Sébastien Lemieux and his team develop bioinformatics algorithms and artificial intelligence methods for the analysis of transcriptomics, proteomics, and high-throughput chemical screening data.

Research theme

Current sequencing, mass spectrometry, and screening technologies produce data volumes that can only be interpreted through intermediate computational representations. Conventionally, mRNA sequences are aligned to a reference genome, then summarized as expression levels for some 20,000 genes. The molecules in a compound library are described by the graph of chemical bonds between their atoms.

These representations are convenient, but they rest on human-defined categories that are incomplete and continually revised. Genome annotation leaves out non-canonical transcripts, novel isoforms, and gene rearrangements, often precisely where such events distinguish one tumor from another. Chemical structure, for its part, is a poor foundation to predict biological activities.

The laboratory is interested in minimizing the information lost during these transformations. Its work focuses on representations built directly from the measured signals, and on the deep learning architectures able to extract biologically meaningful information from them.

Research objectives

The Functional and Structural Bioinformatics Research Unit divides its efforts between two missions. On one hand, Sébastien Lemieux’s team develops data pre-treatment techniques in order to increase accuracy.

It also seeks to create tools to better extract biologically sound knowledge from heterogeneous data sources. The group notably tries to more accurately measure the expression levels of messenger RNA, as well as to be able to identify splicing events or the presence of mutations.

The raw data that Investigators work with are often altered by the experimental biases stemming from each of the steps leading to the production of data. By uncovering these biases one-by-one, then by measuring the impact of each one on the data, Investigators are in a position to integrate corrective measures into their algorithms in order to minimize their impact.