Methodological considerations for a deep learning‐based prescription framework for traditional Chinese medicine

We read with great interest the article by Mao and colleagues (Mao et al., 2025), who present the DA-TCMPO (Data Augmentation for TCM Prescription Optimisation) deep learning framework. We commend the authors for proposing a risk-contained, incremental prescription optimisation strategy that leverages attention-enhanced diffusion-based augmentation and noise-resilient latent representation learning. This low-risk refinement paradigm offers valuable methodological inspiration for controllable TCM prescription optimisation within intelligent clinical decision support systems (CDSS). However, several methodological boundaries deserve further discussion to fully interpret the clinical translation and applied reliability of this framework. First, the framework's ability to evaluate herb–herb interactions remains implicit and non-extractable. Although the Double Attention and Diffusion (DAD) module captures latent-space correlations among embeddings, these embedding-level associations cannot be equated with mechanistic interpretation of pharmacological synergy or antagonism. Critically, herb–herb interactions are foundational to the pharmacological identity of traditional Chinese medicine (TCM) multi-herb formulas, where formula interactions directly shape formula attributes and systemic therapeutic effects. In TCM prescription design, interaction assessment must adequately incorporate molecular, signalling and metabolism-related interaction features. The current framework does not yet support explicit extraction, interaction categorisation or exportable interpretation of true herb-compatibility intensity or underlying mechanisms, which underscores the unmet major requirement for traceable, explainable interaction modules in AI-driven TCM prescription systems (Varghese, 2020). Future frameworks should advance beyond latent embeddings by explicitly modelling herb–herb interactions through compound-level target overlap and pathway convergence-based interaction matrices, enabling mechanistically interpretable interaction modelling. Second, although the CH (Chinese Herbal Prescriptions for Diseases) dataset covers extensive sources, its prescription pool includes a substantial proportion of ancient and classical TCM formulas, the compositions of which reflect historical herb availability and resource constraints rather than modern clinical practicality. These classical formulas show notable limitations in current use, and their clinical value differs substantially across patient and disease contexts (Li et al., 2015). Therefore, datasets constructed from real-world clinical prescriptions may offer greater clinical representativeness than those primarily derived from ancient or classical TCM formulas. Separately, across all prescriptions collected in the CH database, clinical utility shows notable variability, with some formulations demonstrating broad applicability and others where therapeutic efficacy may not be consistently reliable. This variability should be understood as carrying unequal clinical contributions. The current model adopts uniform encoding of symptoms and herbs without incorporating evidence-informed weighting or explicit herb role constraints. For instance, TF-IDF (term frequency-inverse document frequency)-driven symptom-specific importance scoring has been shown to improve clinical relevance alignment and remedy prioritisation in TCM corpora (Wang et al., 2010), supporting the adoption of evidence-aware, non-uniform feature weighting as a feasible enhancement path for TCM-oriented CDSS model design (Salmi et al., 2024). Third, herb dose equivalence is essential when comparing optimised versus original TCM formulas, because increased dosage alone can elevate therapeutic effects generally. In this study, both the CYKKL-1 (6 g Zingiber officinale Ganjiang substituted with 15 g Codonopsis pilosula Dangshen and CYKKL-2 (6 g Ligusticum chuanxiong Chuanxiong substituted with 10 g Wolfiporia extensa Fuling) formulas raised the total herbal dosage relative to the original CYKKL (Changyankang granule) formula, making efficacy gains difficult to attribute to model-guided optimisation rather than dose escalation. Moreover, dextran sodium sulfate (DSS)-ulcerative colitis (UC) mice received 3-times the clinically equivalent human dose for 10 days; under such high-dose conditions, dosage increments exert a disproportionately amplified impact on systemic therapeutic readouts, magnifying the apparent separability between CYKKL-2 and CYKKL without providing proportionate insight into formula-specific optimisation value. Consequently, the apparent benefit of CYKKL-2 could stem from the increased dose effect, rather than a true indicator of model-driven prescription superiority. To address this issue, dose equivalence constraints or dose-normalised efficacy evaluation should be applied to distinguish genuine model-driven optimisation from dose-related effects under high-dose conditions. Fourth, organ–symptom alignment in the model's in vivo validation strategy requires closer attention. In the CH dataset, intestinal-related symptom features appear less frequently than signals linked to organs such as the heart or kidneys. The study selected a DSS-induced ulcerative colitis mouse model for biological validation. The DSS model may have been selected based on CYKKL's clinical use intent and treatment priorities. The intestinal phenotype is not strongly represented in CH's symptom-frequency profile, making dataset-to-phenotype alignment less direct. Consequently, the validation outcomes may reflect the model's performance on a less represented organ category. Additional biological validation in disease models corresponding to organ systems more prominently represented in the CH dataset (e.g. cardiac or renal systems) is required to substantiate the generalisability of the model outputs. We hope these perspectives will inform the ongoing refinement of model-guided TCM prescription design systems and enhance their clinical applicability. Mingze Ma: Conceptualization; writing—original draft; data curation. Yingjie Jia: Writing—review and editing; supervision. Ruiyu Mou: Supervision; writing—review and editing; methodology. The authors received no specific funding for this work. The authors declare that, to their knowledge, they have no financial, personal or professional relationships that could be construed as a potential competing interest regarding the work described in this manuscript.