Trainer Engine

Predict molecular properties and boost the efficiency of machine learning workflows.

Trainer Engine makes chemical, physical and biological activity predictions available by streamlining learning from input data with high accuracy, reliability and confidence at scale. The framework simplifies sharing models, and managing the machine learning lifecycle.

Translate data to prediction

The biological, chemical and physical properties of molecules are encoded in their molecular structure. The grand challenge is to discover the relationship between structural properties and the measured activity. Where data is measured, collected and curated for a series of compounds, there is an opportunity to find the hidden relationships.

Complete

From input data to implementation of validated models.

Smart

Chemical structure normalization, high-quality and customizable descriptors.

Convenient

Rich feedback and visualization for model optimization.

Reproducible

Central model repository to support selecting production grade models.

Integrable

Access to predictions from a built-in graphical interface, Design Hub or other design platforms.

Predictive

Successful models built on bio-activity, ADMET and phys-chem targets.

Machine learning

Trainer Engine offers automatized, high performant, and configurable descriptor generation on normalized chemical data. It provides a wide range of machine learning algorithms including Random Forest, Gradient Boosted Trees, Support Vector Machine, and Logistic Regression. Model performance is automatically evaluated, and the most important statistical parameters are calculated both for regression and classification cases. Feature selection is supported by seamless re-training, based on feature importance in the case of Random Forest. Calibrated error is calculated using the conformal prediction framework. Applicability domain assessment is enabled by returning the most similar structures and corresponding activity data from the training set.

Analyse classification

Visualization

The collection of generated models is accessible from the central service in order to benchmark their prediction power and provide insights into their behavior. Trainer Engine stores the models in a repository to ensure reproducibility and comparison of their parameters conveniently. The configurable analysis view comes with a classification and a regression layout presets with optimized tables, charts and molecule visualizations.

Trainer Engine Analyze

Usage

Single / batch mode
Trainer Engine provides services for novel predictions in single or batch mode (through SDF file upload).
REST API
For programmatic access, we recommend using the REST API interface to automatize the machine learning workflow and integrate predictions into design tools like Design Hub or other third-party applications.
Optimized for model building
Trainer Engine graphical user interface offers a rich set of tools to build, validate and compare models. It is loved by computational chemists.
Simple prediction interface
It comes with a lightweight prediction application optimized for end-users, the Playground.

Articles

Leveraging data as future insight

This short Cheminfo Story episode presents predictive model building for passive permeability using publicly available data.

Chemical consistency: from principles to applications

The goal of the presentation is to highlight the effect of chemical normalization on investigating correlations and building predictive models.

Automation of building reliable models.

Summary of descriptor selection and hyperparameter optimization for a bioactivity data set including >150 ChEMBL targets. As a case study we also present detailed results on hERG model development.

Prediction driven design of hERG liability free compounds.

The new ChemAxon ADMET plugin group exploits the power of machine learning methods on curated data sets to support the drug design and medicinal chemistry optimization with reliable models and predictions.

Trainable models

This webinar gives insight into recent development of ChemAxon's Calculators and Predictors towards trainable models and new ADMET models.