Table 1.
The Meridians and their example herbs.
Each Meridian is linked to a particular Organ which is characterized by its Elements and Quality of Yin or Yang. TCM considers a disease a result of loss of balance in the Yin and Yang, which can be restored using herbs that target particular Meridians.
Fig 1.
Herb-compound network shows the associations between herbs (green rectangles) and their active compounds (purple circles), which were used to determine the Herb-Feature and the Compound-Meridian matrices from the Herb-Meridian and Compound-Feature matrices. The features of herbs and compounds were determined from the chemical fingerprints and ADME properties. Machine learning methods were utilized to predict the Meridian classes for herbs and compounds respectively, by parameter optimization, model selection and feature selection.
Fig 2.
Herb-Meridian and Compound-Meridian distributions.
(A-B) The color bars at the bottom left represent the frequency of herbs or compounds for each of the seven major Meridians, which can be further collapsed into subclasses depending on whether an herb or a compound is shared by one or several Meridians. The vertical bars show the frequency of herbs or compounds for a particular subclass of Meridian combination, as indicated by the connected lines below the x-axis between the Meridians. (C-D) The Jaccard coefficients between the Meridian pairs at the herb and the compound levels. The size of blue circles on the upper diagonal shows the degree of the similarity.
Fig 3.
Evaluation of the machine learning model predictions.
(A) The overall Balanced accuracy for the seven Meridians. Dashed line indicates the level of 0.65. (B) The Balanced accuracy at the three data levels (compound-level, herb-level before and after ADME filtering). (C) The balanced accuracy for the four machine learning methods at the compound level. (D) The balanced accuracy for the ADME and fingerprint feature types at the compound level. Wilcox rank sum test. *: p < 0.05; **: p < 0.01; ***: p < 0.001; ****: p < 0.0001.
Table 2.
The balanced accuracy that was achieved for each Meridian at the compound level by Random Forest using all the available features.
Fig 4.
Important features determined at the compound-level prediction of Meridian.
(A) The distribution of importance scores for the top 59 features as compared to all features. (B-C) The bi-clustering of the importance scores for the 27 ADME features and 32 fingerprints.
Table 3.
The AUPRC ratio that was achieved for each Meridian at the compound level by Random Forest using ADME features only.