To limit the combinatorial explosion in pathway prediction, enviPath allows to learn and apply relative reasoning models. Rule-based relative reasoning models are available, as well as multi-label machine learning algorithms. If employed during pathway prediction, models predict probabilities for each applied rule and the corresponding educt and product compound, and can thus be used to truncate the predicted pathway.
Pre-trained relative reasoning models for the packages provided by us are available in enviPath.
Moreover, the user can train own such models on any collection of packages. Hence, new models can reflect exactly the conditions of the selected packages, and predicted probabilities are automatically adapted to the type of data used as input in the training process. In more detail, relative reasoning models will be trained on all compounds, rules and reactions that are available within the selected package/s.
Applying relative reasoning
Relative reasoning can be applied for pathway prediction by selecting/creating a dedicated setting when initializing the pathway prediction. Within each setting the relative reasoning model and a probability threshold can be selected (all edges with a predicted probability below this threshold will be pruned).
Currently, two different types of relative reasoning are available.
Rule-Based Relative reasoning
In rule-based relative reasoning, the possibility to apply a rule depends on the presence of other applicable rules. Practically, this requires additional rules for the prioritization of rules and the resolution of conflicts. These meta-rules, or relative reasoning rules, express that some reactions take priority over others, and vice versa, and that some reactions only occur if others are not possible. In enviPath, we provide trained relative reasoning rules on the enviPath core package. However, users can train their own relative reasoning rules using their own packages.
Machine Learning-Based Relative Reasoning
Individual machine learning models can be trained based on the structure of a set of compounds and all transformation rules triggered by them. These models can predict probabilities for all transformations of a new compound. As the learning problem clearly is a so called multi-label classification problem, we extended the machine learning approach using multi-label classifiers to improve the prediction.