An Integrated Approach For Validating and Delivering In Silico Human Liver Microsome Clearance
Models To Discovery Scientists
Ignacio Aliagas, Alberto Gobbi, Cyrus Khojasteh, Jane Kenny, Man Ling Lee and Daniel F. Ortwine
Discovery Chemistry, Genentech, Inc., 1 DNA Way South San Francisco, CA 94080
Metabolism is one of the major clearance pathways of drugs from the body. High-throughput liver microsome and hepatocyte stability assays are now in routine use in the pharmaceutical industry. Such assays produce a substantial amount of consistent data, permitting predictive computational models to be developed. Use of such models to forecast the stability of compounds ahead of synthesis could be a valuable tool for prioritizing compound synthesis and compound design. We have developed a model to predict clearance from human liver microsome (HLM) cells using a dataset of >20,000 compounds. The model output is expressed as categories and probabilities. The cross-validated Q-square was 0.50 for the human liver microsome model with a RMS error of 3.4. The R-square for a prospective set of 5,500 newly synthesized compounds was 0.41 with a RMS error of 3.8. This model, along with others that predict various DMPK assay end points, have been integrated into a number of desktop tools available to discovery scientists.
Calculations are also performed on all registered compounds and results uploaded to our corporate database. All models are updated on a regular basis.
In addition to desktop availability of all models, we have developed easy to use tools to evaluate model performance on individual compounds, chemical series, or project-wide sets of compounds. These tools provide easy-to-interpret charts and plots, giving discovery scientists an instant assessment of how well the models are working for their compounds. These capabilities, along with strong support from management for their use, has resulted in an across the board decrease in the percentage of synthesized compounds that are unstable in HLM cells.
We will present statistics from the HLM model using the validation tools we developed to generate charts and plots to show prediction accuracy on compounds not in the training set. We will also show a three year history of measured HLM stability data, demonstrating that if the model is consistently used as a filter in synthesis decisions, an overall decrease of HLM in vitro clearance can be achieved.