Compound Danshen Formula (CDF) is a widely used Traditional Chinese Medicine (TCM) which has been extensively applied in clinical treatment of cardiovascular diseases (CVDs). However, the underlying mechanism of clinical administrating CDF on CVDs is not clear. In this study, the pharmacological effect of CDF on CVDs was analyzed at a systemic point of view. A systems-pharmacological model based on chemical, chemogenomics and pharmacological data is developed via network reconstruction approach. By using this model, we performed a high-throughput in silico screen and obtained a group of compounds from CDF which possess desirable pharmacodynamical and pharmacological characteristics. These compounds and the corresponding protein targets are further used to search against biological databases, such as the compound-target associations, compound-pathway connections and disease-target interactions for reconstructing the biologically meaningful networks for a TCM formula. This study not only made a contribution to a better understanding of the mechanisms of CDF, but also proposed a strategy to develop novel TCM candidates at a network pharmacology level.

P-glycoprotein (P-gp), a drug efflux pump, affects the bioavailability of therapeutic drugs and plays a potentially important role in clinical drug-drug interactions. Classification of candidate drugs as substrates or inhibitors of the carrier protein is of crucial importance in drug development. Accurate classification is difficult to achieve due to two major factors: i. The extreme diversity of substrates and the presence of multiple binding sites complicate the understanding of the mechanisms behind and hinder the development of a true, conclusive quantitative structure-activity relationship (QSAR) for P-gp substrates. ii. Both inhibitors and substrates interact with the same binding site of P-gp, as a result, it is not surprising that both share many common structural features. In this work, an unsupervised machine learning approach based on the Kohonen self-organizing maps (SOM) was explored, which incorporated a predefined set of physicochemical descriptors encoding the key molecular properties capable of discerning a substrate from an inhibitor. The SOM model can discriminate between substrates and inhibitors with an average accuracy of 82.3%. The current results show that the SOM-based method provides a potential in silico model for virtual screening.