The Genetics Podcast

The episode is packed with substantive technical content about delivery mechanisms for genetic medicine, including specific explanations of LNP biology, machine learning optimization approaches, and tissue-specific targeting strategies. However, there are some sections with repetition (e.g., reiterating the delivery challenge multiple times) and the host's closing remarks add minimal new insight, preventing a higher score.

“The real question is what we're trying to do is deliver nucleic acid so that we can reprogram or complement a mutation in a cell. And we're packaging the nucleic acid because the nucleic acids themselves would be degraded by the body...hundreds of millions of years, the human body has been pushing out foreign nucleic acids to prevent that from happening.”

“Within three or four rounds we're at a very high forming lipid nanoparticle. And that in vivo work is done quite efficiently and the formulation done quite efficiently...Four rounds is about six months.”

The guest presents original frameworks around modularized LNP design, the machine learning feedback loop for optimization, and the explicit separation of tropism (ionizable lipids) from targeting (external ligands). The thinking on exploiting the corona hypothesis and the concept of 'trans-endothelial delivery' are fresh. However, the core LNP and viral vector comparison is standard industry knowledge, and some concepts (e.g., 'endosome as a jail') are pedagogical restatements rather than novel.

“You take advantage of that inactivation by then disassembling and then fusing with the endosome and then getting the nucleic acid out. So that's kind of an engineering of that system.”

“We built was a vehicle that would get to, say, the spleen and detarget from the liver, because you don't want it going to the liver if it doesn't have to. And then we could add in targeting agents on top of the lipid nanoparticle that then specify which cell type you're going to get into.”

Dr. Jagesh Shah is an exceptionally well-qualified guest with deep expertise spanning systems biology at Harvard Medical School, senior scientific leadership at two funded biotech companies (Cobalt at Flagship, Sana), and now as CSO at Mirai Bio. He has clearly built and shipped technologies at scale, demonstrated by published work on LNP optimization and partnerships with major pharmaceutical players. He speaks with authority grounded in both academic rigor and practical industry execution.

“I was a faculty member, I was a basic scientist...and that turned into a sabbatical for me. And that then turned into the first company that we started, which was Cobalt at flagship Pioneering and then merge into Sauna.”

“I left Sana in 2024. I found that that itch for in vivo delivery was just something that I had not...And I came in and we really tried to reinforce this idea of the in vivo delivery, the ML and try to really bring that to the table.”

The episode includes concrete technical specifics: ionizable lipids as the linchpin of LNPs, four-component formulations, mouse biodistribution screening across named tissues (liver, spleen, pancreas, lung, heart, bone marrow), six-month timelines for three-to-four optimization rounds, 12-16 KB mRNA cargo capacity, and named therapeutic targets (CAR-T, B cell depletion in monkeys, GLP-1 delivery, dystrophin complementation). However, proprietary constraints limit disclosure of exact molecular structures, formulation ratios, and clinical trial data, preventing a perfect score.

“these lipid nanoparticles are put together with four or sometimes more components. And the proportion of each one of these is what we refer to as formulation.”

“we've played with some fairly big mRNA so self amplifying mRNA which has been the basis of a recent vaccine that is about 12 to 16 KB of MRNA.”

Patrick Short asks intelligent, technically informed follow-up questions that probe the guest's thinking (e.g., on ML exploitation vs. exploration, modularity of LNPs, the role of protein corona), and he occasionally pushes back or reframes (e.g., on the platform business model analogy). However, he misses some opportunities to challenge claims or dig into failure modes. For instance, he doesn't push back on whether six months to optimization is actually fast, or probe the limits of the in vivo mouse model for predicting human biodistribution. Some questions are somewhat leading rather than genuinely open.

“Are you able to multiplex them in some way? Like, because I'm imagining you have this. When you use LLMs, you can set the heat, which will tell you just how out of the box you want it to think.”

“Are there parts of the process where the computational methods are resisting biology in some ways where it's hard to predict and you need a lot more wet lab data?”