Seminars-Gene regulatory network inference from time series: machine learning approaches

EVENT : C3BI Seminars – Seminars-Gene regulatory network inference from time series: machine learning approaches


Speaker : Florence D’Alché-Buc, from Laboratoire de Statistiques et Applications, Télécom ParisTech, Paris Time : 2PM Date :   June 9th, 2016 Location : Room Jules Bordet  – METCHNIKOFF (67) ,Institut Pasteur, Paris


We consider the well known problem of gene regulatory network inference from time series measurements (gene expression, protein concentration) under the angle of Machine Learning. We address two different instances of this problem.

In the first instance we assume no information about the structure and no perturbed data: we develop a nonparametric dynamical model and a learning algorithm to identify the biological dynamical system and its network structure from the data. It gives us the opportunity to present a novel family of vector-valued models called operator-valued kernel based models. These models extends the well know family of scalar-valued kernel to multiple output prediction as well as structured output prediction.

If the results exhibit very good results compared to other methods, it is clear that the bottleneck of these methods is the lack of additional measurements especially, perturbation data that allow to identify the underlying system accurately. We prone the idea that when the goal of experiments is to estimate a model, data have to be chosen consequently. This introduces the second instance.

In this case, we address the design of experiments together with the identification of a parametric model (differential equations) of the network. The idea is to help the biologist to choose the experimental data, especially perturbation data like knock out, knock down, as well as the sampling rate of the measurements with a limited budget for experiments.The experimental design is thought as a one-player game and is solved using a active learning procedure that allows to both optimize the number and the cost of experiments to perform and the network parameters to be identified. On both problems, we show numerical results on the well known benchmarks of the DREAM Challenges. To conclude, new perspectives about dynamical modeling and machine learning and their use in computational biology are drawn.

N.B. The first work is a joint work with Néhémy Lim and Cédric Auliac, the second work is a joint work with Adel Mezine, Véronique Letort and Artemis Llamosi.


Due to security policy in Institut Pasteur, please register before if you plan to come to this meeting