PMC

Part of the clustering of the full set of 173 generated rules.
Notes: We performed rule alignment on each subfamily (indicated by red rectangles in the dendrogram). Two interesting rules are indicated by (*).

Definition of neighbouring residues.
Notes: For the mutated residue, Asn180 of protein Q13496, a sphere of radius 10 A° is drawn with the residue in the centre. Any residues that lie within the sphere are defined as neighbours.

Part of a screenshot with four induced rules obtained using Aleph with noise = 0.5%, minpos = 5, nodes + 50,000.
Notes: Users can click on the + icon to see the covered examples. The keyword “sub_family_conservation” was used as a filter in this screenshot.

Construction of background knowledge from MSV3d.
Notes: Each mutation in the database is identified by a unique identifier ‘id’ and the values of each. Modeh defines the head of a hypothesised clause, while Modeb declares the predicates that can occur in the body of a hypothesised clause. The asterisk * in the mode declarations indicates that the corresponding predicate can be called many times during the construction of a hypothesised clause.

Mutation data model.
Notes: Each missense mutation is characterised by physico-chemical features (size, charge, polarity, hydrophobicity, etc), evolutionary information and 3D structural features. In addition, it may have one or more than one neighbouring residues, each of which can belong to a single class, based on Koolman’s classification.

Main steps for an ILP application include: (i) mutation selection from MSV3d, (ii) definition of negative/positive examples in the training set, (iii) background knowledge creation, (iv) selection of the ILP system, (v) selection of the ILP parameters (number of nodes, noisy..) and optimization of the predicates in the background knowledge, (vi) model evaluation using K-fold cross validation, and (vii) the final rules used for interpretation.