Pengyi Yang

Health and life scientists frequently rely on quantitative experts to perform complex data analyses, yet interpretation of these results often becomes a bottleneck due to communication barriers across disciplines. Large Language Models (LLMs) have the potential to act as intermediaries that translate analytical outputs into accessible scientific narratives, but their ability to reliably interpret outputs from real-world biomedical data analyses across disciplines remains unclear. In this study, we investigate how state-of-the-art LLMs behave when integrated into real bioinformatics analysis workflows as report-generation assistants for interdisciplinary teams. We evaluated LLM-generated summaries using both automated and human evaluation frameworks to ensure holistic evaluation. Automated assessment employed multiple-choice questions designed using Bloom’s taxonomy to assess multiple levels of understanding, while human evaluation tasked scientists to score summaries for factual consistency, lack of harmfulness, comprehensiveness, and coherence. All generally produced readable and largely safe summaries, confirming their value for first-pass translation of technical analyses; however frequently misinterpreted visualisations, produced verbose summaries and rarely offered novel insights beyond what was already contained in the analytics. Our findings suggest that LLMs are best suited for easing interdisciplinary communication rather than replacing domain expertise and human oversight remains essential to guarantee accuracy, interpretative depth, and the generation of genuinely novel scientific insights.

Single-cell transcriptomics technology offers unprecedented insights into molecular heterogeneity. However, capturing sample-level representations that reflect both systemic and cellular states remains challenging, especially when disease annotations are mostly available as coarse sample-level labels. Here, we introduce Phenoverse, an interpretable deep learning framework that learns sample-level disease state representations through cell type-aware residual encoding, prototype learning, and Perceiver-based aggregation. Applied to independent single-cell transcriptomic cohorts of COVID-19, Alzheimer’s disease, and systemic lupus erythematosus, totaling over 5 million cells, we demonstrate that learned sample representations enable disease state prediction and encode a continuous spectrum of disease severity on unseen data that correlate with multiple clinical and pathological measures, despite being trained solely on binary phenotype labels. Further, we demonstrate that trajectory-derived genes reveal cross-cohort molecular programs and show consistently higher reproducibility than traditional case-control comparisons. Finally, prototype learning provides intrinsic model interpretability and enables the characterization of cell type-specific disease states. Taken together, Phenoverse offers an interpretable disease-phenotyping approach to dissecting sample heterogeneity, and our results highlight its utility in translating complex single-cell transcriptomic data into patient-level biological insights.

Retinoic acid (RA) is a key morphogen in human retinal development, activating transcriptional programs that drive retinal progenitor differentiation and photoreceptor development, ensuring proper spatial organisation within the neural retina, essential for vision. Despite its well-established role in retinal patterning, the concentration-dependent effects of RA on human retinal cell fate specification and the regional definition of the primate macula and peripheral retina remain poorly understood. Here, we show that temporal and dosage-dependant modulation of RA during human retinal organoid differentiation induces distinct changes in retinal cell abundance, maturation, and organisation. Single-cell transcriptomics and protein analysis reveal that RA dosage influences the relative abundance and maturation of photoreceptors and retinal interneurons. Spatial transcriptomics analyses demonstrates that low RA levels biases retinal organoids toward a macular-like regional identity, whereas high RA levels promotes peripheral-like development. Collectively, our findings emphasise the critical role of RA signalling in retinal maturation and regional specification. This study elucidates mechanisms involved in human retinal development and we anticipate that controlled RA modulation in retinal organoids provides a strategy to refine disease modelling of inherited retinal disorders and enhance the specific generation of photoreceptors suitable for transplantation and regenerative therapies.

Single-cell omics enable the dissection of cellular heterogeneity, yet the high dimensionality, inherent noise, and sparsity present significant challenges. These challenges are amplified for rare cell populations, which are often difficult to annotate reliably but can be central to development and disease. As single-cell assays increasingly capture multiple molecular layers, the integrative analysis of such multimodal data further increases complexity. Here, we propose Hydra, a deep generative framework based on an ensemble of variational autoencoders for effective learning of unimodal and multimodal single-cell omics data. Hydra implements interpretable modules for capturing cell-type-specific molecular signatures. The ensemble of such interpretable modules enables reproducible feature selection and robust cell-type annotation, with particular effectiveness for rare populations. We benchmarked Hydra on a repertoire of 21 datasets, including unimodal and multimodal single-cell omics data. Our results demonstrate that Hydra offers comparable to superior performance to several state-of-the-art methods. Finally, we highlight the utility of Hydra in robustly annotating brain cellular subtypes and preserving disease-relevant signatures using our previously published dataset that profiles Alzheimer’s disease.

Insulin action on the skeleton is essential for bone development and whole-body energy metabolism, however a global view of signaling in this tissue is lacking. Furthermore, whether there are signaling differences that drive the gene-specific activation under insulin-resistant (IR) or ageing conditions is unknown. Here, we perform a phosphoproteomic analysis of insulin signaling in the bones of young, lean, insulin-sensitive versus old, obese, IR mice revealing a rewiring of phosphorylation. We target dysregulated phosphoproteins in a zebrafish functional genomic screen of bone development and mineralization revealing candidates important for skeletal formation. One of these is ALF Transcription Elongation Factor 4 (AFF4), the core scaffold of the Super Elongation Complex and we show that phosphorylation of S831 on AFF4 is an insulin-dependent substrate of P70S6K and attenuated in aged, IR bone. Phosphorylation of S831 is defective in IR osteoblasts and associated with reduced transcriptional elongation at discrete locations in the genome. Mechanistically, we show phosphorylation of S831 increases recruitment of chromatin remodelers, ENL/AF9 to crotonylated histone via the YEATS domain, and promotes gene-specific activation. Our analysis identifies regulators of insulin action on the skeleton, further uncovering a mechanism of IR via locus-specific changes in transcriptional elongation and gene activation.

Chemotherapy often kills a large fraction of cancer cells but leaves behind a small population of drug-tolerant persister cells. These persister cells survive drug treatments through reversible, non-genetic mechanisms and cause tumour recurrence upon cessation of therapy. Here, we report a drug tolerance mechanism regulated by the germ-cell-specific H3K4 methyltransferase PRDM9. Through histone proteomic, transcriptomic, lipidomic, and ChIP-sequencing studies combined with CRISPR knockout and phenotypic drug screen, we identify that chemotherapy-induced PRDM9 upregulation promotes metabolic rewiring in glioblastoma stem cells, leading to chemotherapy tolerance. Mechanistically, PRDM9-dependent H3K4me3 at cholesterol biosynthesis genes enhances cholesterol biosynthesis, which persister cells rely on to maintain homeostasis under chemotherapy-induced oxidative stress and lipid peroxidation. PRDM9 inhibition, combined with chemotherapy, results in strong anti-cancer efficacy in preclinical glioblastoma models, significantly enhancing the magnitude and duration of the antitumor response by eliminating persisters. These findings demonstrate a role of PRDM9 in promoting metabolic reprogramming that enables the survival of drug-tolerant persister cells.

Health and life scientists routinely collaborate with quantitative scientists for data analysis and interpretation, yet miscommunication often obscures the interpretation of complex results. Large Language Models (LLMs) offer a promising way to bridge this gap, but their cross-discipline interpretative skill remains limited on real-word bioinformatics analyses. We therefore benchmarked four state-of-the-art LLMs: GPT-4o, o1, Claude 3.7 Sonnet, and Gemini 2.0 Flash, using automated and human evaluation frameworks to ensure holistic evaluation. Automated assessment employed multiple choice questions designed using Bloom’s taxonomy to assess multiple levels of understanding, while human evaluation tasked scientists to score summaries for factual consistency, lack of harmfulness, comprehensiveness, and coherence. All generally produced readable and largely safe summaries, confirming their value for first-pass translation of technical analyses, however frequently misinterpreted visualisations, produced verbose summaries and rarely offered novel insights beyond what was already contained in the analytics. Our findings suggest that LLMs are best suited for easing interdisciplinary communication rather than replacing domain expertise and human oversight remains essential to guarantee accuracy, interpretative depth, and the generation of genuinely novel scientific insights.