In this thesis, we address the black-box nature of deep learning models for molecular toxicity prediction as well as propose methods for aggregating various chemical features to have an improved accuracy. An ideal toxicity prediction model is characterized with high accuracy, capable of handling descriptors/features diversity, ease of training and interpretability. Considering these attributes of an ideal model, in the first quarter of this thesis we present a novel hybrid framework based on decision trees (DT) and shallow neural networks (SNN). This method paves a path to feature interpretability while enhancing the accuracy by selecting only the relevant features for model training. Using this approach, the run-time complexity of developed toxicity model is substantially reduced. The idea is to create a contextual adaptation of the models by hybridizing the decisions trees to enhance the features interpretability and accuracy both. In the later quarters of this thesis, we argue for the idea of effective aggregation of chemical knowledge about molecules in toxicity prediction. Molecules are represented in various data formats such that each format has its own specific role in predicting molecular activities. We propose various deep learning ensemble approaches to effectively aggregate different chemical features information. We have applied these methods to quantitative and qualitative molecular toxicity prediction problems and have obtained new stateof- the-art accuracy improvements with respect to existing deep learning methods. Our ensembling methods also prove helpful in making the model’s prediction robust over a range of performance metrics for toxicity prediction.