Cost-sensitive and parametrized Hopfield Networks for gene function prediction

To deal with this problem we developed COSNet (COst Sensitive neural Network), a novel cost-sensitive family of parametrized Hopfield networks, whose characteristics can be summarized as follows (Frasca et al. 2013, Bertoni et al. 2011):

1. Class labels and neuron states are conceptually separated. In this way a class of Hopfield networks is introduced, having as parameters the values of neuron states and the neuron thresholds.

2.The parameters of the network are learned from the data through an efficient supervised algorithm, in order to take into account the unbalance between positive and negative node labels.

3. The dynamics of the network is restricted to its unlabeled part, preserving the minimization of the overall objective function and the a priori information and significantly reducing the time complexity of the learning algorithm.

A regularized version has also been proposed to deal with extremely unbalanced functional classes (Frasca et al. 2012) and we are currently working on SINTNET, a COSNet-inspired unbalance-aware method for integrating multiple sources of bio-molecular data (Frasca et al. 2013).

Variants of COSNet exploit gene multi-functionality to improve gene function predictions (Frasca 2015), and a new multi-parametric Hopfield network (Hopfield multi-category -- HoMCat), designed to take into account a priori-known "categories" of neurons within networks, has been successfully applied to the multi-species protein function prediction problem (Frasca et al. 2015).