Lang Li

BACKGROUND Evidence for medication use in pregnancy and lactation is sparse, and clinicians and patients are increasingly encountering large language model (LLM)–generated summaries of drug information. Little is known about how end users perceive such summaries or what they need from them. The Maternal and Pediatric Precision in Therapeutics Knowledge Portal (MPRINT-KP) indexes more than 750,000 maternal and pediatric pharmacology papers, and researchers are developing LLM-generated drug summaries to complement it. OBJECTIVE Explore the information needs of pregnant and lactating individuals and the clinicians who care for them and identify user-centered design suggestions for LLM-generated drug summaries. METHODS We conducted a qualitative descriptive study using semi-structured video interviews (20–45 minutes) with 8 individuals (all pregnant) and 3 clinical providers (a clinical pharmacist, a maternal-fetal medicine physician, and an obstetrician-gynecologist) in February 2026. Participants reviewed a roughly 4000-word, web-formatted LLM-generated summary of methadone use in pregnancy and lactation, synthesized from 52 full-text articles cited in the LactMed methadone record (CASRN: 76-99-3) and commented on its content and format. Transcripts were analyzed using inductive thematic analysis, and findings were organized by the three study aims into six themes, from which we derived design suggestions. RESULTS Safety was ranked as the top concern by 9 of 11 participants; all 8 patients framed safety in terms of risk to the infant or fetus. Most participants cross-checked medication information across several sources. No participant expressed unconditional trust in AI-generated information and non would act on it without human verification. If their clinician endorsed it, 5 of 8 patient would be more receptive to it. Participants perceived the summary as clinician-oriented: At least one unfamiliar term was flagged by 7 of 8 patients, and 7 of 11 judged it too long. Clinicians found the language appropriate but wanted dosing (3/3), indications and risk–benefit framing (2/3), monitoring parameters (1/3), and noted that the summary lacked the clinical context they would give in counseling. Eight of 11 found the in-text blue hyperlinked citations disruptive or confusing and 7 of 11 wanted tables or figures. Six themes emerged across the three aims: information needs and source preferences (Aim 1); safety comes first, and safety means the baby, and conditional trust and triangulation (Aim 2); and content, formatting, and visual aids (Aim 3). CONCLUSIONS A single LLM-generated summary could not meet the needs of both patients and clinicians. Design suggestions include audience-specific layers (a plain-language, infant-focused patient version and a clinical version with dosing and monitoring), footnote-style traceable citations, diverse presentations, a short summary at the top, explicit AI-generation and human-review disclosure, and risk–benefit context. For drugs like methadone, where stopping treatment increases maternal risk, a patient-facing layer should pair infant risk with maternal benefit of continuation of treatment. These findings are based on a small sample and a single drug and should be tested with other medications and a larger, more diverse sample.

Motivation: Identifying effective gene combinations is critical for developing combinatorial cancer therapies. Synthetic lethality (SL) has been widely used to characterize genetic interactions based on cell viability, but predictive models are often limited by the availability of combinatorial double-knockout (CDKO) data for individual cancer cells. How to borrow information from CDKO experiments in other cancer cells to improve prediction in a targeted cancer cell remains an important challenge. Results: We investigated the CDKO effects of paired genes using both SL as an indirect measure and gene combination effect as a direct measure. We developed TL-iGeneCombo, a transfer learning approach, to borrow data from CDKO experiments in other cancer cells for the prediction of SL and gene combination effect in a targeted cancer cell. Applying TL-iGeneCombo improved the accuracy of SL classification by 19% and increased the Spearman correlation of gene combination effect scores by 29%. Using data and models built from the first CDKO experiment in a prostate cancer cell, 22RV1, we further designed gene pairs for a second CDKO experiment in the same cell. It validated the superior prediction of gene combination effect, but not SL prediction. Availability and implementation: The software is freely available at https://github.com/BirkanGokbag/TL-iGeneCombo under the GNU General Public License v.3.0.

Pediatric developmental safety assessment presents distinct challenges in drug development. Children undergo dynamic changes in organ systems and drug disposition pathways during development, and drug‐related effects may be delayed or manifested differently than in adults. Pharmacovigilance data, primarily from adults, is not always predictive of the response in children. Historically, juvenile animal studies have been used for nonclinical pediatric drug safety assessments exclusively, but the reduction in animal testing is a goal of public health agencies. To address this concern, the Office of Clinical Pharmacology at the U.S. FDA and the Triangle CERSI (Center of Excellence in Regulatory Science and Innovation) convened a workshop entitled: “Pediatric Developmental Safety Assessment and New Approach Methodologies (NAMs),” attended by experts from regulatory agencies, academia, and industry. Discussions focused on three overarching topics: (i) Secondary Pharmacology Screening Tools for Assessing Pediatric Developmental Safety; (ii) Non‐clinical Tools for Pediatric Developmental Safety Assessment; and (iii) Computational and Data‐Driven Tools for Pediatric Developmental Safety. These topics were examined within the framework of the S11 Nonclinical Safety Testing In Support of Development of Pediatric Pharmaceuticals and E11A Pediatric Extrapolation guidances, which promote structured, question‐based pediatric extrapolation in drug development and authorization. The workshop concluded that NAMs hold considerable promise to advance the assessment of pediatric drug developmental safety.

Objective: Pregnant and breastfeeding women, along with children, represent populations persistently underrepresented in clinical research, leading to uncertainty in clinical decision-making. This study investigated the landscape of pharmacology research concerning pharmacokinetics (PK), pharmaco-epidemiology (PE), and clinical trials (CT) within maternal and pediatric patient populations. Materials and Methods: A pharmacology research landscape in maternal and pediatric patient populations was analyzed using a novel co-learning framework applied to 38 million PubMed abstracts. It incorporated an evolving annotation guideline for abstract labeling, uncertain sampling, annotation error calibration, and large language model training. The co-learning framework enabled human annotators and language models to iteratively learn from each other, optimizing the classification of PK/PE/CT papers. Analysis was restricted to medications prescribed for maternal and pediatric patients, focusing on publication evidence. Results: Co-learning significantly increased the F1-scores for classifying the CT, PE, and PK abstracts from (0.81, 0.84, 0.73) to (0.96, 0.94, 0.93), respectively. On the external validation set, the F1-score increased from (0.81, 0.60, 0.42) to (0.95, 0.88, 0.91) for CT, PE, and PK, respectively. The models identified 376 231 and 879 936 abstracts relevant to maternal and pediatric patients, respectively. In pregnant or postpartum women, 22%–66% prescribed drugs lacked either PK or PE/CT evidence. Among pediatric patients, 8%–40% prescribed drugs had no or weak PK or PE/CT evidence. Discussion: Co-learning is a more effective approach than active learning in obtaining feedback from human annotators for optimizing the language model. With sufficiently labeled samples, a language model such as BioBERT outperformed large language models. Conclusion: In both maternal and pediatric pharmacology research, the absence of PK evidence was notably more frequent compared to PE and CT. This significant finding calls for more pharmacokinetics studies within pediatric and maternal patient populations.

Objective: Extract infant breast milk feeding status from electronic health record clinical notes to support observational drug-safety studies. Materials and Methods: Infant notes (19 489; 228 infants) were weakly labeled by large language model (LLM) and validated on 1045 manually reviewed notes. Using these labels, we built a rule-based medspaCy classifier from a curated feeding term set and compared it with ClinPhen and fine-tuned Bidirectional Encoder Representations from Transformers (BERT) document classifiers. Maternal postpartum notes (23 mothers; 616 partitions) were manually labeled to compare BERT with prompt-based language models. Results: LLM labels achieved 99.62% accuracy (macro F1 99.32%). medspaCy achieved 94.36% accuracy and 90.40% balanced accuracy (macro F1 0.90) and was applied to 52 278 mother–infant dyads. On maternal notes, the prompt-based model improved Positive-class F1 vs BERT (0.96 vs 0.82). Discussion: LLMs were most effective for scalable labeling; deterministic rules supported transparent large-scale inference. Conclusion: Weak supervision plus lightweight rules can convert sparse breastfeeding documentation into structured variables at-scale.

Objectives: To harmonize pregnancy medication records across heterogeneous real-world data sources and assess the comparability of recorded medication utilization. Materials and Methods: We analyzed pregnancy medication records from MarketScan administrative claims, the Indiana Network for Patient Care health information exchange, and the Ohio State University Medical Center electronic health record (EHR) system. A deterministic, rule-based pipeline resolved drug names to RxNorm ingredient concepts and mapped them to Anatomical Therapeutic Chemical (ATC) classes. We evaluated mapping coverage and compared utilization using shared ingredient counts, Jaccard similarity across rank thresholds, correlations, and therapeutic-class distributions. Results: Of 3094 ingredient components across sources, 3085 (99.7%) resolved to RxNorm concepts; 93.9% of resolved components carried an ATC code (93.4%-94.4% by source). The 3 sources shared 759 ingredients, with pairwise Jaccard similarities of 0.68-0.70 across complete lists. However, claims and EHR shared only 2 of their top 10 medications (Jaccard similarity, 0.11), with similarity increasing steadily as more medications were included in the ranked lists. Anatomical Therapeutic Chemical level-2 analysis identified 15 source-enriched classes. Medications identified in only one source had low recorded utilization. Discussion: High mapping coverage enabled comparison using standardized medication concepts, but substantial utilization differences remained. These patterns were consistent with differences in care settings and medication capture, although their individual contributions could not be isolated. Conclusion: This framework supports pregnancy medication research by enabling cross-source comparisons and identifying utilization differences requiring source-specific interpretation. Terminology standardization should be accompanied by consideration of how each source generates and captures medication records.

Motivation: Chemical toxicity assessment is critical for drug development and environmental safety. Computational models have emerged as a promising alternative to animal testing and now play a significant role in efficiently evaluating new chemicals. To address the urgent need for user-friendly machine learning tools in computational toxicology, we developed ToxiVerse, a public web-based platform. Results: ToxiVerse provides automatic chemical bioprofiling, curated toxicity datasets, and a predictive modeling interface designed for researchers who lack programming expertise. The platform comprises three integrated modules: (i) Bioprofiler, which provides chemical descriptors by combining chemical-bioactivity data from PubChem assays with a machine learning-based data gap-filling procedure; (ii) Database, which hosts ∼50 000 curated chemicals covering diverse toxicity endpoints; and (iii) Cheminformatics, which enables dataset upload, chemical curation, and automatic generation of quantitative structure–activity relationship models for toxicity prediction. Availability: The tool is accessible at www.toxiverse.com, and source code is available at https://github.com/zhu-research-group/toxiverse.

Background Long COVID affects a substantial proportion of the over 778 million individuals infected with SARS-CoV-2, yet predictive models remain limited in scope. While existing efforts, such as the National COVID Cohort Collaborative (N3C), have leveraged electronic health record (EHR) data for risk prediction and identification, accumulating evidence points to additional contributions from social, behavioral, and genetic factors. Methods Using a diverse cohort of SARS-CoV-2-infected individuals (n > 17,200) from the NIH All of Us Research Program, we investigated whether integrating EHR data with survey-based and genomic information improves model performance. Results Our multi-scale approach outperforms EHR-only model’s area under the receiver operating curve 0.736 (95% CI: 0.730, 0.741), achieving an area of 0.748 (0.741,0.755). Among the top predictors, active-duty service status, and self-reported fatigue are the most informative survey features. Conclusions These findings highlight the importance of incorporating multi-scale data to improve risk stratification and inform personalized interventions for long COVID. However the relative increase in accuracy is modest, and the cost of collecting genetic and survey data should be considered before implementation.

Background/Objectives: Regulated cell death (RCD), a process that relies on a series of molecular mechanisms, can be targeted to eliminate superfluous, irreversibly damaged, and potentially harmful cells. In this research, we want to better understand how the cell death pathway contributes to cancer therapy. Methods: We studied 1150 cancer cells in the Dependency Map (DepMap) database for 12 distinct cell death pathways and assessed their gene essentialities. Genes which are essential in 90% or more of cancer cell lines are called always essential, or partial essential if falling into (10%, 90%), or rare essential if they are essential in less than 10% of cancer cell lines. Results: Overall, among these 12 cell death pathways, 23, 47, and 549 genes were classified as always essential, partial essential, and rare essential, respectively. In two cell death pathways, Parthanatos, and Pyroptosis, all genes were rare essential. Among the other ten cell death pathways, Apoptosis, Autosis, Necroptosis, Efferocytosis, Ferroptosis, Mitotic cell death, Autophagy, Lysosome-dependent cell death, MPT-driven necrosis and Immunogenic, there are (10, 1, 13, 6, 3, 9, 11, 1, 1, 0) partial essential genes, and (2, 0, 3, 1, 1, 13, 4, 0, 0, 1) always essential genes. Conclusions: These cell death pathway essential genes could be viable targets for therapeutic drug development for cancer therapies.