Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 8 appraisals
Journal of chemical information and modeling
Functionally Guided Graph Learning for Robust Cross-Patient Cell-Type Annotation in Single-Cell RNA Sequencing
Cross-patient cell-type annotation in single-cell RNA sequencing (scRNA-seq) remains challenging due to pronounced interpatient heterogeneity and distribution shifts across patient-specific cellular contexts. Conventional annotation approaches often rely on proximity-driven graph construction or expression similarity, which may introduce spurious cell-cell connections and lead to unstable knowledge transfer across patients. To address this limitation, we propose PathoGraph, a functionally guided graph learning framework for robust cross-patient cell-type annotation. The proposed method integrates KEGG-7-based biosemantic graph structure learning with cross-patient representation adaptation. Specifically, pathway-derived functional semantic profiles are incorporated to refine patient-specific cell graphs, encouraging biologically coherent neighborhoods and suppressing noise introduced by purely expression-based similarity. Based on the refined graphs, a cross-patient representation adaptation mechanism further aligns embeddings between labeled reference patients and unlabeled query patients to facilitate reliable annotation transfer. Experiments on three cross-patient scRNA-seq data sets, including leukemia, breast invasive carcinoma, and colorectal cancer data sets, demonstrate that PathoGraph achieves stable annotation performance across 32 directed reference-to-query transfer tasks. Across all tasks, PathoGraph obtained an average ACC of 84.28% and an F1-score of 84.08%, showing competitive and stable performance compared with representative marker-based, correlation-based, and model-based annotation methods. Ablation studies further show that removing the biosemantic graph learning module reduces the average accuracy to 83.48%, highlighting the importance of functional-guided graph refinement. In addition, post hoc functional relevance analyses in immune-cell and cancer-associated contexts suggest that the learned cell-cell graphs capture biologically relevant neighborhood structures beyond expression-driven proximity. The source code and processed data are publicly available at: https://github.com/LiYuechao1998/PathoGraph.
29 July 2026
Read appraisal →Journal of chemical information and modeling
Unveiling Large-Scale Kinase-Centric Protein-Protein Interactions through a Knowledge-Informed Workflow
Protein phosphorylation regulates signaling, yet atomic-level substrate specificity remains elusive due to sparse structural data and phosphorylation-site-insensitive deep-learning predictors. Here we present a pipeline reformulating kinase-substrate modeling as a Bayesian inference problem. By integrating curated data sets and literature evidence parsed by Large Language Models, we converted diverse biological knowledge into structural restraints for the restraint-guided deep-learning model GRASP. For EGFR, BRAF and JNK1, we obtained 336 new phosphorylation-site-specific structure candidates refined by molecular dynamics. These models recapitulate known features, such as JNK1's hydrophobic docking groove, and enabled a Virtual Position Scanning Peptide Array (V-PSPA) to map recognition patches and derive sequence preferences. Cross-referencing predicted interfaces with AlphaMissense pathogenicity scores reveal that the interaction types and distances to the catalytic pocket significantly influence pathogenicity scores. A comparison with clinical mutation data sets further connects pathogenic mutations to the kinase-substrate interface. This high-resolution, high-throughput pipeline can be broadly applicable to kinase specificity studies and general drug discovery.
28 July 2026
Read appraisal →Molecular genetics and genomics : MGG
The use of artificial intelligence in advancing molecular biology in Africa: a narrative review
Artificial intelligence (AI) is rapidly becoming a core methodological pillar of molecular biology and precision medicine, and Africa is a uniquely consequential setting for this transition because the continent combines the world's greatest human genomic diversity with the most severe underrepresentation of that diversity in the datasets and reference resources on which AI models are built and benchmarked. This narrative review examines, for a genetics and genomics readership, where AI-driven methods are already strengthening African molecular biology, where the supporting evidence remains preliminary, and what is required to translate technical capability into scientifically robust and equitable benefit. The central argument is that AI is especially consequential in African molecular biology, not simply because it automates analysis, but because it can help unlock insight from African genomic diversity, pathogen biology, and clinically relevant multi-omics data that remain underrepresented in global models. Across core molecular domains, AI is accelerating protein structure prediction, high-throughput variant calling and pan-genomic reference construction, genome-wide association analysis, transcriptomic interpretation, drug discovery, and CRISPR guide design. African initiatives such as H3Africa, the African Genome Variation Project, H3ABioNet, and the H3D Centre show that locally generated datasets and African-led computational pipelines can already support meaningful discovery, from improved variant interpretation to structure-guided therapeutic prioritization. At the same time, persistent barriers remain, including underrepresentation of African genomes in training data and reference genomes, uneven computational infrastructure, limited interdisciplinary training, fragmented governance, and the risk that AI-derived benefits will remain inaccessible to the populations whose data enable them. We conclude that the future impact of AI in African molecular biology will depend less on adopting global tools in the abstract and more on building African-led datasets, validation pipelines, governance frameworks, and translational pathways that make molecular discovery both scientifically robust and equitably useful. Looking ahead, the central perspective offered by this review is that Africa's exceptional genomic diversity should be treated as a scientific asset rather than an analytical liability: realising this will require population-representative pan-genome references, sustained computational capacity, and governance structures that ensure African populations are not only the source of the underlying data but also the principal beneficiaries of the discoveries it enables.
23 July 2026
Read appraisal →Cell systems
BulkFormer: A large-scale foundation model for bulk transcriptomes
Foundation models are transforming transcriptome analysis, yet most RNA sequencing (RNA-seq) foundation models are pretrained on sparse single-cell RNA-seq data, which typically captures only ∼3,000 genes per cell. This creates a need for models tailored to bulk transcriptomes, a modality that profiles ∼16,000 genes per sample and supports clinical and tissue-level analyses. Here, we present BulkFormer, a foundation model for bulk transcriptome analysis. BulkFormer contains ∼150 million parameters, covers 20,010 protein-coding genes, and is pretrained on 581,503 human bulk RNA-seq profiles. Its hybrid encoder combines a graph neural network to model explicit gene-gene relationships with a Performer module to capture global expression dependencies. Across five downstream tasks, BulkFormer outperformed existing single-cell foundation models while requiring substantially lower training cost. These results identify pretraining data modality as a key determinant of foundation model performance and establish BulkFormer as a framework for bulk transcriptome modeling. A record of this paper's transparent peer review process is included in the supplemental information.
17 July 2026
Read appraisal →Nucleic acids research
SNPnexus: an enhanced web platform for large-scale and multi-sample variant analysis (2025 update)
SNPnexus is a long-standing web-based platform for the functional annotation and prioritization of genetic variants. Since its previous release, SNPnexus has undergone substantial backend re-engineering resulting in major performance improvements and a complete restructuring of the underlying datasets. A redesigned interface now enables larger queries and multi-sample analysis enabling comparative workflows such as identifying shared pathogenic variants in disease cohorts and divergent mutations in cancer evolution studies. SNPnexus increased its capacity to 150 000 variants per query and introduced pre-annotation filters that allow users to restrict analyses to selected genes or genomic regions, improving efficiency while enabling targeted interrogation of high-priority targets. SNPnexus integrates updated and expanded annotations across both GRCh37 and GRCh38, covering genomic consequences, in silico pathogenicity predictions, population allele frequencies, evolutionary conservation, regulatory elements, biological pathways, and clinical associations. The refreshed result interface provides interactive visualizations for single-sample and cohort-level outputs, together with advanced filtering and export options. Optional user accounts now support query history and real-time job monitoring while preserving full, unregistered access for all users. SNPnexus remains free and open to all users without login requirements at https://snpnexus.org/.
13 July 2026
Read appraisal →Briefings in bioinformatics
Bridging local-global transmembrane protein contexts with contrastive pretraining for alignment-free pathogenicity prediction.
Predicting the pathogenic consequences of protein mutations is a cornerstone of precision medicine, yet it remains a formidable challenge for transmembrane proteins (TMPs), a clinically vital class of drug targets. Existing computational methods are often hampered by their reliance on evolutionary data and fail to model TMP-specific biophysical constraints. Here, we introduce Memo-Patho, a deep learning framework for robust, alignment-free pathogenicity prediction of TMP variants. The core innovation is a within-protein, label-informed supervised contrastive pretraining strategy that learns sequence-encoded biophysical signatures distinguishing pathogenic and benign variants by directly comparing them within the same protein context. By fusing sequence-level representations from protein language models with local structural proxies derived from sequence, Memo-Patho achieves accurate predictions without multiple sequence alignments or experimental structures. Across diverse TMP benchmarks and under protein-level group splits, Memo-Patho consistently outperforms leading predictors, achieving up to 0.93 accuracy, and it transfers to an independent KCNQ1 ion-channel cohort without re-training. Its resource-efficient, alignment-free design enables routine large-scale screening when evolutionary or structural data are sparse. Conceptually, Memo-Patho addresses a key gap by directly learning discriminative, sequence-anchored signatures pertinent to TMP-specific constraints, offering a principled and generalizable foundation for research-use clinical variant triage and proteome-wide mutation-effect modeling.
6 July 2026
Read appraisal →American journal of human genetics
A transparent and generalizable deep-learning framework for genomic ancestry prediction
Accurately characterizing genetic ancestry is critical for ensuring reproducibility and fairness in genomic studies and downstream health research. This study aims to address the prediction of ancestry from genetic data using deep learning, with a focus on generalizability across datasets with diverse populations and on explainability to improve model transparency. We adapt the Diet Network, a deep-learning architecture proven to be effective in handling high-dimensional data, to learn population ancestry from single-nucleotide polymorphism (SNP) data using the populational Thousand Genomes Project dataset. Our results highlight the model's ability to generalize to diverse populations in the CARTaGENE, Montreal Heart Institute, and All of Us biobanks and that predictions remain robust to high levels of missing SNPs. We show that, despite the lack of North African populations in the training dataset, the model learns latent representations that reflect meaningful population structure for North African individuals in the biobanks. To improve model transparency, we apply Saliency Maps, DeepLift, GradientShap, and Integrated Gradients attribution techniques and evaluate their performance in identifying SNPs leveraged by the model. Using DeepLift, we show that the model's predictions are driven by population-specific signals consistent with those identified by traditional population-genetics metrics. This work presents a generalizable and interpretable deep-learning framework for genetic-ancestry inference in large-scale biobanks with genetic data. By enabling more widespread genomic ancestry characterization in these cohorts, this study contributes practical tools for integrating genetic data into downstream biomedical applications, supporting more inclusive and equitable healthcare solutions.
4 July 2026
Read appraisal →Molecular cell
Rapid customization of base editors via machine learning-powered combinatorial mutagenesis
Being able to control the complementarity and hindrance between target DNA and base editor proteins enables precise, bystander-free editing. Here, we combined combinatorial mutagenesis with machine learning to analyze and engineer these interactions at scale. By profiling DNA motif preferences across 160,000 evoAPOBEC1 and 64 million TadA variants in human cells, we used as little as 0.004% of the mutational landscape to make predictions. This identified variants with motif-specific activity and eliminated residual adenine editing in cytosine base editors. In correcting >800 disease-associated mutations, our variants outperformed previous versions in precluding unintended edits at purine motifs, achieving undetectable bystander edits in 50% of cases. Additionally, a pre-trained, structure-based deep learning model predicted functional TadA variants with 63% success across 2026 variants spanning 26 amino acid sites, without experimental data and in a single prediction round. These approaches streamline the re-engineering of base editors for enhanced precision tailored to specific targets.
23 May 2026
Read appraisal →