The FDA's internal AI assistant, Elsa, just got a new brain. Not because the old one stopped working, but because the White House told every federal agency to drop Anthropic's Claude after a fight with the Pentagon over AI in autonomous weapons. The replacement? Google's Gemini (maybe).

Here's what makes this really messy. Elsa isn't a standard chatbot. It's a custom-built system (retrieval-augmented generation, basically AI that pulls from the FDA's own document stores to answer questions) wired deep into how reviewers analyze drug submissions.

Swapping the foundation model means revalidating the entire pipeline: how documents get retrieved, interpreted, and summarized. And Elsa already had a hallucination problem before the rush job.

Now, sponsors are being told to mark everything confidential and ask the FDA which model is even reviewing their data.

This is the part nobody in drug development planned for: your regulatory process is now downstream of geopolitics. The AI tools we're building into the most consequential decisions in medicine can be ripped out overnight, not for scientific reasons, but political ones.

How do you build a durable AI infrastructure for drug review when the foundation can change with an executive order?

For more details: Full Article

Pharos iBio, the South Korean AI drug discovery company developing Rasmotinib for acute myeloid leukemia, appears to have benefited from investor optimism around potential licensing discussions at Bio China.

No transaction was announced, but the combination of business development outreach, big pharma meetings, and China market exposure was enough to lift expectations.

All others, no real news I could see in the past week to explain some double digit change.

Select the right answer! (See explanation below and source)

Acelot registered three Phase 1 trials in Australia for ACE-2223, a first-in-class AI-designed small molecule targeting misfolded TDP-43 in ALS. The trials cover single and multiple ascending doses, formulation bioavailability, and food effect studies, with dosing expected to begin as early as March 2026. Acelot's platform uses generative AI and molecular dynamics simulations to design molecules against misfolded proteins, a target class historically considered undruggable. [Trial 1] [Trial 2] [Trial 3]

Insilico's AI-discovered CKD anemia drug ISM4808 dosed its first patient in a Phase 1 trial just three months after being out-licensed to TaiGen Biotechnology. The oral HIF-PHD inhibitor was designed using Insilico's Chemistry42 platform and targets the Greater China market, where anemia affects more than one in seven CKD patients. Insilico is also independently advancing a second Chemistry42-derived PHD inhibitor, garutadustat, which entered Phase 2 for inflammatory bowel disease in January 2026. [Link]

Lawrence Livermore scientists used machine learning on 11,000 veterans' health records to identify existing drugs that may extend survival in ALS. Published in The Lancet Digital Health, the study combined causal inference with ML to evaluate 162 medications and found 27 with statistically significant effects on mortality risk, with statins and PDE5 inhibitors among the classes showing consistent positive associations. The team plans to open-source its software pipeline so other researchers can apply the methods to their own datasets. [Link]

An ARPA-H funded project is combining human liver organoids with AI to predict drug toxicity at 90% accuracy, up from roughly 50% with traditional methods. The DATAMAP initiative, led by Inductive Bio and including UMichigan's TorchBio, Amgen, Cincinnati Children's, and Baylor, uses robotically generated organoid imaging to run roughly 20,000 toxicity tests per batch in an afternoon. The five-year goal is to enable FDA regulatory approval for first-in-human trials based on computational safety data rather than animal models. [Link]

A team at the University of Illinois just simulated the entire life cycle of the simplest free-living cell on Earth. Not a sketch of it. Not a snapshot. The whole thing.

All 493 genes being transcribed, every ribosome assembling and bumping through a packed cytoplasm, a chromosome replicating and splitting into two daughters, metabolism churning through nucleotides, and a membrane growing lipid by lipid until the cell pinches in half. In three spatial dimensions, tracked across time, at 10-nanometer resolution.

This, is all for a bacterium that lives and dies in under two hours.

To understand why this matters, you need to know what JCVI-syn3A actually is. It's a bacterium with a synthetic genome that's been stripped to only the genes required for independent life. No regulatory frills. Fewer genes of unknown function than almost any organism studied.

It was engineered at the J. Craig Venter Institute as a kind of biological ground truth - the minimum viable cell. If you want to build a computational model of a living system from the bottom up, this is where you start.

And the computational architecture is nothing short of amazing. The 4D whole-cell model (4DWCM) hybridizes four simulation methods running on two GPUs simultaneously. Reaction-diffusion equations handle spatially localized chemistry on a cubic lattice. Brownian dynamics on a separate GPU simulate the chromosome as a polymer of over 54,000 beads, complete with condensin-like loop extrusion and topoisomerase-mediated strand passage. Ordinary differential equations drive metabolism. A stochastic chemical master equation governs transcription and tRNA charging.

All of this synchronizes through a hook algorithm that interrupts the main simulation every 12.5 milliseconds of biological time. Each simulated cell cycle takes four to six days of wall-clock time. The 50 replicates presented here required roughly 15,000 GPU hours on NVIDIA A100s.

The predictions land close to real data. The model predicts a doubling time of 105 minutes, matching the experimentally measured value. It recovers the origin-to-terminus DNA replication ratio seen in sequencing.

It finds that roughly 55% of ribosomes and 70% of RNA polymerases are active at any given moment. And because every protein, RNA, and ribosome is tracked as an individual particle in 3D, the model automatically predicts how molecules partition between daughter cells after division. That partitioning turns out to be stochastic, approaching a binomial distribution.

Every newborn cell is slightly different from its sibling.

Now, the caveats. This is a bacterial cell with fewer than 500 genes. Human cells have over 20,000 and are orders of magnitude larger. The model still can't simulate polysomes (multiple ribosomes reading the same mRNA at once), which limits accuracy for large proteins. It relies on an artificial 12 piconewton force to physically push daughter chromosomes apart because the real partitioning mechanism in Syn3A is unknown. And testing multiple parameter sets is prohibitively expensive when each run takes nearly a week.

But here's what connects this to the broader current in drug discovery. The concept of an "AI Virtual Cell," championed by CZI, Recursion, Ginkgo, and others, envisions computational systems that can simulate how cells respond to perturbations like drugs, gene knockouts, or disease mutations.

Most of those efforts use AI foundation models trained on massive omics datasets to make statistical predictions about cellular behavior. What the Illinois team has built is something different and complementary: a physics-based, mechanistic simulation where every reaction follows known kinetics and every molecule obeys diffusion laws.

The paper itself draws this connection, noting that virtual cells built from AI/ML and whole-cell models like the 4DWCM could serve as complementary methods, with AI assembling cell states from data and physics-based models predicting the chemical and physical processes driving changes between those states.

The question nobody has answered yet is whether these two approaches, statistical and mechanistic, can actually meet in the middle at a scale that matters for human disease.

Physics-based models give you causality, but can't scale. AI models scale beautifully but struggle to tell you why a cell responded the way it did. Somewhere between a 493-gene bacterium simulated on two GPUs and a 20,000-gene human cell simulated on a foundation model, there's a convergence point that could change how we understand drug response at the cellular level. We just don't know how far away it is.

Sources: [Research Article]

AML is a cancer of immature myeloid cells, which build up in the bone marrow and interfere with normal blood cell production. [Source]