Daphne Koller on Machine Learning, Biomedicine, and the Future of Drug Discovery

This podcast features Daphne Koller, a pioneer in machine learning and co-founder of Coursera, discussing her pivot to applying AI to human health through her company, Insitro. Koller addresses profound questions about curing all diseases and extending human lifespan, emphasizing the complexity of biology and the current limitations in understanding fundamental disease mechanisms. She highlights that while some diseases are relatively well-understood, the majority, including conditions like Alzheimer's and schizophrenia, remain largely a mystery, often being heterogeneous collections of mechanisms rather than single ailments. Koller advocates for an increased "healthspan"—a longer period of healthy, active life—as a more realistic and desirable goal than immortality, acknowledging the natural processes of aging and cellular wear and tear.

A central theme is the transformative role of machine learning in drug discovery, a field where its impact has historically been limited by a scarcity of suitable data. Koller explains that Insitro's innovative approach is to actively generate large-scale, high-quality datasets specifically designed for machine learning. This involves leveraging advanced bioengineering and cell biology techniques to create "disease in a dish" models. These models overcome the limitations of traditional animal models, which often fail to translate findings to humans because the induced disease mechanisms in animals differ significantly from those in human patients. By using induced pluripotent stem cells (iPS cells) derived from human patients, Insitro can create specific cell types (e.g., neurons, cardiomyocytes) that carry the patient's unique genetics, allowing for the study of disease manifestation at a cellular level.

Koller details the technological advancements enabling this data generation, including the ability to revert human cells to stem cell status using Yamanaka factors and then differentiate them into target cell types. She also discusses the use of CRISPR gene editing to introduce pathogenic mutations, creating direct comparisons between healthy and diseased cells from the same genetic background. The data collected from these models is highly quantitative, utilizing techniques like single-cell RNA sequencing to measure gene activity and super-resolution microscopy to visualize subcellular structures, effectively transforming "squishy" biological phenomena into digital data.

Finally, Koller outlines how machine learning processes this vast amount of data to uncover patterns, identify molecular subtypes of diseases, and predict potential interventions. This less hypothesis-driven approach allows for the discovery of novel drugs or CRISPR-based therapies that can revert diseased cells to a healthy state. While cautious about making definitive promises, Koller expresses optimism that this methodology holds significant promise for diseases with a strong genetic basis, reproducible cellular models, and those contained within a manageable number of cell types, ultimately aiming to address fundamental challenges in human health by understanding and intervening in disease mechanisms at an unprecedented scale.

"it's time for me to turn to another critical challenge the development of machine learning and it's applications to improving human health."

"curing disease is very hard because oftentimes by the time you discover the disease a lot of damage has already been done."

"Alzheimer's is probably closer to zero than to 80... there is an increasing number of people who believe there's a traditional hypotheses might not really explain what's going on."

"breast cancer is really not one disease it is multitude of cellular mechanisms all of which ultimately translate to uncontrolled proliferation."

"I think all of us aspire to an increased health span I would say which is an increased amount of time where you're healthy and active and feel as you did when you were 20."

"what we are doing it in seat rows actually flipping that around and saying here's this incredible repertoire of methods that bile engineers cell biologists have come up with let's see if we can put them together in brand-new ways with the goal of creating data sets that machine learning can really be applied on productively."

"the problem is is that oftentimes the way in which we generate the disease and the animal has nothing to do with how that disease actually comes about in a human."

"the ability for us to take a cell from any one of us you or me revert that's a skin cell to what's called stem cell status which is a what if it was called a pluripotent cell that can then be differentiated into different types of cells."