Interpretable Models of Antibiotic Resistance with the Set Covering Machine Algorithm

source: GoogleTechTalks     2017年2月16日
A Google TechTalk, 13 Feb 2017, presented by Alexandre Drouin.
ABSTRACT: Antimicrobial resistance is an important public health concern that has implications in the practice of medicine worldwide. Accurately predicting resistance phenotypes from genome sequences shows great promise in promoting better use of antimicrobial agents. For instance, treatment plans could be tailored for specific individuals, likely resulting in better clinical outcomes for patients with bacterial infections.
Sparse machine learning algorithms are appealing tools in this context, since they make use of a concise set of features, which can be further interpreted by domain experts. However, in extremely high dimensional settings, which are common in genomics, the main challenge remains resistance to overfitting.
In recent work, the Set Covering Machine (SCM) algorithm has been used to obtain concise, expert interpretable, models of antibiotic resistance for 6 pathogenic bacterial species. Known and validated resistance mechanisms were recovered within minutes of computation. An empirical benchmark showed that the SCM compared favorably to more complex learning algorithms (e.g., L1-SVM), both in terms of accuracy and sparsity. Moreover, a theoretical analysis of the method revealed that the SCM has an uncharacteristically strong resistance to overfitting in genomic contexts.
In this talk, I will present the SCM algorithm, along with an efficient implementation for genomic data (https://github.com/aldro61/kover). I will rely on theoretical results and an application to antibiotic resistance to demonstrate that this algorithm is well-suited for predictive modeling in genomics."
ABOUT THE SPEAKER: Alexandre Drouin is a PhD candidate in Machine Learning and Computational Biology at the Université Laval, advised by Prof. François Laviolette.

No comments: