This reminds me of a problem I worked on while at Sandia National Labs: how to better inform nuclear reactor operators in a Fukushima-like disaster.

The Earthquake forced the reactor into a SCRAM procedure, shoving control rods into the mix to capture neutrons and slow the fission reaction, but the tsunami shut off the backup diesel power behind the pump. With the lights off and pump no longer providing cool water, the operators ran around with a physical battery like an EKG trying to restart circulation. What they missed, however, was information that might’ve informed when to abandon this pursuit. From simulations of nuclear meltdowns, we found key sensors within the reactor pressure vessel were essential to identifying the buildup of hydrogen gas (from super hot water interacting with the zirconium cladding on the control rods), and had the operators noticed these signals in the deluge of sensor data in the reactor, they could’ve dropped the battery and ran away before boom.

The task informs the data, and heterogeneity (in space, time, across cells) relevant for one task may be irrelevant for another. Another example is the method my mom and I used to improve the efficacy of chemo for her pancreatic cancer: heterogeneity in the phase of the cell cycle across cancer cells affects the efficacy of chemotherapies like nucleoside analogs that inhibit DNA synthesis during the S phase of the cell cycle. Where we can exert forcing on the system, such as with fluctuations in blood glucose that modulate the rate of cell cycle progression, we may be able to learn from and exploit the heterogeneity in cell cycle phase, maximizing the number of cells in the S phase of the cell cycle at the time of infusions. Here, the task was curing my mom, and the relevant data ended up being a few peepholes of ancillary, often supplementary, graphs in papers that suggested variation in glucose concentrations can lead to significant variation in the rate at which cells progress out of G1 and into S phase. While using a customized glucose-forcing regime in conjunction with nucleoside analog infusions, my mom showed one of the most extreme responses to chemotherapy, her stage IV tumor disappearing by her 17mo post-diagnosis scans.

With loops in automated labs, there’s often the foundation model task - just explore the space widely, capturing significant variation to improve model robustness. However, the economics may be better in the diagnostics or therapeutics tasks, in which case the economic value of a loop depends on the reliability of biomarkers as assays of efficacy and safety (let alone absorption, diffusion, metabolism, excretion, toxicity in vivo).

Loved reading your thoughts! My new favorite phrase: “heroic assumptions” 😂🫡