Probabilistic reasoning in intelligent systems: networks of plausible inference
Probabilistic reasoning in intelligent systems: networks of plausible inference
Elements of information theory
Elements of information theory
Diffusion Kernels on Statistical Manifolds
The Journal of Machine Learning Research
Arevir: a secure platform for designing personalized antiretroviral therapies against HIV
DILS'06 Proceedings of the Third international conference on Data Integration in the Life Sciences
A new metric for probability distributions
IEEE Transactions on Information Theory
| Hi-index | 0.00 |
This chapter provides a review of the challenges machine-learning specialists face when trying to assist virologists by generating an automatic prediction of an outcome of HIV therapy. Optimizing HIV therapies is crucial since the virus rapidly develops mutations to evade drug pressures. Modern anti-HIV regimens comprise multiple drugs in order to prevent, or at least delay, the development of resistance mutations. In recent years, large databases have been collected to allow the automatic analysis of relations between the virus genome other clinical and demographical information, and the failure or success of a therapy. The EuResist integrated database (EID) collected from about 18500 patients and 65000 different therapies is probably one of the largest clinical genomic databases. Only one third of the therapies in the EID contain therapy response information and only 5% of the therapy records have response information as well as genotypic data. This leads to two specific challenges (a) semi-supervised learning --- a setting where many samples are available but only a small proportion of them are labeled and (b) missing data. We review a novel solution for the first setting: a novel dimensionality reduction framework that binds information theoretic considerations with geometrical constraints over the simplex. The dimensionality reduction framework is formulated to find optimal low dimensional geometric embedding of the simplex that preserves pairwise distances. This novel similarity-based clustering solution was tested on toy data and textual data. We show that this solution, although it outperforms other methods and provides good results on a small sample of the Euresist data, is impractical for the large EuResist dataset. In addition, we review a generative-discriminative prediction system that successfully overcomes the missing value challenge. Apart from a review of the EuResist project and related challenges, this chapter provides an overview of recent developments in the field of machine learning-based prediction methods for HIV drug resistance.