Background: Antimicrobial resistance among Gram-positive veterinary pathogens has increased the need for more rational and experimentally grounded strategies to prioritize tylosin derivatives for antibacterial development. This study aimed to develop and externally a calibrated, decision-oriented screening workflow for tylosin-derived antibacterial analog prioritization against Staphylococcus aureus and Streptococcus agalactiae. Methods: Publicly available minimum inhibitory concentration (MIC) data were used to construct organism-specific multilayer perceptron models. Model performance was evaluated using a five-fold out-of-fold framework. Calibrated activity probabilities were then combined with precision-oriented thresholds and similarity-based applicability-domain constraints to define prospective Go/No-Go decisions. To examine external transferability, six synthesized tylosin derivatives (A1–A6) were prospectively predicted and subsequently tested using in vitro MIC assays. Results: Internal validation showed stronger and more stable performance for S. aureus than for S. agalactiae. However, closed-loop external validation revealed distinct organism-specific decision behaviors. For S. aureus, the workflow assigned Go decisions to five compounds and included the only experimentally active analog, A6, but also generated several false Go decisions for inactive analogs. For S. agalactiae, all six compounds were classified as No-Go under the primary decision rule, whereas MIC testing showed that four analogs were experimentally active, indicating conservative under-selection in a low-data extrapolation setting. Conclusions: Calibrated probabilities and applicability-domain analysis can improve the transparency and diagnostic value of antibacterial prioritization, but they do not guarantee robust external activity prediction when training-set coverage, threshold transferability, or chemical-space representation is limited. Overall, this study provides an early-stage, risk-aware closed-loop decision-support framework for tylosin-derived analog prioritization and highlights the need for iterative model updating, probability recalibration, and larger external validation before such workflows are used as high-confidence prospective screening tools in veterinary antibacterial discovery.

The purpose of this study was to use data mining and network pharmacology to determine drug patterns for bovine viral diarrhea (BVD). The frequency, properties, tastes, meridian tropism, and functions of prescription data were gathered and examined from four literature databases (2004–2024). Cluster analysis was used to find prescription patterns, and the Apriori algorithm (SPSS Modeler 18) was used to find associations. A total of 391 literature-derived prescription records were included in the analysis, involving 189 distinct herbal medicines and a cumulative herb-use frequency of 2031 occurrences. These herbs primarily enter the liver meridian and were categorized as cold, warm, or neutral. The predominant tastes were bitter, pungent, and sweet. Five frequently recorded herbs and 14 co-occurrence patterns among herbs were extracted. According to traditional Chinese medicine (TCM), spleen–stomach deficiencies and damp-heat pathogens are linked to BVD. These medication patterns were mainly associated with heat-clearing, detoxification, spleen-strengthening, and Qi-regulating strategies in TCVM theory. Targets were screened, PPI networks were constructed, and enrichment studies for core herbs (Baitouweng, Huangbo, Huangqin, Qinpi, and Zhizi) were performed using network pharmacology. The binding affinities between disease targets and active components were further assessed using molecular docking. The findings provide a descriptive summary of medication patterns and generate preliminary hypotheses regarding potential compound–target–pathway associations involved in the symptomatic and supportive use of TCM for BVD.