Response to “From promising signatures to clinically credible prediction: Strengthening microbiome–metabolome prognostic research”

Response to "From promising signatures to clinically credible prediction: Strengthening microbiome-metabolome prognostic research" Dear Editor,We thank Yang and Cheng for their thoughtful and constructive Letter to the Editor regarding our review, "Microbiome-metabolome multi-omics biomarkers for infectious disease prognosis: Current evidence, AI-driven integration, and translational challenges" [1].We appreciate their careful consideration of the methodological and translational challenges of moving microbiome-metabolome research from promising biological associations to clinically useful prognostic applications.The authors appropriately emphasize the importance of transparency in identifying and grading evidence.Our article was designed as a critical narrative review, rather than a systematic review or formal evidence synthesis.The qualitative categories used to characterize the maturity of biomarkers and predictive approaches were intended to provide an overview of the field's current state rather than constitute a formally validated evidence-grading system.Nevertheless, we agree that providing more explicit operational definitions for these categories would further improve transparency and reproducibility.Future evidence syntheses should incorporate clearly prespecified search and selection strategies and, where appropriate, structured frameworks for assessing evidence quality.We also agree with the distinction between the feasibility of applying machine-learning methods to high-dimensional omics data and establishing a clinically validated prognostic model.Importantly, we did not intend to equate applying machine-learning algorithms to omics datasets with clinically validated prediction models.Rather, we discussed these approaches as emerging strategies whose clinical utility remains to be established through rigorous validation.As emphasized in our review, the current literature remains constrained by relatively small cohorts, heterogeneous study designs, limited external validation, reproducibility concerns, and challenges in model interpretability.Future studies should therefore provide comprehensive information on cohort size, outcome events, feature-selection procedures, resampling and validation strategies, discrimination, calibration, and incremental value beyond established clinical predictors.Contemporary frameworks such as TRIPOD + AI and PROBAST + AI provide useful guidance for transparent reporting and assessment of prediction models developed using regression or artificial intelligence methods [2,3].Temporality and treatment-related confounding are equally important.Microbiome and metabolomic profiles can change substantially in response to antimicrobial therapy, corticosteroids, nutritional interventions, intensive-care exposure, and disease progression.Consequently, without a clearly defined prognostic time zero and appropriate temporal sampling, an apparent prognostic signature may reflect treatment effects or evolving disease severity rather than baseline risk.