Method Research For The Prediction Of Small Biological Molecular-pathway Interactions

Drug discovery is a long and complicated process. In addition to find out relations between biological molecules and protein targets, more important thing is to identify the correct metabolic or conduction pathways for these molecules to play the biggest role. In other words, we need to minning associations between biological molecules and pathways. However, traditional experimental methods for the identification of molecular-pathway interactions were limited by experiment conditions, cost, et al, and machine learning methods can just make up the deficiencies in these aspects. In this article, we selected biological molecules and pathways as research objects, by collecting all kinds of biological knowledge, predicted the interactions between small molecules and pathways. What’s more, predicting the association between molecules and pathways can also provides valuable information for the following clinical experiments. The main contents are as follows:1. This section predicted the drugs-pathways interactions from KEGG database via multiple feature fusion, which all the pathways and drugs were associated with cancer disease. In detail, we integrated drug chemical structure information, drugs’ functional groups information into drug feature profiles; integrated mean value and variance value of gene expression data based on NCI-60, gene’s GO information into pathway feature profiles; in additional, topology information of drug-pathway interaction network were added into the feature profile. Three semi-supervised methods were selected to predict the drug-pathway interactions. Results shows that the performance was improved by the feature fusion, several potential drug-pathway interactions were validated in authoritative database.2. This section predicted the interactions between compounds and pathways based on improved Rotation Forest (RF) algorithm. We broadened research scope and screened 78 pathway based on cMap gene expression data, and screened 147 compounds from cMap’s 1309 molecules, then built the compound-pathway dataset. The feature profiles were also obtained specifically based on the cMap database. Then we suggested a improved ensemble learning method based on RF algorithm, in which relief method was used for feature projection and graph-based semi-supervised method was used for basic classify, and the new method was called RGRF algorithm for short. Results shows that RGRF method was performed better than original Rotation Forest method, and several predicted compound-pathway interactions can be find in recent literature.3. This section implements a visual prediction-tool based on RGRF algorithm. This tool can provide all cMap’scompounds’basic information in KEGG database, and predict the associations between all the cMap’s compounds and pathways, then give out the list of related pathways according to their predicted probability scores.