Hello. Thank you for joining me in the ACMG Exhibition Hall. My name is Kristen Sun, and I am a Senior Scientist at IMAGINE. I come to you with a background as a Certified Genetic Counselor and Diplomate of the American Board of Medical Genetics and Genomics, who spent almost 10 years in a clinical genetics laboratory setting. In that time, I saw changes in the types and quantities of testing ordered and the shift from manual to more automated processing. I am a staunch believer that the survival and prosperity of a lab depends on the ability to accelerate processes through the use of artificial intelligence and automation in wet lab and dry lab processes. Today, I will review the results of a qualitative study of laboratories who have already incorporated AI into their clinical workflow. We would like to shift the conversation around AI into thinking of optimal technologies for AI adoption. Accuracy studies are great, and we've performed them ourselves. However, no one in medicine today is relying on the self-driving or self-analyzing algorithms. Once AI has done its part, we need to find the best way to incorporate it into our workflows, both clinical and research. So we asked our customers a series of questions about their experience implementing AI into their workflow. As an overview, I'll give a brief introduction and take a few moments to demystify buzzwords often heard around AI technologies. I will then review the survey results. But first, a brief overview of Imagine, the company. Imagine was founded in 2016 with the goal of creating an AI-first approach to dramatically scaling genomic interpretation. We have a large team of geneticists in our R&D and a strong heritage of rare disease discovery. Our AI was the first automated interpretation solution in the market some three years ago, and it's the first and only explainable AI. It's been in daily use in clinical and research labs for the past three years. I also want to talk a little bit about our motivation, why we need AI in genomic medicine. The International Rare Diseases Research Consortium, a global consortium of patient groups, industry, and academia, set as one of its 10-year goals that by 2027, any patient with a suspected rare disease will be diagnosed within a year if their disorder is known in the medical literature. Looking beyond the rare disease use case, analysts are predicting that when we reach the $100 genome, up to 50 to 60 million people will have their DNA sequenced every year. Are our current tools and processes sufficient to handle this dramatic volume increase? Those of you who have been in the field for a while are familiar with how manual everything can be in genomics. From rooms full of paper charts literally stacked to the ceiling, and the manual extraction of phenotypes during paper requisitioning, the ordering physician has their work cut out for them. At the beginning of the laboratory workflow, many samples were, and in some cases, were not sequenced, but were sequenced by a geneticist. Many samples were, and in some cases still are, accessioned by adding stickers to individual tubes with handwritten labels for tracking. Samples used to be set up one at a time with manual pipetting. Once the data was collected, data analysis was performed in Excel or in a homebrew software, with lots of linkouts and manual exploration of data resources in multiple external websites. And let's be real, a lot of laboratories are still dealing with pieces of this workflow. But to survive in the future, a lab needs to find a way to make these processes more efficient, to save money, and ultimately, to diagnose more patients. Nowadays, requisitioning is partly automated from the EMR record. We have robots mixed in with laboratory technologists for accessioning, barcoding, and setting up tests in large batches. But approaches to data analysis are often unchanged. If we want to support genomic medicine at scale, and we need to do this, we need the same automation we've incorporated in the wet lab to cross over to the dry lab. Much of that will be in the form of artificial intelligence that can transform data collection, synthesis, and analysis, leaving scientists more time to focus on areas where their clinical judgment is irreplaceable. In 1980, Steve Jobs said, when we invented the personal computer, we created a new kind of bicycle, a new man-machine partnership. We can't imagine lives without our computers or our smartphones. The old analog asynchronous way is unfathomable today. We believe that in the very near future, we'll feel the same way about artificial intelligence. AI will be the basis of a new scientist-machine partnership that will help us analyze tens of millions of genomes a year and deliver life-changing answers to patients. But in order to develop this meaningful partnership with AI, we need to understand how to incorporate it into our genomic analysis workflows. AI is an overly used term today, so let's spend a moment demystifying some of the buzzwords that are commonly used to describe AI. We can distinguish between artificial intelligence, which has been around since the 50s, machine learning, which was developed in the 80s, and deep learning, which has only been around for a few years. I've taken these definitions from Eric Topol's book, Deep Medicine. Artificial intelligence is the broad concept of machines that have the ability to achieve goals like humans via a constellation of technologies. In our field, an example of AI is polyphen, an in silico tool that uses empirical rules to predict functional effects of human SNPs. Machine learning is a computer's ability to learn from examples and experiences rather than predefined rules and without explicit programming. An example of a machine learning tool is Revel, which uses a random forest classifier model to predict the pathogenicity of rare missense variants. Deep learning is a subset of machine learning algorithms that permit the software to train itself by processing multilayered networks of data. An example of a deep learning model is DAN, which is a deep neural network. Deep learning allows DAN to capture nonlinear relationships between features and achieve a higher area under the curve than machine learning models trained for the same purpose. Now, all of these examples I just cited are narrow AI solutions. That is, models developed to perform a single task, for example, predicting variant pathogenicity. At the opposite end of the spectrum, we have generalized AI, currently only seen in movies like How, the heuristically programmed algorithmic computer, which is the machine that can perform any human task that a human can. We're not quite there yet. But AI technologies are moving up from narrow AI. We're now able to blend AI technologies like neural networks, machine learning, and natural language processing in order to mimic human intelligence. This category of AI is typically called cognitive intelligence and is able to make decisions in complex and context-rich situations. The IMAGINE platform is a cognitive intelligence solution. The last concept I'd like to review is xAI, or explainable AI. We are firm believers that AI in medicine cannot be a black box. xAI is a model that elucidates the decisions of an AI model in a way that's easier for humans to understand. It allows the analyst to detect any errors or bias.

artificial intelligence

laboratory workflows

genomics

clinical settings

research settings