Home- Research
- Conferences
- Program & Schedule
- Location
- KM in R&D Workshop
- Sponsors & Exhibitors
- Call for Contributions
- Bursary Awards
- Poster Session
- Program 2007
- Knowledge Management Workshop
- ADMET
- Beger, R
- Cherkasov, A
- Contrera, J
- Helma, C
- Hopfinger, A
- Jamois, E
- Klon, A
- Madden, J
- Pelletier, D
- Poroikov, V
- Richard, A
- Tomaszewski, J
- Program 2006
- Photo Gallery
- Workshops & Training
- Program
- Exhibition
- Registration
- About
- Overview
- Contact
- Support
- Schedule
|
|
|
|
|
|
|
|
|
| About Eric Jamois (Strand Life Sciences) |
|
Eric Jamois, holds a Ph.D. in Organic Chemistry from North Carolina State University. He has 15 years of experience with leading life sciences informatics companies in Europe and US. He started his career with Tripos in 1991 as a Business Unit Manager in Paris. His more recent positions were as Associate Director, Business Development at Accelrys and Director of Business Development at Inpharmatica. He is now Vice-President of Business Development at Strand Life Sciences.
|
|
Towards Cognizant Data Models for SAR and Modeling of ADME/Tox Properties
Eric Jamois*, Kas Subramaniam and Ramesh Hariharan , Strand Life Sciences
We place significant demands on SAR models in terms of training set selection, statistical behavior, interpretability, usage traceability and predicting abilities. Recently, a considerable emphasis has been placed on assessing the true prediction scope of models such as those used in the prediction of ADME/Tox endpoints. That is, defining a space where we can expect a model to perform within given accuracy guidelines. Although internal statistics can be captured as metadata, these do not generally provide sufficient information to forecast external predictions.
We present herein the status of our work towards a data model capturing larger amounts of information. We introduce the concept of a “cognizant” model, that is, a model that is self aware of its predictive ability and other performance related metrics. This new data model also provides an open access to the underlying training set, thereby providing easier update mechanisms when new structures and data becomes available.
We believe that there are several practical applications of this concept. For example, models could be queried initially regarding their ability to predict on a given chemotype. This may be particularly useful when dealing with large collections of models and searching for a suitable one. With appropriate modeling of prediction error, we can quickly identify models which predict within given accuracy guidelines.
|
|
|
|
|
|
|
|
|
|
|
