Dear Editor, We read with great interest the recent publication by Dai et al entitled “A Multi-Omics Pipeline Integrating Machine Learning and Spatial-Cellular Analysis Identifies SASH1 as a Prognostic Biomarker and Therapeutic Target in Head and Neck Squamous Cell Carcinoma”1. The study presents a sophisticated integration of multi-omics data, machine learning algorithms, and spatial transcriptomic profiling to identify potential prognostic biomarkers in head and neck squamous cell carcinoma (HNSCC). Its comprehensive analytical framework – spanning computational discovery to cellular and protein-level validation – demonstrates a strong contribution to the advancement of precision oncology. While the study provides valuable insights, several methodological and interpretative aspects warrant further discussion and clarification. First, the study utilized four machine learning algorithms – LASSO, SVM-RFE, XGBoost, and Boruta – to identify key predictive genes. While such an ensemble strategy may enhance model robustness, this advantage could also introduce algorithm-selection bias. Given the wide range of available approaches, including Elastic-Net regression2, Random Forests, Gradient Boosting Machines3, and deep neural networks4, the rationale for selecting only these four algorithms remains insufficiently justified. Each algorithm is based on distinct assumptions regarding linearity, feature interdependence, and noise tolerance. Without a systematic rationale or stability analysis across alternative models, there is a potential risk that SASH1 was highlighted as a consequence of this specific algorithmic configuration rather than reflecting its intrinsic biological significance. Second, the study provides limited comparative evaluation of model performance and interpretability. Specifically, presenting metrics such as classification accuracy, area under the curve values, or the consistency of feature importance across models would have enhanced methodological transparency and reproducibility. Furthermore, model interpretability – particularly the biological reasoning underlying algorithm-derived feature rankings – was not sufficiently addressed. Incorporating explainable AI techniques, such as SHAP analysis5, could have elucidated how SASH1 contributes to the model’s decision-making process and whether its predictive relevance extends beyond mere statistical association. Third, although the inclusion of single-cell and spatial transcriptomic analyses is technically valuable, several important limitations should be acknowledged. The spatial data were generated using the 10× Genomics Visium platform, which features a spot diameter of approximately 55 μm6, thereby capturing transcripts from multiple adjacent cells. This constraint inherently limits single-cell resolution and may introduce signal contamination among tumor, stromal, and immune compartments. As a result, the spatial maps exhibit extensive overlap and color blending of annotated cell types – particularly within the tumor core – making the deconvolution results difficult to interpret. Moreover, the excessive overlay of text labels further compromises figure readability. Reorganizing these plots – for instance, by grouping related cell types, simplifying legends, and overlaying gene-expression heatmaps delineated by clear tissue boundaries – could substantially improve interpretability. Fourth, at the single-cell level, SASH1 expression was predominantly detected in nonmalignant fibroblast and endothelial populations rather than in malignant epithelial clusters. This pattern suggests that if SASH1 is not tumor-cell-specific, its prognostic significance may derive from stromal regulation rather than intrinsic tumor suppression. Clarifying whether SASH1 functions primarily within cancer cells or in the surrounding microenvironment is critical to establishing its translational relevance. Finally, in terms of experimental validation, the verification was limited to Western blot analysis of SASH1 expression. Although the observed protein-level downregulation supports the computational findings, no functional assays were conducted to demonstrate a causal link between SASH1 depletion and malignant behaviors, including proliferation, migration, and epithelial–mesenchymal transition. In summary, this study represents a meaningful step toward biomarker discovery in HNSCC. Moving forward, clarifying the algorithmic rationale, enhancing model interpretability, improving the visualization of spatial data, as well as experimentally validating the functional mechanisms, could further enhance the methodological rigor and translational relevance of the findings.
Xu et al. (Mon,) studied this question.