Data Science Seminars: Bioinformatics focus

Speaker: Michela Francesconi

October 27^th, 2026 | 2:30 pm
DEIB, PT2 Meeting Room (Bld. 20A)

For further information please contact:  Silvia Cascianelli |  silvia.cascianelli@polimi.it

Abstract

Tuesday, October 27^th, 2026 at 2:30 pm a new appointment of Data Science Seminars: Bioinformatics focus will take place in DEIB PT2 Meeting Room (Building 20A) organized by the Data Science for Bioinformatics group.

The seminar will be held by Michela Francesconi, PHD Student in Bioengineering, on the following subject: "Longitudinal Multi-Omic Biomarker Discovery for Immunotherapy Response in NSCLC".

The PEOPLE trial (NCT03447678) enrolled patients with PD-L1-low/negative NSCLC receiving first-line pembrolizumab and profiled them at baseline and after the first immunotherapy cycle. Four complementary domains were collected: circulating immune populations (flow cytometry), gut microbiome (shotgun metagenomics), CT radiomics, and clinical variables. The trial reflects the typical challenges of real-world oncology data: a small cohort, heterogeneous and partially missing modalities, imbalanced outcomes, and a patient biology that changes under treatment.
We built a modelling framework for this setting. Each modality goes through redundancy filtering and feature selection, followed by repeated nested cross-validation across several classifiers, with permutation testing to separate real signal from chance. The immune profile gave the most consistent predictive signal for response and progression-free survival. On-treatment sampling outperformed baseline, which highlights the value of longitudinal data. Clinical variables predicted early progression. The microbiome showed only weak and largely non-reproducible signals, and radiomics showed none.
To handle the high-dimensional, sparse microbiome, we used consensus network modules. We tested their robustness to preprocessing choices and confounders such as concomitant medications, and used them to probe cross-omic relationships. Finally, we are developing integration strategies that tolerate missing modalities, including cross-modality de-correlation and late fusion. Overall, the work argues for an honest view of what small, heterogeneous datasets can support, where robustness and negative results matter as much as predictive performance.