Amad Diouf Email & Phone Number

Paris, Île-De-France, France

Villejuif, Île-De-France, France

POC - Biomarker discovery in breast cancer for hard-to-treat cases including triple-negatives- Collected the study-related genomic datasets from public repositories, applied preprocessing including feature engineering, and selected an effective restriction strategy to isolate the sub-population of interest followed by a robust validation step. Impact : full data analysis process resulted in major expansion of the initially available samples pool, with more than half of features common to all populations, adding the compelling option of data integration further in the project. - Executed Machine Learning Engineering with model building focused on Regression, enacted Single task (Scikit-learn) and Multitask (SPAMS) learning strategies, designed a stringent score-based feature selection analysis rewarding stability and presence to optimize candidates list. Impact : initial set of candidates features reduced to a minimum, with a quarter of the reduction on the account of the feature selection analysis.- Realized collection of all usable future-proof data (RNAseq), guaranteeing continuity of the project after the proof of concept (microarrays). Impact : reporting to stakeholders listed a pool in which complete cohorts samples represented the third, and for incomplete populations was attached the full contact list for missing information requests.Key notes - Entire cycle of data manipulation, from collection to insights.- Regression : Single task (Scikit-learn) and Multitask (SPAMS).- Feature selections analysis.- Reporting to stakeholders.- Data types analysed - Field : GEX (microarrays) - Oncology.- Paving the way to future developments.

Région De Marseille, France

Classification benchmarking on high dimensional omics data.- Accomplished a concise selection of classification methods geared towards interpretable predictive models, while including for performances comparison non-interpretable methods with high interest in the state-of-the-art practices.Impact : 5 algorithms selected : Random Forest, XGBoost, linear kernel SVM, rbf kernel SVM, Deep Neural Network.- Evaluated predictive models on performance metrics and feature selection qualities (size for conciseness and list for future analysis possibilities).Impact : Rankings showcased the decision trees models dominating with major performances, pushing out short feature selections, while having less computational cost than the Deep Learning method tested.- Designed and developed the pipeline SLATE for automated evaluation of models performance in supervised learning. SLATE incorporates 4 capabilities (data management during analysis, model evaluation through metrics, feature selections evaluation through size and list, automation with a one-line launcher and consolidated results delivery), supports the 5 classification algorithms benchmarked while offering an ease to include additional methods in the future. Impact : SLATE offers a reduction of up to 75% in time of preparation for classification models evaluation, with built-in quality of life aspects to the user.Key notes- Full focus on model building in Machine Learning Engineering.- Classification : Random Forest, XGBoost, Deep Neural Network, SVM (linear kernel), SVM (rbf kernel).- Benchmarking : Ranked the 5 algorithms on multiple aspects.- Data types analysed - Field : CNV, GEX, SNV, METH - Oncology.- Paving the way to future developments.

Master’S Degree (Msc.), Bioinformatics, Knowledge And Data

Engineering Degree, Biological/Biosystems Engineering