We designed an optimized machine learning pipeline aiming at identifying short molecular biomarker signatures. The identified signatures can be used for tasks like Breast Cancer BC screening or predicting a reponse to Neoadjuvant Chemoterapy NAC treatment in BC. The pipeline is based on Random Forest RF method. The core of the pipeline is a combinatorial testing of all the combinations possibly generated from a set of k genes. Basically, the pipeline starts by a feature selection procedure aiming atof reducing the dimensionality of the dataset from m to k (k<<m). All the 2^k-1 combinations are generated, and exhaustively tested for their prediction performance in terms of AUC, sensitivity and specificity. The pipeline has been already used to identify circulating microRNA-based signatures for BC screening [P. Freres et al. 2016]. Our pipeline is an extended and optimized version of the pipeline used in this study. The optimization concerned the robustness of the identified signatures and the reduction of the computational time. We incorporated three modules aiming at: i- Optimizing the Feature Selection FS process by identifying the most robust and stable FS method, ii- Optimizing the modelization module by comparing a panel of 15 RF methods and identifying the most stable and fastest RF method, and iii- Designing and implementing a fitering and reduction strategy of combinatorial tests. Each module is tested individually, and ongoing work is being done to assemble all the modules in one complete pipeline.

In this study, we've investigated and explored the robustness of feature seletion methods (or gene selectors) in RNA-seq data. The robustness here is defined as the sensitivity of the selectors to perturbations in the dataset. Data pertubations can come from instance perturbation and/or changing of RNA-seq normalization method. We designed a comparison pipeline and applied it on three RNA-seq datasets acquired from different sources. The datasets are normalized by using 17 state-of-the-art RNA-seq normalization methods, and feature selection is performed using a panel of 17 selection methods. We introduced novel scores SPR/SPN based on Kuncheva stability index to measure the impact of data perturbation on the stability of gene selectors. Furthermore the selectors are evaluated in stability in the context of short biomarker discovery, i.e. the ability of the selectors to be robust in the selection of top ranked genes significantly relevant to the classification problem. We found the selectors behave differently according to our stability scores SPR/SPN for the same dataset, and each selector showed different stability scores across the three datasets. Furthermore, the selectors are more impacted by the change of normalization methods than the change of resampling partition. Finally, for short biomarker signature discovery, we found that simple univariate test for Differential Expression DE like KW and GLM-LRT showed high stability scores, and can compete with complex wrapper and embedded selectors.

Biomarker signature discovery remains the main path to develop clinical diagnostic tools when the biological knowledge is still missing for a pathology. The ability to find the best and short signature relies on the robustness of the models that can be built on such set of molecules. Machine Learning (ML) is among the fields of choice to build such models, and multiple ML methodologies are often compared for their average AUC performance prior usage. Here we describe two AUC Hyper-Stability scores, the HRS and the HSS, as complementary metrics to the average AUC, that should better help researcher in the choice of the best ML strategy. To emphasize the interest of this metrics over the ML methodologies, we compared 15 different Random Forests implementation as an example. Additionally, the modelization time of each ML strategy was computed to further help in the decision of the strategy. Our findings show that the Random Forest implementation should be chosen according to the dataset and the classification being evaluated. Therefore, no Random Forest implementation can be used universally forany classification and on any datasets, but each of them should betested for their average AUC performance and AUC stability prior to analysis.

We designed a score allowing to improve the computational time of the whole BSD pipeline. After performing the feature selection to select the most relevant k genes, all possibly generated 2^k-1 combinations are tested for their performance in AUC. To avoid an exhaustive testing of all the combinations which is a computationaly comsuming process O(2^k), we designed a strategy allowing to reducing the solution space by filtering out less performing combinations, and keeping only certainof them. Those combinations are then evaluated to be potential solutions for the problem. This strategy is applied before any calculation of the performance of the combination space. Basically, the dataset is splitted on two parts, and a Mean Decrease in AUC MDAUC on the test partition is calculated for each gene selected by the FS process. This reflects the contribution of the genes. This contribution is then calculated at the signature level, and each signature score is compared to a threshold on basis of what less performning signatures are discarded from downstream analysis. The threshold is determined on basis of visual inspection of the correlation between the signature scores and the AUC performances of a sample of signatures randomly selected from the original solution space. We applied our strategy on the same miRNAs dataset used in [P. Freres et al. 2016]. Similar results (the identified best signatures and their performances) are obtained by filtering out 86% of signatures and processing the remainnig 14% signatures as potential solutions. Neverthless, the startegy is still basic and more tests are needed. In addition, more work about the value of the threshold should be done towards and automatized filtering strategy.

We tried to validating the best signatures for BC in Belgian patients found in [P. Freres et al. 2016] on a Rwandan Cohort of 157 samples (73 Healthy vs 84 Cancer). The results showed a drastic reduction in performance of the signatures. Then we tried to design a merged signature able to predict BC in both Belgian and Rwandan patients by merging the two datasets, and applying the BSD pipeline. 133 signatures have been identified as candidates with a performance on test data of AUC>=0.86. The best signature is then validated separately on an independent Belgian cohort with AUC=0.76, and an independent Rwandan cohort with AUC=0.81. Furthermore, the signature is validated on 4 BC subtypes cohorts: Luminal-A, Luminal-B, TNBC, and ER-HER2+, and on 2 cohorts startified by petient age (Under 50 and Over 50). Finally, we designed a set of short predictive signatures for BC screening in Rwandan women by using the Rwandan cohort only. 83 signatures have been identified as candidates with a performance of AUC>=0.85. The best signature has been validated on an independent Rwandan cohort leading to a validation performance of AUC=0.83.

In this work, 22 ER+ Breast Cancer and their paired adjacent non-malignant tissues were analyzed by strand-specific- RNA-seq. All the analysis has been done from the QC of the reads to the summarization of genes per conditions. The count matrix containing the expression of the genes has been used to study the preservation of coexpressed modules between conditions using network approaches (WGCNA). The approach has been applied on DE genes only. We've used the healthy tissues to build the reference gene coexpression network. 16 modules containing highly coexpressed genes have been identified in the reference network. Quatitative assessment of the preserved modules in the ER+ tissues has been determined using a Z-zcore calculated as the average of module density and module connectivity. Among the 16 modules, 2 of them showed a low preservation and have been considered in pathways enrichment analysis because they may highlight dysregulation pathways in the tumor tissues (acquired or lost when compared to the healthy network). Pathways Enrichment Analysis have been then performed on the genes belonging to the two modules using ConsensusPathDB.

This project was in collaboration with the Laboratory of Tumors and Development Biology (LBTD-GIGA Cancer). The project aimed at identifying three classes of novel long intergenic noncoding lincRNAs that are diffrentially expressed between cell types (CISPLATIN and PARENT) and cell subtypes. I got stranded RNA-seq reads after sequencing 8 parent and 8 cisplatin cell types (bulk RNA-seq). All bioinformatics analysis of sequencing data including the reference-guided genome assembly and the prediction of novel lincRNAs has been performed.Novel lincRNAs discovered by the method are then classified as NONCODE for lincRNAs that are identified as such but not annotated yet, and bona fide lincRNAs transcripts which are not annotated and not identified yet. Furthermore, the method allowed to classify novel lincRNAs as single exonic SE and multi-exonic ME transcripts. 1026 lincRNAs are identified as NONCODE, and 1489 lincRNAs as bona fide, among them 29 are classified as ME, and 1460 as SE. Differential Expression analysis was then performed on NONCODE and bona fide lincRNAs using multiple comparisons.