Case Studies

Custom Modeling

admetis brings powerful data mining and dynamic visual inquiry methods to enable the building of custom models from scratch. Predictive models can be built starting just with structures and known properties. The Workflow Browser guides the user through a logical sequence of steps that need to be performed on a dataset for building custom models and running predictions. These workflows can be further customized and configured to suit specific customer processes.

admetis offers Decision Tree, Neural Network and Support Vector Machine algorithms for classification models. For building regression models, Multiple Linear Regression and Neural Network methods are available. The user can evaluate multiple algorithms and select the one that is the most predictive.

admetis provides multiple interactive views of data. These views are dynamically linked, i.e., selection in one view is propagated to all other open views. They are highly customizable and can easily be exported into most common applications. Data can be filtered dynamically using slider-bars and check boxes. The highly enhanced scripting engine allows scripting almost anything.

Complex functions or even entire workflows can be scripted and integrated with external programs. This approach has been applied to designing custom models for various Pharma organizations. Some of the case studies described below illustrate the value of this approach.

Case Study I: Reduction of expensive in vitro toxicity testing

A Pharma organization had interesting hits from their screens. However, hERG binding was a major issue within their libraries. This testing resulted in the organization spending about $600-$1000 per compound. They tested truTox for its predictive ability on their compounds (from multiple therapeutic programs). truTox was able to identify 94% of binders in a proprietary dataset of compounds. New compounds are run through truTox first and prioritized to minimize in vitro testing.

Case Study II: Prioritising Leads for Animal Studies

A Pharma organization had a bottleneck at the animal studies stage with the result that promising leads had been delayed in the past. The challenge was how does one pick promising (good clearance) leads prior to performing animal studies? The proposed solution was to build a mouse AUC predictor based on proprietary lead-like structures.

All leads are run through the predictor first and re-prioritized. Promising compounds are picked first for animal studies. This has resulted in demonstrated time and effort savings within the organization.

Case Study III: Prioritising Hits for Animal Studies - CNS penetration

An organization is interested in CNS targets; hence their drugs must penetrate the blood-brain-barrier (BBB). This resulted in many hits to be tested in animals. Strand developed a BBB penetration predictor based on proprietary lead-like structures as well as public domain compounds.

The model was able to predict 80% of all compounds predicted within the correct class. 100% of BBB penetrants and non-penetrants were predicted either correctly or in the adjacent classThe model is now used within organization prior to animal studies as a prioritization tool.