction work forever. Not so, says Hansen. He elaborates, “Knowing the protein structure does not mean that you’re able to find drugs faster than other people necessarily. We’ve [been collecting] structures for a very long time. [We’ve entered] over 65,000 structures in the Protein Data Bank. We don’t have companies that look at structures and then tell you what the drug is. There’s a lot of steps between knowing a protein structure and being able to define what a drug is.”

Previous articlePharma Leverages AI to Elevate Digital-Only Operations
Next articleBeating COVID-19 Means Tipping the Balance of an Age-Old Struggle (PART I)