Inferring Gene Regulatory Relationships in Drosophila Using Cell Morphology Data

Oaz Nir, Massachusetts Institute of Technology

We have made rapid progress in developing computational techniques to infer gene regulatory relationships from the massive amounts of phenotypic data collected from knockout and overexpression experiments for genes implicated in Drosophila cell morphology and migration. Our approach involves multiple levels of modeling and computational approaches. First, we develop a mathematical model that allows us to use phenotypic data to discover cascades of regulatory relationships. Second, we implement software to carry out the computations necessary for making these regulatory inferences. Third — and this is really the key step for the success of this approach — we use techniques from machine learning to train the parameters of our model on known data (here, PPI network data for the genes of interest).

Initial results show promise for our modeling approach. A preliminary study involved carrying out only the first two steps outlined above. That is, we developed a model for predicting regulatory relationships that we did not train on known interaction data. The parameters for this model were selected on the basis of several reasonable modeling assumptions. This primitive model managed to predict key regulatory relationships, particularly at the level of GTPase interaction with GAPs and GEFs (the model did not perform adequately for regulatory relationships further downstream in cascades). However, by carrying out the training step of our approach, we expect greatly improved predictive strength for regulatory relationships at all levels of these cascades.

Regulatory predictions from phenotypic data will be used in conjunction with predictions from transcriptional data. Since these data are somewhat orthogonal, we expect that pooling predictions inferred from both datasets will lead to improved accuracy in discovering regulatory relationships.

Abstract Author(s): Oaz Nir, Chris Bakal, Bonnie Berger