AI Health
Friday Roundup
The AI Health Friday Roundup highlights the week’s news and publications related to artificial intelligence, data science, public health, and clinical research.
August 1, 2025
In this week’s Duke AI Health Friday Roundup: machine learning models unlock “undruggable” targets; “virtual lab” of AI assistants develops nanobodies; lead linked to memory problems many years after exposure; why anteaters keep coming; Bluesky’s performance among scholars; AI casually defeats “not a robot” CAPTCHA; LLM hallucinates a diagnosis, other details into patient medical record; much more:
AI, STATISTICS & DATA SCIENCE
- “Current methods of making drugs almost exclusively rely on finding drug “keys” that plug neatly into troublemaking protein “locks” and thus inherently require the protein target to have a structure. Finding binders — i.e., drugs — for disordered proteins is hard when what is supposed to be a “lock” shape is a single piece of slippery spaghetti….But as machine-learning-powered protein structure and design tools came online over the past few years, this problem suddenly became solvable.” At STAT News, Brittany Trang reports on recent work, including a paper published last week in Science, that builds on new machine-learning approaches to address formerly “undruggable” targets.
- “Our randomized controlled trial demonstrates the benefits of ASCE in both OSCE performance and self-perceived emotional preparedness. Notably, our findings highlight ASCE as a tool willingly adopted by students to enhance their clinical skills development….Consistent with expectations, ASCE training largely enhances emotional preparedness by offering structured practice opportunities. Surprisingly, despite the inherently high-pressure context of the examination’s oral format and direct evaluator observation, we observed a small but significant reduction in stress levels.” A research article published in NEJM AI by Lavigne and colleagues presents findings from a randomized trial evaluating an AI-assisted training system to prepare medical trainees for structured exams designed to evaluate their clinical competence.
- “…we expand the capabilities of LLMs for science by introducing the Virtual Lab, an AI-human research collaboration to perform sophisticated, interdisciplinary science research. The Virtual Lab consists of an LLM principal investigator agent guiding a team of LLM scientist agents through a series of research meetings, with a human researcher providing high-level feedback. We apply the Virtual Lab to design nanobody binders to recent variants of SARS-CoV-2.” A research article by Swanson and colleagues (available as an accepted but uncorrected manuscript at Nature) describes the creation of a “lab” of LLMs with an LLM “PI” that, with human oversight, design novel nanobodies for COVID (H/T moorejh.bsky.social).
- “…the patient noticed the entry that had introduced the diabetes diagnosis was listed as a summary that had been ‘generated by Annie AI.’ The record appeared around the same time he had attended the hospital for a severe case of tonsillitis. However, the record in question made no mention of tonsillitis. Instead, it said he had presented with chest pain and shortness of breath, attributed to a ‘likely angina due to coronary artery disease.’ In reality, he had none of those symptoms.” Fortune’s Beatrice Nolan reports on a UK patient who was invited to a diabetes screening based on information that an AI had hallucinated and inserted into his patient record.
- “Computational models of signaling networks provide frameworks for predicting how molecular cues guide cell decisions. But they are typically limited by manual curation from incomplete literature. Here, we test whether general-purpose large language models (LLMs) generate accurate models of signaling networks.” A research paper by Tewari and colleagues, available as a preprint from bioRxiv, evaluates the performance of large language models tasked with modeling cellular signaling networks.
BASIC SCIENCE, CLINICAL RESEARCH & PUBLIC HEALTH
- “Other animal groups contain examples of unrelated species evolving similar traits, a phenomenon called convergent evolution. Take crustaceans, which have evolved a crablike body plan at least five different times to online notoriety. But the anteater example is striking; crustaceans’ five so-called carcinizations occurred over a couple of hundred million years, but the dozen iterations of the anteater body plan popped up in just 66 million years, less than half the time.” Science’s Jake Buehler explores nature’s curious insistence on re-creating the anteater in various similar forms – again and again and again. (Maybe because coming up with a different design is such aard vark?)
- “Once contact with a pathogen has occurred, it might be too late for the immune system to react. Here, we asked whether anticipatory neural responses might sense potential infections and signal to the immune system, priming it for a response.” A research article published in Nature Neuroscience by Trabanelli and colleagues reports findings from a virtual-reality based study that shows an anticipatory activation of the immune system when test subjects perceived visibly ill VR avatars entering their personal space.
- “The findings add to the evidence connecting poor health outcomes with earlier life lead exposure. ‘The precipitous decline in atmospheric lead exposure in the last quarter of the 20th century may help to explain the declining incidence of dementia in the U.S.,’ Brown and colleagues noted.” An article by MedPage Today’s Judy George reports on a study recently presented at the Alzheimer’s Association International Conference that found exposure to atmospheric lead (including from vehicle exhaust) in early life is linked to memory problems many years later (H/T @ryanmarino.bsky.social).
- “Three key findings emerge. First, even modest daily step counts were associated with health benefits. Second, 7000 steps per day was associated with sizeable risk reductions across most outcomes, compared with the reference of 2000 steps per day. Third, even though risk continued to decrease beyond 7000 steps per day, it plateaued for some outcomes.” Perhaps not surprising, as there never was any real basis for it, but a new systematic review published in Lancet Public Health by Ding and colleagues suggests that the ubiquitous benchmark of 10,000 steps may not be the optimal goal for daily exercise.
COMMUNICATION, Health Equity & Policy
- “On Friday, OpenAI’s new ChatGPT Agent, which can perform multistep tasks for users, proved it can pass through one of the Internet’s most common security checkpoints by clicking Cloudflare’s anti-bot verification—the same checkbox that’s supposed to keep automated programs like itself at bay.” At Ars Technica, Benj Edwards has bad news for anyone relying on those “I am not a robot” checkboxes to hold back the ‘bots.
- “…this study presents the first large-scale analysis of scholarly article dissemination on Bluesky, exploring its potential as a new source of social media metrics. We collected and analysed 87,470 Bluesky posts referencing 72,898 scholarly articles from February 2024 to April 2025, integrating metadata from the OpenAlex database. We examined temporal trends, disciplinary coverage, language use, textual characteristics, and user engagement. A sharp increase in scholarly activity on Bluesky was observed from November 2024, coinciding with broader academic shifts away from X. Posts primarily focus on the social, environmental, and medical sciences and are predominantly written in English.” A research paper by Zheng and colleagues, available as a preprint from arXiv, presents an analysis of BlueSky as a vehicle for disseminating information about science in the wake of a significant uptick in its scholarly user base.
- “…concerns have been raised about [social media’s] putative societal impacts, ranging from undermining mental health and exacerbating polarization to fomenting violence and disrupting democracy. Despite extensive research, consensus on these effects remains elusive, with observational studies often highlighting concerns while randomized controlled trials (RCTs) yield conflicting or null findings. This review examines how the complexity inherent in social systems can account for such discrepancies, emphasizing that emergent societal and long-term outcomes cannot be readily inferred from individual-level effects.” A review article by Bak-Coleman and colleagues, available as a preprint on arXiv, addresses the complexity of untangling the impacts of social media consumption.
