Domain ontology core entities for spectroscopic profiling data.

用于表征和集成多种谱图数据的本体模型

---

---

The domain ontology defines the core entities and their associations. 
A.	The “DataSet” is a collection of multiple “Spectrum” instances. The spectrums in one dataset are generated for the same purpose (e.g. classify milk brands or identify a specific genuine geo herb), by the same measurement modality (Raman or MALDI-TOF-MS), with the same data preprocessing methods (i.e. filtering, averaging, peak identification, baseline drift removal, etc.), and must have the same data dimensions (i.e. peak number).
B.	The “Spectrum” is a basic unit to represent a specific spectroscopic data. It is usually a final processed spectrum data (e.g. averaged and filtered from multiple scans) that can be used directly for successive data analysis, other than the original raw data.
A spectrum object contains an array of X values (e.g. wave number for Raman, or m/z for MALDI-TOF-MS), and an optional y label (used for supervised data analysis). 
C.	A “DataSet” can be exported as a matrix or tabular form, which is directly importable by major scientific data analysis platforms, such as MATLAB, R, or Python. In the following manuscript, we will show how to use this intermediate data format to drive the data analysis workflow.
D.	Each “Spectrum” instance has multiple “Log” items, which track the status change in its life cycle. The ontology defines several phases for the spectrum data life cycle, including generate, preprocess, curate, analyze and report. 
E.	Each “Spectrum” instance can be serialized to an mzML (for MS) or JCAMP-DX (for vibrational spectroscopy) file or deserialized from an external one. For third-party instruments and systems (e.g. Agilent, Bruker, Horiba, Shimadzu, Thermo, Waters, etc.), such standard file formats can be used to exchange and share spectrum data. Our team is also developing our own MALDI-TOF-MS hardware that will directly transmit data in the designed format.
F.	The “Pipeline” is a set of algorithm elements organized to achieve complex data analysis tasks. A typical pipeline for spectroscopic data contains several preprocessors (e.g. filter, normalization, dimension reduction) and one regressor or classifier.
G.	Each “Algorithm” instance represents a specific algorithm component used in the pipelines. The algorithms belong to several categories, such as baseline drift removal, averaging filter, feature scaling, feature selection, classifier, regressor, etc.
H.	Each algorithm can have multiple implementations from different scientific platforms and programming languages. The implementation can be machine-interpretable scripts, source codes or compiled binary files. Engineers can either call existing libraries to implement algorithms or upload their own implementations.
I.	Each “Pipeline” object is targeted to a specific data set and analysis purpose. The final state of a pipeline instance is usually a statistical model (e.g. logistic regression, SVM or neural network) trained by the related data set. The model with its various parameters is persisted to a model file (e.g. .mat file for MATLAB or pickle file for python), and can be reloaded back to runtime.
J.	A trained pipeline can predict/analyze a new sample data of the same topic, and generates both a human-readable report and a computer-processable structured report for further decision support uses.