I often receive newsletters from subscriptions I maintain on subjects that might interest me (as opposed to those which do–IT) and I got two recently that piqued my interest.
The first: a paper published in Nature talks titled “Agenerative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial”, while very technical in nature, was best summarized in the popular press as “First AI-Designed Drug Nears Final Trials Before Approval”.
Now, this is of interest.
Given that traditional drug development timelines cost about $2 billion, AI accelerates candidate nomination by shortening preclinical phases to 12–18 months, according to Insilico’s Pharma.AI.
Ground News
Insilico Medicine is a Hong Kong startup using AI to design new drugs and, in collaboration with a pharmaceutical company, Hisun Pharmaceutical Co., Ltd., plans to start Phase 3 trials sometime in 2026.
Phase 1 [is] for safety in small groups, Phase 2 [is] for efficacy/side effects in patients, Phase 3 [is] for large-scale efficacy/comparison) followed by regulatory review (FDA) and finally, post-market Phase 4 monitoring for long-term effects…
Google AI
The savings in time, and the ability to get to clinical trials quickly, is the most amazing part of the story:
Whereas traditional early-stage drug discovery typically requires 2.5 to 4 years, more than 20 of Insilico’s internal programs initiated between 2021 and 2024 achieved PCC nomination in just 12 to 18 months on average, with only about 60–200 molecules synthesized and tested per program.
EurekAlert!
The use of AI in this case is an accelerator for finding disease target proteins, modeling and screening compounds that might bind to those targets, predictions of absorption, and delineation of production processes.
Should this particular drug get through the remaining clinical trials and FDA acceptance, we will have had a major demonstration of the value of AI to medical science.
The second newsletter notice: also in Nature, a paper about the application of AI to diagnosing a host of diseases based on data gathered during sleep.
Polysomnography (PSG)—the gold standard for sleep analysis—captures rich physiological signals but is underutilized due to challenges in standardization, generalizability and multimodal integration. To address these challenges, we developed SleepFM, a multimodal sleep foundation model trained with a new contrastive learning approach that accommodates multiple PSG configurations.
Nature
The summary: the AI agent, named SleepFM, uses data gathered using standard Polysomnography protocols (wearing equipment during sleep that captures brain, heart, respiratory, eye and muscle signals.
The predictive results so far are nothing short of amazing.
From one night of sleep, SleepFM accurately predicts 130 conditions with a C-Index of at least 0.75 (Bonferroni-corrected P < 0.01), including all-cause mortality (C-Index, 0.84), dementia (0.85), myocardial infarction (0.81), heart failure (0.80), chronic kidney disease (0.79), stroke (0.78) and atrial fibrillation (0.78).
Nature
There was also work done to apply SleepFM to already captured sleep study records–in this case from 25 years of Stanford Sleep Medicine Center records–and the results are not as stellar there, but the researchers are working on further interpretation. They are also working on a wearable to record the sleep data.
Amazing what AI has done for our health so far.