We are moving our newsletter to Substack for a better experience!
In Week #245 of the Doctor Penguin newsletter, the following papers caught our attention:
1. Spatial Proteomics. A successful example of using spatial proteomics, a technique that analyzes protein expression while preserving information about cell types and their location within tissue, to identify treatments for a fatal condition without effective therapy.
Nordmann et al. identified and validated JAK inhibitors as a treatment for Toxic Epidermal Necrolysis (TEN), a fatal drug-induced skin condition with a one-third mortality rate and no effective therapy. Using Deep Visual Proteomics (DVP), a tool that combines high content imaging, AI-guided cell segmentation and classification, and laser microdissection of individual target cells coupled with ultra-sensitive mass spectrometry-based proteomics, they analyzed protein patterns in specific cells from skin samples. The analysis revealed that the JAK/STAT pathway was markedly upregulated in TEN patients, which could be served as an actionable therapeutic target. After validating the effect of JAK inhibitors in vitro, they treated seven TEN patients with the inhibitors, including three who were previously unresponsive to high-dose systemic corticosteroids. All seven patients responded well to the inhibitors and were discharged in good health. This study demonstrates a successful translation of cell-type-resolved spatial proteomics findings into an effective treatment modality.
Read paper | Nature
2. Medication Errors. Drug-related errors are a leading cause of preventable patient harm in clinical settings. Among these, there are two critical types of drug administration errors: vial swap errors, where medication is incorrectly drawn from the wrong vial into a syringe, and syringe swap errors, where the drug is labeled correctly but the wrong syringe is administered to the patient.
Chan et al. developed and tested a wearable camera system that can detect medication errors in clinical settings before drugs are administered to patients. The system uses a head-mounted 4K camera and deep learning algorithms to identify and classify drug labels on syringes and vials, specifically to catch "vial swap" errors. The algorithms can detect vial swaps in real-time from videos wirelessly streamed to a local edge server with a GPU, enabling real-time auditory or visual feedback to alert providers about medication errors prior to drug administration. Testing was conducted across 13 healthcare providers in 17 operating rooms over 55 days at two hospitals. The system achieved 99.6% sensitivity and 98.8% specificity in detecting vial swap errors across 418 drug preparation events. Notably, the mean time between when a syringe was selected by a provider and when the drug was given was 9.9 ± 7.2 s from the data collected in this study, while the system took less than 25 ms on an NVIDIA A40 GPU. This suggests the system could be used to identify syringe swaps prior to drug injection in drug delivery events recorded in the patient dataset.
Read Paper | npj Digital Medicine
3. RNA Foundation Model. Natural genomic sequences exhibit limited diversity due to evolutionary constraints, with high sequence similarity within species. Current genomic foundation models apply generic self-supervised learning strategies borrowed from text and vision domains, such as masked language modeling and next token prediction. However, these approaches fail to incorporate biological knowledge, leading models to focus on reconstructing non-informative regions of sequences rather than learning meaningful biological representations.
Fradkin et al. developed Orthrus, an RNA foundation model that learns through a biologically motivated contrastive learning approach. The model is pre-trained on mature RNA sequences by maximizing embedding similarity between two types of related sequences: splice isoforms (different RNA versions from the same gene) across 10 species, and orthologous transcripts (related genes that evolved from a common ancestor) from over 400 mammalian species. This approach allows Orthrus to learn both evolutionarily related sequences and sequences responsible for shared functions between splicing isoforms. By training on mature RNAs with high functional importance and sequence conservation, Orthrus specifically focuses on sequence regions with high information content. The model outperforms existing approaches on five mRNA property prediction tasks while requiring only a fraction of fine-tuning data—demonstrated by achieving strong performance with as few as 45 labeled examples for RNA half-life prediction. This result also shows self-supervised pretraining can address the data efficiency challenges in genomics.
Read Paper | bioRxiv
4. LLM. How are healthcare applications of large language models (LLMs) currently evaluated?
In this systematic review, Bedi et al. analyzed 519 studies published between January 2022 and February 2024 examining healthcare applications of LLMs. The analysis revealed several key limitations in current evaluation approaches. First, only 5% of studies used real patient care data, with most relying on medical examination questions, clinician-designed vignettes, or expert-generated questions. Second, evaluations predominantly focused on medical knowledge tasks and diagnostic capabilities, while administrative tasks such as billing codes, prescriptions, clinical referrals, and note-taking remained understudied. Third, while 95.4% of studies assessed accuracy, few examined fairness, bias, toxicity, or deployment considerations. Among the specialties, internal medicine, surgery, and ophthalmology were the most frequently studied specialties, while nuclear medicine, physical medicine, and medical genetics received minimal attention. Based on these findings, the authors recommend six improvements for future LLM evaluations: using real patient data, standardizing evaluation methods, prioritizing high-impact administrative tasks, addressing specialty gaps, conducting financial impact assessments, and establishing better frameworks for quantifying bias and reporting failure modes.
Read Paper | JAMA
-- Emma Chen, Pranav Rajpurkar & Eric Topol