BTEP AI in Biomedical Research @ NIH Seminar Series

The accessibility of artificial intelligence/machine learning (AI/ML) tools has taken off in recent years. This democratization of advanced analytics has the potential to revolutionize predictive toxicology, especially for applications that generate massive, multimodal data. Realizing this promise will require tools tuned to learn from trusted sources that can evolve as new data emerge. This talk will describe such efforts at NIEHS using data that range in scale from lab-based behavioral experiments to epidemiological-scale geospatial data.

Recent advances in artificial intelligence (AI) have revolutionized the use of hematoxylin and eosin (H&E)-stained tumor slides for precision oncology, enabling data-driven approaches to predict molecular characteristics and therapeutic outcomes. In my talk, I will present ENLIGHT–DeepPT, a novel two-step AI framework. The first step, DeepPT, leverages deep learning to predict genome-wide tumor mRNA expression from H&E slides. The second step, ENLIGHT, utilizes these inferred expression values to predict patient response to targeted and immune therapies. We validate this framework across 16 cohorts from The Cancer Genome Atlas (TCGA) and independent datasets, demonstrating successful prediction of true responders in five patient cohorts spanning six cancer types, with a 39.5% increased response rate and an odds ratio of 2.28.

In addition, I will introduce DEPLOY, a deep learning model designed to enhance the diagnosis of central nervous system (CNS) tumors by predicting tumor categories from histopathology slides. DEPLOY integrates three components: a direct classifier based on histopathology images, an indirect model that predicts DNA methylation profiles for tumor classification, and a model that uses patient demographics. Trained on a dataset of 1,796 patients and tested on independent cohorts of 2,156 patients, DEPLOY achieves 95% overall accuracy and 91% balanced accuracy. These results underscore the potential of DEPLOY to assist pathologists in classifying CNS tumors rapidly, offering a promising tool for improving diagnostic precision in clinical settings.

Digital Pathology has advanced significantly in the past decade, evolving beyond assisting pathologists to now informing molecular and genetic properties of tumors. Recent breakthroughs in machine learning and AI, particularly in big data and image analysis, have ushered in a new era of capabilities. This talk will present these advancements and demonstrate how our lab is developing tools to apply them, focusing on improving precision diagnostics and treatment of pediatric tumors. We'll conclude by exploring applications of these techniques to analyze and understand reproductive aging, showcasing the broad potential of next-generation digital pathology in medical research.

Dr. Turkbey will discuss radiology, pathology, and multimodal AI models his NCI lab has developed for prostate cancer diagnosis and prognosis prediction.

Although generative artificial intelligence (AI), a’la ChatGPT, is receiving a lot of “attention” these days, there are many other options for using AI to support biomedical research.  In order to help analyze and interpret single cell genomics data, we have found that AI approaches that retain “explainability” are especially useful in providing functional insights into the underlying biological systems being studied, in this case, the human brain.

Artificial Intelligence (AI) is becoming increasingly ubiquitous in biomedical research, enabled by large datasets, new algorithms, and hardware improvements. In this session, Dr. Brian Ondov will introduce the basic principles of AI and describe how its various forms can help researchers in different ways, including image classification, sequence-based prediction, generative models, and language understanding.