Research Project
9846761
Project Summary The objective of ?MegaPredict? is to enable scientists to generate predictions for a natural product (or any molecule) and identify targets for efficacy assessment as well as identify any potential liabilities. We are building on our previous work which has compiled a comprehensive collection of datasets for structure-activity data for a broad variety of disease targets and other properties, in a form ready for model building. All of these models utilize the many sources of curated open data, including ChEMBL, ToxCast etc. We have developed a prototype of MegaPredict that utilizes Bayesian algorithm and ECFP6 fingerprints to output a list of prioritized ?targets?. We realize that neither the algorithm or the descriptors may be optimal therefore we propose to address this as we validate MegaPredict and develop a product over this proposal. Our team is suitably qualified to develop the software needed and we will leverage our large collaborator network to assist us in validating the activity of compounds. We will initially create a script to take a natural product and score it against many thousands of machine learning models then rank the outputs to propose efficacy targets. We will use over 12,000 ChEMBL derived target-assay / bioactivity groups extracted from the ChEMBL v24 database, as well as EPA Tox21 measurements and other public datasets, using methodology that we have already partially developed. We can repeat this process for over 200 published compounds and access the outputs versus what is known. We intend to compare how the approach performs with synthetic drugs or drug-like compounds as well as natural products. We will assess whether other machine learning algorithms and molecular descriptors can improve predictions. As we generate machine learning models such as Linear Logistic Regression, AdaBoost Decision Tree, Random Forest, Support Vector Machine and deep neural networks (DNN) of varying depth we will assess the predictions for natural products and compare with the Bayesian approach. We will compare ECFP6 with other 2D, 3D descriptors and physicochemical properties in order to identify the optimal combination for generating predictions for natural products and compare how this differs for synthetic compounds. We will validate our predictions for natural product efficacy assessment. We will work closely with multiple academic groups to generate predictions for at least 20 natural products of interest against over 20 different targets or diseases. Our goal will be to identify potential targets that were previously unknown and then generate in vitro data inhouse or with academic collaborators. Develop a prototype user interface for input of a structure, processing an input molecule and output of prioritized targets and liabilities. We have developed multiple software prototypes (e. Assay Central, MegaTox, etc.) previously and will ensure a user-friendly interface and develop new visualization methods and algorithms for prioritizing potential predicted targets based on the outputs of thousands of machine learning models. In Phase I, we will use the software internally with collaborators to rapidly prototype it. In Phase II we will develop a commercial product, and greatly expand our validation by building a larger network of academic and industry partners that would help to prioritize features of most relevance. Using the machine learning models which we have for natural products is limited because ECFP6 fingerprints cannot distinguish between these very different classes of molecules. But this provides us with an opportunity by going for a "pharmacophore" style approach (ideally without using 3D conformations directly). We will therefore focus on developing a ?3D shape-based fingerprint? or developing a novel ?2D fingerprint? that captures the ?3D shape? for natural- and druglike molecules. Currently, the public datasets in ChEMBL and PubChem etc. are made up of mostly druglike molecules, but if we have fingerprints that can compare drug-like and natural product-like molecules then we can likely reliably use our MegaPredict models for natural products as well. We can also attempt to rank natural products with our ChEMBL models or we can look through catalogs of druglike compounds using models derived from natural products. That would be an important innovation. Additionally, in Phase II it would be important to see if we could find uses for natural products with any of the 7000 rare diseases. Developing software that predicts potential natural product drug interactions with various targets could be useful to regulatory organizations as well as the pharmaceutical industry and may broaden utility of being able to more effectively mix natural product and druglike compounds in models will have a profound effect on the value of cheminformatics in this arena.
