ENPICOM

Six months ago, getting a workflow from 20 clicks to 5 made scientists happy. Today, five clicks is too many. That's how fast expectations are shifting inside big pharma, according to Melody Shahsavarian, Senior Director of Data Strategy and Digital Transformation at Eli Lilly and the guest on our new Models & Molecules episode 🎙️ Its also our first in-person one, recorded in San Diego. Melody started in the wet lab, spending eight years on antibody discovery until the sheer volume of data pulled her into the digital side. Before Lilly, she led NGS and AI for biologics discovery at Sanofi. We talked about agentic AI and what it really takes to drive digital transformation in a large pharma organization, so that it turns into productivity and, eventually, medicines reaching patients faster. The episode is packed with practical insight: why digital transformation fails if IT drives it alone, why scientists adopted LLMs faster than almost any tool before them, what makes an agent production grade, and how Lilly is validating the tools scientists now build themselves. Melody was refreshingly practical about all of it, including the risks of agentic workflows, which are already showing up today. We closed, as always, with the contrarian question. Melody's answer is one she finds herself defending almost daily. Worth a listen.

Everyone is talking about AI agents. Few people can tell you what it takes to make them work inside a 150-year-old pharma company. Our next Models & Molecules episode is with someone who can. Its our first in-person episode, recorded in San Diego, with a guest from Eli Lilly. The topic is digital transformation and agentic AI in big pharma. I went in with a few assumptions. I came out with fewer. Out Tuesday. Stay tuned 🎙️

Science is meant to be transparent and subject to review and criticism. Industry, where more and more science is now happening, treats methods as trade secrets. Iddo Weiner argues that sharing models, datasets, methods, and even negative results is both a moral imperative and a way to earn trust with customers who are scientists themselves. Watch/listen here: https://lnkd.in/er5FDt3q

Bigger models and more compute eventually run into a data wall: there isn't enough well-labeled experimental data to post-train on (with structural data, binding affinity measurements etc.) In Episode 7 of #ModelsnMolecules, Iddo Weiner, CSO and co-founder of Converge Bio, shares his take on data bottlenecks in antibody design, and on how active learning turns a fixed experimental budget into a way of generating the right data instead of simply more of it. Watch/listen here: https://lnkd.in/egwQBrkR

When frontier models don't stay frontier for more than a month 😅, and open weights keep becoming more of a trend, I do wonder: what about bio? Is bio following the same path? And if so, where is the moat? I put that to Iddo Weiner, co-founder and CSO at Converge Bio, in the new episode of Models & Molecules. His answer isn't the comfortable one. Not the model, he's very clear about that. And not really the data either, which is where most people land. Data is a bit better than the model as a defensible position, but only a bit. His take: the moat is being able to put all of it together. State of the art model under the hood, whoever built it. Your own data on top. An application layer people actually want to use. And enough proximity to the customer that you understand what they are trying to do. He thinks nobody is close to solving that yet. He also makes a case for open science, far from idealism, concrete and pragmatic. Plus: what if the most impressive looking results in AI for bio are just homology in disguise? And why better models will mean more wet lab work, not less.

A smaller, more diverse dataset can outperform a billion data points focused on a single thing. On Ep 6 of #ModelsnMolecules, we get into the data strategy behind Boltz and what sets companies apart in this highly competitive field. As Talip Uçar explains it, models learn better when they can draw on commonalities across different domains and modalities. So it's not the amount of data that matters, but the diversity and the nature of it. If you missed it, watch or listen to the full conversation here: YouTube → https://lnkd.in/ePieqkVU Spotify → https://lnkd.in/eFpqGnqW Apple Podcasts → https://lnkd.in/eKyePZj2