Publications
Research articles published in journals and conferences, and papers under review. Click the paper icon to access an article. Full list on Google Scholar.
Under review
Foundation model-guided cohort optimization enables robust and efficient biomedical discovery
Submitted, 2026
Sparse multi-landmark supervision for longitudinal EHR survival prediction
Submitted, 2026
Journal articles
In silico generation of gene expression profiles using diffusion models
BMC Bioinformatics, vol. 27, 2026
Abstract
RNA-seq data is used for precision medicine (e.g., cancer predictions), which benefits from deep learning approaches to analyze complex gene expression data. However, transcriptomics datasets often have few samples compared to deep learning standards. Synthetic data generation is thus being explored to address this data scarcity. So far, only deep generative models such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) have been used for this aim. Considering the recent success of diffusion models (DM) in image generation, we propose the first generation pipeline that leverages the power of said diffusion models. This paper presents two state-of-the-art diffusion models (DDPM and DDIM) and achieves their adaptation in the transcriptomics field. DM-generated data of L1000 landmark genes show better predictive performance over TCGA and GTEx datasets. We also compare linear and nonlinear reconstruction methods to recover the complete transcriptome. Results show that such reconstruction methods can boost the performances of diffusion models, as well as VAEs and GANs. Overall, the extensive comparison of various generative models using data quality indicators shows that diffusion models perform best and second-best, making them promising synthetic transcriptomics generators.
GAN-based data augmentation for transcriptomics: survey and comparative assessment
Bioinformatics, vol. 39, Supplement 1, pp. i111–i120 (ISMB 2023 proceedings), 2023
Abstract
Transcriptomics data are becoming more accessible due to high-throughput and less costly sequencing methods. However, data scarcity prevents exploiting deep learning models’ full predictive power for phenotypes prediction. Artificially enhancing the training sets, namely data augmentation, is suggested as a regularization strategy. Data augmentation corresponds to label-invariant transformations of the training set (e.g. geometric transformations on images and syntax parsing on text data). Such transformations are, unfortunately, unknown in the transcriptomic field. Therefore, deep generative models such as generative adversarial networks (GANs) have been proposed to generate additional samples. In this article, we analyze GAN-based data augmentation strategies with respect to performance indicators and the classification of cancer phenotypes.This work highlights a significant boost in binary and multiclass classification performances due to augmentation strategies. Without augmentation, training a classifier on only 50 RNA-seq samples yields an accuracy of, respectively, 94% and 70% for binary and tissue classification. In comparison, we achieved 98% and 94% of accuracy when adding 1000 augmented samples. Richer architectures and more expensive training of the GAN return better augmentation performances and generated data quality overall. Further analysis of the generated data shows that several performance indicators are needed to assess its quality correctly.All data used for this research are publicly available and comes from The Cancer Genome Atlas.
Deep transfer learning for blended source identification in galaxy survey data
Astronomy & Astrophysics, vol. 657, A98, 2022
Abstract
We present BLENDHUNTER, a proof-of-concept deep-transfer-learning-based approach for the automated and robust identification of blended sources in galaxy survey data. We take the VGG-16 network with pre-trained convolutional layers and train the fully connected layers on parametric models of COSMOS images. We test the efficacy of the transfer learning by taking the weights learned on the parametric models and using them to identify blends in more realistic Canada-France Imaging Survey (CFIS)-like images. We compare the performance of this method to SEP (a Python implementation of SEXTRACTOR) as a function of noise levels and the separation between sources. We find that BLENDHUNTER outperforms SEP by ∼15% in terms of classification accuracy for close blends (< 10 pixel separation between sources) regardless of the noise level used for training. Additionally, the method provides consistent results to SEP for distant blends (≥10 pixel separation between sources) provided the network is trained on data with noise that has a relatively close standard deviation to that of the target images. The code and data have been made publicly available to ensure the reproducibility of the results.
Conference papers
Frugal Generative Modeling for Tabular Data
ECML-PKDD 2024, Lecture Notes in Computer Science vol. 14948, pp. 55–72, 2024
Abstract
This paper presents a generative modeling approach called GMDA designed for tabular data, adapted to its arbitrary feature correlation structure. The generative model is trained so that sampled regions in the feature space contain the same fraction of true and synthetic samples, allowing true and synthetic data distributions to be aligned using a frugal and sound learning criterion. The merits of Gmda in terms of the usual performance indicators (pairwise correlation errors, precision, recall, predictive performance) are on par with or better than the state-of-the-art approaches for tabular data based on VAEs, GANs, or diffusion models. The key point is that it provides generative models with one or more orders of magnitude that are more frugal than baseline approaches.
Identifying GANs Blind Spots in Transcriptomic Data Generation
Junior Conference on Data Science and Engineering (JDSE), 2024
Abstract
Generative Adversarial Networks (GANs) show promise for data augmentation. However, they can overfit high-dimensional transcriptomic data, failing to capture biological diversity. This paper introduces a new metric that reveals discrepancy patterns between real and generated data, often overlooked by standard metrics. Our evaluation opens new perspectives for GAN-based data augmentation, balancing a priori the diverse modes of the real data and aligning a posteriori the generated data.
ML-CI: Machine Learning Confidence Intervals for Covid-19 forecasts
BayLearn – Machine Learning Symposium; also presented at JDSE 2021, 2021
Abstract
Epidemic forecasting has always been challenging and the recent Covid-19 outbreaks emphasizes it. We introduce a novel approach to address the problem of evaluating confidence intervals (CI) of time series prediction forecasts for compartmental models, using machine learning. We evaluate our approach using real data of the Covid pandemic on 27 countries. Compartmental models were trained taking into account non pharmaceutical governmental measures. A Random Forest regressor was trained, using various engineered features, to predict the forecasting error for various horizons on synthetic data, then applied to estimate CI on real data forecasts. Our method outperforms baselines using forecast likelihood as metric.