Andrew Satz, EVQLV’s CEO, was published in a Life Science Leader article on how AI can be used in pharma to automate, analyze, and/or accelerate tasks as well as how to leverage AI to assist and augment the work of team members. Written for leaders in the biotech and pharmaceutical industries, the article details strategies and tactics around how decision-makers should think about and use AI to transform an organization, its members, and its outcomes.
“Examining AI as automated, assistive, analytical, accelerated, or augmented intelligence is a framework for thinking about how to apply these tools to data in ways that are aligned with an organization’s goals. This framework is designed to allow those who do not routinely build or employ AI to avoid the overhyped rhetoric of AI and consider the practical outcomes of employing it.”
Full text available below.
Magazine Article | May 10, 2022
Demystifying AI In Pharma
By Andrew Satz
Artificial Intelligence (AI) is a term so overused and misunderstood that it has lost its meaning. As a data scientist who spends a significant amount of time thinking about how and where to apply machine learning to challenges in life sciences and the pharmaceutical industry, when the majority of individuals say AI, what I feel they really mean is technology that can simulate human intelligence. By that definition, everything from a simple calculator to a recurrent generative adversarial network (and beyond) could be considered AI. But thinking of AI as an algorithmic method may not be the best way to recognize its value.
Economist Theodore Levitt once said, “No one wants a drill. What they want is the hole.” The same can be said about AI; it is less about the tool than what it can do. Instead of thinking about AI as a smart machine or some theoretical statistical algorithm, think of it as a framework consisting of five other words starting with “A”: automated, assistive, analytical, accelerated, and augmented.
In pharma, automation can be applied to a whole host of operations, including clinician engagement, clinical trials, and laboratory processes. For example, using data from the Centers for Medicare and Medicaid, automated systems can track which doctors are performing certain procedures and use that data to draw conclusions about certain therapeutics or procedures to improve the lives of patients. In clinical trials, wearables can automatically track vitals, location, and sleep schedules in ways that are much more accurate and consistent than self-reported data. In drug discovery, robotic lab automation can be used to produce and test novel candidates by mimicking manual and labor-intensive tasks.
One of the main concerns surrounding automated intelligence is job loss, which is a distinct possibility even for highly trained laboratory scientists and experts. Yet the transition to utilizing automated intelligence allows these individuals to design laboratory experiments at scale. Since most of the manual labor at early stages of the experimental pipeline will be performed by automated intelligence, laboratory scientists are afforded more opportunities to leverage their experience and creativity in designing careful and smart experiments. While automated intelligence can perform laboratory experiments faster and with higher accuracy, it lacks the capacity to distinquish causation from correlation and generate hypotheses on its own. Skilled scientific thinkers, who can generate independent hypotheses and experiments, combined with fast machines, allow a significant scale-up of the number of lab experiments. If done strategically, tactically, efficiently, and with compassion for individuals who make up the workforce, automated intelligence can increase innovation and productivity while minimizing job loss, reducing inefficiencies and minimizing failure rates.
The amount of biological and clinical data is growing at an exponential rate. For example, at EVQLV, we generate nearly 200 data points per therapeutic candidate. When working with collaborators, we may provide them with 1,000 potential candidates, which means we are delivering over 80,000 data points, which is far too many for any human being to comprehend. This is where assistive intelligence can help by using, for example, data clustering. Clustering is an algorithmic method of grouping a set of objects in such a way that objects in the same group are more similar to each other than to those in other groups. So, while automated intelligence supports performance of a physical task, assistive intelligence enables mental tasks to be performed faster. Assistive intelligence also includes supporting cognitive tasks that would be nearly impossible without a computer, such as calculating the square root of 4,080,400 (it’s 2020, in case you were wondering).
In medicine, computer-assisted detection is meant to support (not replace) radiologists by prereading scans and prioritizing the potential seriousness of the case, rather than the order in which the patients were scanned. Taken a step further, instead of the tedious process of delineating areas of interest on medical images, assistive intelligence can highlight a region of interest for a radiologist. In the case of cancer, this would enable quick, accurate, and consistent quantification of the metastases in the scan, which can set a very precise baseline for impact of therapies or assist in classification for treatment plans.
There are legitimate concerns that assistive intelligence may cause cognitive decline in humans. For example, there is definitely a ”dumbing down” when it comes to the ability to spell as a result of autocomplete. Yet assistive intelligence offers us significant cognitive advantages and advances in exchange for some cognitive losses. While one may not learn how to spell, it does not remove the ability to communicate. Additionally, the ability to spell may become an antiquated skill, like writing in cursive. This is one of the many challenges of change. When it comes to assistive and automated intelligence, it is important to be thoughtful about whether we are giving up too many old skills in exchange for the benefits of the new skills.
General Eric Shinseki once said, “If you dislike change, you’re going to dislike irrelevance even more.” So, while humanity will most certainly lose skills because of the changes caused by assistive intelligence, we will absolutely achieve new and innovative abilities. For example, while autocomplete impacts the ability to spell, it can be used to increase the speed at which doctors fill out EHRs, allowing them to see additional patients. In the case of the radiologist, if 50 percent of the scans have the same outcome, the assistive intelligence will get really good at reading those scans, but AI will be really bad at the outliers. That’s where human intuition comes into play, requiring a well-trained and experienced individual. Human cognition is essential, and assistive intelligence can support the advancement of human intellectual ability.
Mathematician Clive Humby said, “Data is the new oil.” Like oil, in order to become valuable, data must be extracted, cleaned, and processed so that it can be transformed into information. That information can then be converted into knowledge to allow individuals and organizations to take action. Analytical intelligence offers the capability to transform data into usable and relevant knowledge.
In the case of the laboratory, while automated intelligence can generate 80,000 data points, analytical intelligence allows the data points to be transformed into an interpretable format. One way to do this is to visualize the data, with a pie chart for example, which allows human experience and inference to be combined with information. That combination becomes knowledge and provides the ability to make a decision. Terms such as “business intelligence” are examples of the conversion of data into information that allows individuals or teams to make a decision.
But imagine that instead of the lab producing 80,000 data points, it produces 80 billion data points. That is much harder to visualize in a way a human can interpret for decision-making purposes, especially beyond three dimensions. This is where additional analytical intelligence layers can come into play, because the computer can use statistical inference and perform probabilistic analysis on the data, which is something that humans do instinctively. Analytical intelligence can convert the data into knowledge, allowing it to make a “decision” in the limited scope of whatever the analytical intelligence is programmed to perform. Those decisions are driven by the data that trains the analytical intelligence. The decisions are made probabilistically across as many data points as are made available to the analytical intelligence. Unlike business intelligence, which requires human interaction to convert information into knowledge, this type of conversion of data into knowledge occurs because the machine is programmed to mimic human decision making.
In pharma and the life sciences, if we consider how many genes or metabolic pathways exist in the human system, analytical intelligence can scan these massive bioinformatics data sets, thus creating a multilayered landscape/matrix of cellular and molecular information. Thereafter, context-dependent analysis will enable identification of relevant cellular and molecular pathways with a higher probability of occurrence of mutations or signaling alterations. The results can then be passed on to life scientists who can use their experience and laboratory prowess to make determinations about those targets and their impact on health in relevant disease models.
Automated, assistive, and analytical intelligence expedite various research and medical processes; however, utilizing this framework, accelerated intelligence provides a way to predict outcomes through various types of modeling. At EVQLV, we use accelerated intelligence for drug discovery to fail as quickly and as often as possible in the computer, where the time and cost are low, instead of failing in the laboratory, where the cost and time are high. However, in some cases, it can be harmful to accelerate, especially in medicine where moving too quickly can be a significant safety risk. Expediting clinical trials harbors the inherent risks associated with candidate drug molecules and therapeutic targets that include side effects and toxicity, to name a couple. In such cases, accelerated intelligence can potentially measure and predict whether a drug will be toxic in certain states, allowing the promotion of studies by limiting clinical study failures. It is the predictive power of accelerated intelligence that could determine the duration of preclinical and clinical trials, identify novel and prevalidated candidates, and facilitate in the removal of therapeutic candidates who might otherwise have received significant time and resources.
Accelerated intelligence is ideally suited for adaptive clinical trials, which are challenging to devise, in part due to the assembly and calculation of prior statistical assumptions. These prior assumptions are built on data about the disease, patients, and medicines, which comes in significant volume and variety. By combining adaptive clinical trial design with accelerated intelligence, the machine can use large amounts of data to calculate prior assumptions into multiple models of the clinical trial. Once this is done, accelerated intelligence can simulate hundreds or millions of trials and predict their outcomes to determine the optimal strategies that would validate or invalidate the endpoints, all at a significantly reduced time and cost. Accelerated intelligence helps to digest all of the data and support the design of various trials that are statistically robust and that regulators will accept. Once the best computationally designed trial options are determined, they can be reviewed by those with experience in the disease and treatment to choose the optimal trial design. This allows the machine to do the calculations with as much data as possible to accelerate the trial, followed by individuals making the decisions with the experience and intuition that the computer lacks.
Automated, assistive, analytical, and accelerated intelligence are meant to augment human behavior. A calculator, which is a form of assistive intelligence, is augmenting the ability to add, subtract, multiply, divide, etc. The calculator mimicks human intelligence and makes the person better at something they already know how to do. When the radiologist is augmented by not having to highlight the tumor in a scan, it allows time to engage in additional critical activities ranging from attending to patients to performing case studies. Google provides easy access to information, augmenting an individual’s ability to sift through vast amounts of information. If we consider the automated, assistive, analytical, and accelerated intelligence tools together, then we can begin to see how augmented intelligence can support the advancement of individuals, organizations, or methods.
Author Michael Pollan said, “Scientists reduce problems to their simplest terms, while engineers are faced with irreducible complexity.” This suggests that scientists and engineers approach similar challenges from entirely different perspectives. By using augmented intelligence to advance science, one can combine the scientific methods of reduction and the engineering approach to complexity. For example, instead of looking at a single factor at a time in vitro, scientists could combine analytical and accelerated intelligence to look at single factors in a series while examining the statistical relationships between those factors. Based on the results, scientists can thereby investigate relevant factors that are related, thus augmenting the workload in a laboratory setting. In drug discovery, this could mean examining potential points of failure or success in tandem at earlier points in time.
Another way to utilize augmented intelligence is in animal studies. Considering the high failure rates of clinical trials, animal studies often offer a minimal predictive power for the outcome of human studies. Unfortunately, this means that, in many cases, testing on and sacrificing animals may not truly advance science in preclinical studies and drug screenings. Animal studies could be augmented with computational models by using retrospective data of each animal study. Analytical and accelerated intelligence can be used to predict the outcomes of a given animal study. With more and more accurate predictions come better and better computational models of the diseases. In order to create scientifically robust computational models of animal studies, augmentation of animal testing (in vivo) is an ideal introductory approach. This provides a path that could lead to replacing animal studies with computational models that are acceptable to the scientific community and regulators.
TRANSFORMING AI FROM A TOOL TO AN OUTCOME
Examining AI as automated, assistive, analytical, accelerated, or augmented intelligence is a framework for thinking about how to apply these tools to data in ways that are aligned with an organization’s goals. This framework is designed to allow those who do not routinely build or employ AI to avoid the overhyped rhetoric of AI and consider the practical outcomes of employing it. At EVQLV, we use AI to accelerate the design, discovery, and delivery of biologics. At Google, AI is used to organize the world’s information and make it universally accessible and useful. At Two Sigma, AI is applied to the data-rich world of finance to discover value. By focusing on which operations to automate, analyze, or accelerate, as well as how best to assist and augment team members, leaders can create strategies and tactics to truly use AI to transform an organization, its members, and its outcomes.
ANDREW SATZ is a data scientist and the CEO/cofounder of EVQLV, a sci-tech company that combines data science, software engineering, and biological data to accelerate the speed at which biologics reach patients.
Source: Life Science Leader