From Models to Medicine

Episode 13: The Microbiome is Messier Than You Think

Jenny Yang [https://www.linkedin.com/in/jennyyang259/] is the co-founder and CEO of Outpost Bio, where her team is working to make human microbiology computable. In this episode, she breaks down why bias in ML models is so easy to miss. High overall accuracy can hide terrible performance on specific subgroups, and in healthcare, that gap has consequences. She traces the problem upstream, from skewed training datasets to the way clinical definitions themselves carry historical bias, and explains the real trade-offs involved in trying to correct for it. We also get into what makes the microbiome such a hard problem is that our microbiomes can differ by up to 90% from person to person. Jenny walks us through how Outpost Bio's "lab in the loop" model tightly integrates wet lab experiments with AI to generate better, less biased data from the ground up, and why rigorous external validation is the thing she'd tell every biotech founder to prioritize before anything else.

Episode 12: Your Life Sciences Data Isn't Ready for AI

Bogdan Knezevic [https://www.kaleidoscope.bio/company] is the CEO and co-founder of Kaleidoscope Bio, and he's seen enough failed AI implementations to know where they almost always break down. In this episode, he walks us through what minimum viable data standardization actually looks like in practice, why consistent naming conventions and structured data entry matter more than people want to admit, and what every biotech CEO should ask their team before writing another AI budget line. We also get into a guardrails conversation and Bogdan is direct about what happens when autonomous agents operate without proper permissions. He closes with a sharp framework for deciding what to build in-house versus hand off, with some great resource hand-offs. * Data vs keys [https://blog.kaleidoscope.bio/unlocking-the-importance-of-structure-data-vs-keys/] * AI needs context [https://blog.kaleidoscope.bio/ai-cant-reason-with-context-it-doesnt-have/] * Iterating your way to success via better data management [https://blog.kaleidoscope.bio/experimenting-your-way-to-success-data-management-in-biotech/] * The data maturity ladder [https://blog.kaleidoscope.bio/where-is-your-biotech-on-the-data-maturity-ladder/] * Before AI there was good data [https://blog.kaleidoscope.bio/before-ai-there-was-good-data/]

Episode 11: From the Gut to the Brain - Rethinking Parkinson's with AI

In this episode of From Models to Medicine, we sit down with Minna Schmidt [https://www.linkedin.com/in/minna-schmidt-80390a97/], a postdoctoral researcher at the Buck Institute for Research on Aging. Minna walks us through Braak's hypothesis and the emerging "brain-first vs. body-first" framing of the disease, explaining how symptoms can appear up to 30 years before a clinical diagnosis is ever made. We also get into the data side of the work. Minna uses a dataset with over 54,000 participants and talks honestly about what AI actually does and doesn't unlock when you're staring down that volume of microbiome and clinical data. She uses LLMs to organize her thinking, speed up literature reviews, and learn basic programming, while being clear-eyed about where the field's biggest bottleneck actually is: not the tools, but the data itself. ---------------------------------------------------------------------- This episode was sponsored by ⁠CleanSpace⁠ [https://www.cleanspaceus.com/]. ⁠CleanSpace⁠ [https://www.cleanspaceus.com/] designs, manufactures, and installs advanced controlled environments—delivering complex projects months faster with guaranteed costs and uncompromising performance. Please contact Chelsea for more information or with any questions at CLauridsen@CleanSpaceus.com.

Episode 10: When AI Gets It Wrong, Patients Pay the Price

In this episode of From Models to Medicine, we sit down with Sal Tejani [https://www.linkedin.com/in/sal-tejani-41418870/], Associate Director for Field Medical Affairs at Regeneron*, who started his career catching dangerous prescription errors at CVS and never lost the instinct for finding the lever that actually moves things. Today that instinct is pointed squarely at AI; how to use it, when to trust it, and when it will absolutely get you into trouble. Sal gives us an honest, practitioner-level view of what AI looks like inside a major pharma company: the tools that are actually useful, the guardrails that are non-negotiable, and the human judgment that no model has figured out how to replace yet. Plus, he closes with a personal story that reframes the whole conversation about why any of this actually matters. This episode was sponsored by CleanSpace [https://www.cleanspaceus.com/]. CleanSpace [https://www.cleanspaceus.com/] designs, manufactures, and installs advanced controlled environments—delivering complex projects months faster with guaranteed costs and uncompromising performance. Please contact Chelsea for more information or with any questions at CLauridsen@CleanSpaceus.com. *Thoughts brought up on this podcast do not represent the views of Regeneron.

Episode 9: Mitochondria, Machine Learning, and a Few Hard Lessons

Rachel Jacobson [https://www.linkedin.com/in/racheldevayjacobson/] has spent her career moving between some of the most demanding corners of life sciences before founding Powerhouse Biology. In this episode, she traces that journey and explains why, after all of it, she keeps coming back to mitochondria. We get into what it actually takes to bridge biology and machine learning inside a lab culture, why asking "stupid questions" across disciplines is a feature and not a bug, and what she had to unlearn from traditional drug development to work effectively alongside ML engineers. We also dig into data design. Rachel makes a sharp case that data passing standard biological QC is not the same as data that's ready for a machine learning model. Uneven plate layouts, cell debris, different scientists handling samples can all create batch effects that quietly break your model before it ever sees a hypothesis worth testing. She connects all of this to a bigger argument about why human biological variability needs to be built into preclinical pipelines from the start, and why ML might finally give scientists the tools to do that seriously.