Fireside Chat, Interview
Digital Biology with insitro's Daphne Koller
- The field is expected to soon reach a state where "everyone's an expert in language models" due to the increasing accessibility of AI concepts in biology.
- Machine learning will be deployed in "truly meaningful" ways within life sciences as datasets expand to support "really interesting" methods.
- A "disproportionate impact" is predicted for the intersection of machine learning and biomedical data compared to other applications of AI.
- The ability to "generate data on spec" using engineered cells is expected to create significant discovery opportunities and novel machine learning challenges like active learning and experimental design.
- Measuring a genome-wide CRISPR screen in "10 or 12 plates in two weeks" is identified as a foundational step toward deciphering genotype-phenotype connections for therapeutic interventions.
- The platform's technology is described as "intrinsically AI enabled," making it impossible to operate without integrated AI for tasks such as cell segmentation and barcode calling.
- The system's "latent space" for human biology is expected to function similarly to large language models, allowing queries on how disease-causing genes alter biological states and how treatments can restore a "healthy state."
- A foundation model approach is anticipated to continuously improve in competitive understanding of biological foundations as more data is fed into the system.
- A systematic process to move from disease decisions to "meaningful interventions" is planned to result in "delivered medicines to patients" by "the end of the decade."
- Improving biological tools, progressing from siRNA to "CRISPR base editing" and now "CRISPR prime" which replaces entire genome regions, are expected to unlock the ability to tackle more diseases over time.
- AI is expected to address clinical trial failures by enabling work in "human and human derived systems" to mitigate differences between mice and humans.
- Machine learning will be utilized to "bridge" cellular data and human clinical data within "representation space" and "genetic space" to interpret complex, high-dimensional biological data.
- Hiring plans include recruiting "translators" with expertise in the middle of machine learning and life science to establish a shared language and vision.
- A culture of open engagement fostering "naive questions" and "naive suggestions" is expected to generate the "best ideas" through rigorous hiring practices.
- Applying machine learning to experiments reveals a signal indicating "what was the technician who actually did the experiments" due to slight variations in cell behavior.
- Significant time will be spent building robots to eliminate human variability, as machines are capable of performing tasks "the same thing over and over again."
- The "next frontier" of AI impact is expected to be in touching the physical world, a challenge deemed significantly harder than building software like chatbots.
- Successfully navigating the physical world frontier is anticipated to demonstrate the "magnitude of the impact" AI can achieve.
- An era of "digital biology" is envisioned where biology is measured at "unprecedented stability and scale," interpreted via machine learning, and engineered using tools like CRISPR to force new behaviors.
- Applications beyond human health are expected in "agriculture" to create crops resistant to drought and severe weather to help "feed 10 billion people."
- Opportunities are anticipated in the environment to potentially improve carbon sequestration using plants or algae.
- The convergence of AI and biology is described as a unique "moment in time" to make a significant global difference using tools that did not exist five years ago.