In silico methods for toxicity prediction

Outline

• In silico toxicology methods
• Lazar framework
• Products and services

In silico methods

• Systems biology/molecular modeling
• Expert systems
• Data driven techniques

Systems biology/molecular modeling

Model individual events (e.g. receptor interactions, (de)toxification) of the adverse outcome pathway

• Mechanistic interpretation

• Calculations/simulations may be very time consuming
• May require a lot of experimental data for parameterization
• Impossible to model mechanisms of complex toxicological endpoints

Examples: VirtualToxLab/Biograf

Expert systems

Formalize expert knowledge about chemicals and toxicity mechanisms and create a software program

• Mechanistic interpretation

• Model creation very time consuming
• Many toxicity mechanisms are poorly understood or even unknown
• Error prone and hard to validate (strong tendency towards overfitting)

Examples: Derek/Lhasa

Data driven

Use all existing data for a particular endpoint and apply machine learning/QSAR algorithms in order to create a prediction model

• Comparably fast
• Applicable for every endpoint with sufficient experimental data
• Sound validation possible

• Applicability domain/model quality depends on experimental data
• Mechanistic relevance has to be extracted from models/descriptors/predictions

Examples: Classical QSARs, Topkat, Multicase, lazar

Lazy-Structure-Activity Relationships (lazar)

Automated read across predictions

• Find similar compounds (=neighbors) with measured activities
• Create a local (Q)SAR model with neighbors as training compounds
• Make a prediction with this model

Lazar estimates the confidence (applicability domain) for each prediction

Chemical Similarity

Can be based on

• Chemical structures
• Chemical properties
• Biological properties
• ...

Lazar uses activity specific similarities

Activity specific similarities

Consider only relevant (i.e. statistically significant) substructures, properties, ... for similarity calculations

Algorithms for finding relevant substructures (by A. Maunz):

• Backbone refinement classes (BBRC)
• Latent structure mining (LAST)

http://lazar.in-silico.ch

http://lazar.in-silico.ch

http://lazar.in-silico.ch

http://lazar.in-silico.ch

Lazar limitations

• Model quality depends on data quality
• Applicability domain depends on learning instances

in silico toxicology gmbh

Open source software and algorithm development

• Predictive toxicology and QSAR models
• Toxicological data mining
• Life science webservices and data warehouses

Why open source?

• Clear and unambiguous documentation of implemented algorithms essential for scientific software (also required by many regulatory guidelines)
• Collaboration with partners, projects and external contributors
• Establishment of international standards
• Security of investment

EU Research projects (FP6/7)

Free products and services

Issue tracker, documentation, ...

Commercial products and services

• Lazar "software as a service" (SaaS): secure access for confidential predictions, batch predictions, ...
• Virtual appliances with lazar software for in-house/desktop installation
• Installation services
• Phone and email support

Commercial products and services

• Virtual toxicity screening of compounds and libraries
• Development of prediction models for new endpoints
• Scientific programming, contract research and consulting

Contact