Related Experiment Video
Updated: Jun 13, 2026
The Visual Colorimetric Detection of Multi-nucleotide Polymorphisms on a Pneumatic Droplet Manipulation Platform
Published on: September 27, 2016
UGP-RAP: A DI-QDa-MS powered deep learning platform for automated identification of ultrafine granular powders
Fei Huang1, Zhen-Wei Li2, Jia-Wei Wang2
1School of Chinese Materia Medica, Nanjing University of Chinese Medicine, Nanjing, 210023, China; Zhongshan Institute for Drug Discovery, Shanghai Institute of Materia Medica, Chinese Academy of Science, Zhongshan, Guangdong, 528400, China.
Background:
The chemical composition of herbal-derived ultrafine granular powders is highly complex, and conventional identification methods often struggle to achieve an effective balance between analytical throughput and identification reliability.
Purpose:
This study aimed to develop a rapid automated platform for the intelligent identification of ultrafine granular powders using direct infusion QDa mass spectrometry (DI-QDa-MS) and cosine similarity-enhanced deep learning.
Methods:
An automated identification platform, termed Ultrafine Granular Powder Rapid Automated Identification Platform (UGP-RAP), was developed by integrating one-dimensional convolutional neural network (1D-CNN), full-ion cosine similarity, and binary cosine similarity within a unified analytical framework. Samples were analyzed by DI-QDa-MS, with per-injection analysis time of approximately 2 min. For the binary cosine similarity approach, the top 200 and 250 most intense peaks in the positive and negative ion modes, respectively, were selected to optimize discriminative performance. A dataset of 530 batches was constructed and split into training, validation, and test sets for model development. External validation was conducted on 63 commercially sourced batches.
Results:
UGP-RAP achieved an overall identification accuracy of 98.41% in external validation, with a prediction confidence of 100%. Across all datasets, the 1D-CNN model consistently outperformed conventional KNN and SVM models, while both cosine similarity-based approaches demonstrated strong discriminative capability in positive and negative ion modes. Through multi-model integration and consensus-based decision strategies, UGP-RAP enabled a high-throughput, low-manual-intervention automated identification workflow with intuitive visualization of prediction results. Overall, the platform provides an efficient and reliable analytical solution for the rapid identification and quality control of ultrafine herbal powders, demonstrating considerable potential for routine analytical applications.