Research Project
9254390
Summary The objective of ?Assay Central? is to compile a comprehensive collection of datasets for structure-activity data for a broad variety of disease targets and absorption, distribution, metabolism, excretion and toxicology (ADMET) properties, in a form that is immediately ready for model building and other forms of analysis using cheminformatics methods. This is aided by the existence of many sources of curated open data, and one in particular, ChEMBL 1, 2 will be used as the nucleus in Phase I. This bioassay data collection is incredibly valuable, but not currently provided in a form that is ready-to-go for use by small research and development (R&D) organisations that do not have their own in-house cheminformatics teams. The effort required to preprocess, filter, merge, validate and normalize the structure and activity data requires a great deal of software expertise and medicinal chemistry domain knowledge, which are key skillsets that are rare and expensive to combine within the same team. Create a script to analyze the databases like ChEMBL, selected parts of PubChem and others 1, 2 and partition it into groups of compatible activity measurements against the same target. We will seed the dataset collection with a set of 1840 target-assay groups that have been recently extracted from the ChEMBL v20 database, as well as EPA Tox21 measurements 3, using methodology that we have already developed (similar to that described in 4). We will build error checking and correction software. We will apply best-of-breed methodology for checking and correcting structure-activity data 5 which errs on the side of caution for problems with non- obvious solutions, so that we can manually identify problems and either apply patches, or datasource-specific automated corrections. We will build and validate Bayesian models with the datasets collected and cleaned. For each of the target-activity groups, we will create a Bayesian model using ECFP6 or FCFP6 fingerprints, and this will be one of the primary outputs from the project. Models will be evaluated using internal and external testing with receiver operator characteristic (ROC > 0.75), the integral of the true-negative-rate ? true-positive-rate curve as well as the enrichment,6 Kappa value and positive predicted value.7 We will develop new data visualization tools as a proof of concept in phase I. We have already begun to explore preliminary visualization methods using multiple models, but these have so far focused primarily on a handful of machine learning models selected from a very large list. New visualization techniques are required to summarize large matrices of data, e.g. a list of proposed structures vs. thousands of target models. In Phase II we will expand by upgrading to newer ChEMBL releases, selectively incorporating screening runs from other databases (such as PubChem 8), These tools will consist of software created explicitly for this project (particularly web-based interfaces), as well as enhanced functionality added to 3rd party tools that we influence (e.g. mobile apps) and open source projects that we have already contributed to (e.g. CDK for fingerprints and Bayesian modelling). We will widely publicise Assay Central at conferences and in papers. Being able to use transparent computational models simultaneously for visualizing activity trends for multiple targets (both diseases and ADMET) removes the burden of curation or purchasing and maintaining expensive software, and drastically simplifies the addition of new data. It also represents a new frontier of drug discovery as a world of small, agile distributed R&D organizations has access to valuable public datasets that can inform their research. Such computational models will assist in drug repurposing efforts internally and with our collaborators while likely identifying new compounds for a wide array of drug discovery projects.
