Background: Text feature ranking is an essential step in biomedical data analysis which directly affects the text prediction models. Traditional feature extraction methods such as mutual information, probabilistic information gain, statistical chi-square and inverse document frequency are used to find the essential key terms and its relationships in document sets of limited size. Semantic relationships plays a vital role in large document sets is a challenging task for text mining applications. In the biomedical applications, predicting the gene-disease based chemical drug is difficult due to large number of training datasets and sparsity issue. In order to overcome these issues, a multi-layered graph based deep learning framework is designed and implemented on the large biomedical gene-disease and chemical drug data. In this framework, an extended version of particle swarm optimization and Multiclass Support Vector Machine (IPSO-MSVM) model are developed to extract the key terms for document analysis. Results: Experimental results are tested on different biomedical document sets using the proposed gene based chemical drug database. Results show that the present framework has better true positivity and accuracy than the stat of art algorithms . Conclusion: Here, a non-linear support vector machine is used in CNN framework to classify the gene disease patterns using the medical datasets. After the analysis of data the gene patterns are identified, based on this the drugs are discovered. So it will more useful for the clinical experts for decision making about the patients and helps to predict the diseases they are prone to and for discovering the drugs based on their genes.