Biology subjects

Curion, F.

Publications and source records attributed to Curion, F..

6 recordsLinked to original sources

bioRxiv2026

Evo 2 as a classification machine: evidence from in-context learning and mechanistic interpretability

In-context learning (iCL) is an emergent capability of Large Language Models (LLMs), allowing them to perform new tasks at inference time using prompt-injected examples. While extensively studied in LLMs, the boundaries of its capabilities and underlying mechanisms remain poorly characterised in genomic Language Models (gLMs). Here, we map the operating regime of Evo 2, a nucleotide-level foundation gLM, across five binary classification tasks spanning biological and artificial sequences. We observe robust iCL on shorter natural sequences (F1=0.902 for miRNA, 0.785 for Toxins), degrading with sequence length, and collapsing at kilobase scale. Strikingly, we find no benefit in model scaling, as the 7B model systematically outperforms the 40B variant. We further find that perplexity, a widely used gLM performance proxy, poorly predicts accuracy. Mechanistic interpretability reconciles these observations: the logit-lens profiling suggests a prediction-generalisation trade-off, while Jacobian Scope indicates models might track prompts' structure rather than the signal-carrying content.

genomics↗

bioRxiv2024

Multimodal weakly supervised learning to identify disease-specific changes in single-cell atlases

To deliver clinically relevant insights from large patient cohorts profiled with single-cell technologies, a key challenge is to relate sample-level and single-cell measurements. We present MultiMIL, a deep learning framework that applies attention-based multiple-instance learning for phenotype prediction and cell state identification. We applied MultiMIL to peripheral blood mononuclear cells from COVID-19 patients, the Human Lung Cell Atlas, and a spatial proteomics breast cancer dataset, demonstrating how our model can be utilized to find phenotype-associated cell states, learn phenotype-informed sample representations, and expand disease signatures.

bioinformatics↗

bioRxiv2023

Identification of cell types, states and programs by learning gene set representations

As single cell molecular data expand, there is an increasing need for algorithms that efficiently query and prioritize gene programs, cell types and states in single-cell sequencing data, particularly in cell atlases. Here we present scDECAF, a statistical learning algorithm to identify cell types, states and programs in single-cell gene expression data using vector representation of gene sets, which improves biological interpretation by selecting a subset of most biologically relevant programs. We applied scDECAF to scRNAseq data from PBMC, Lung, Pancreas, Brain and slide-tags snRNA of human prefrontal cortex for automatic cell type annotation. We demonstrate that scDECAF can recover perturbed gene programs in Lupus PBMC cells stimulated with IFNbeta and TGFBeta-induced cells undergoing epithelial-to-mesenchymal transition. scDECAF delineates patient-specific heterogeneity in cellular programs in Ovarian Cancer data. Using a healthy PBMC reference, we apply scDECAF to a mapped query PBMC COVID-19 case-control dataset and identify multicellular programs associated with severe COVID-19. scDECAF can improve biological interpretation and complement reference mapping analysis, and provides a method for gene set and pathway analysis in single cell gene expression data.

bioinformatics↗

bioRxiv2023

hadge: a comprehensive pipeline for donor deconvolution in single cell

Single cell multiplexing techniques (cell hashing and genetic multiplexing) allow to combine multiple samples, thereby optimizing sample processing and reducing batch effects. Cell hashing conjugates antibody-tags or chemical-oligonucleotides to cell membranes, while genetic multiplexing allows to mix genetically diverse samples and relies on aggregation of RNA reads at known genomic coordinates. We developed hadge (hashing deconvolution combined with genotype information), a Nextflow pipeline that combines 12 methods to perform both hashing- and genotype-based deconvolution. We propose a joint deconvolution strategy combining the best performing methods and we demonstrate how this approach leads to recovery of previously discarded cells in a nuclei hashing of fresh-frozen brain tissue.

bioinformatics↗

bioRxiv2023

Panpipes: a pipeline for multiomic single-cell data analysis.

Single-cell multiomic analysis of the epigenome, transcriptome and proteome allows for comprehensive characterisation of the molecular circuitry that underpins cell identity and state. However, the holistic interpretation of such datasets presents a challenge given a paucity of approaches for systematic, joint evaluation of different modalities. Here, we present Panpipes, a set of computational workflows designed to automate multimodal single-cell and spatial transcriptomic analyses by incorporating widely-used Python-based tools to perform quality control, preprocessing, integration, clustering, and reference mapping at scale. Panpipes allows reliable and customisable analysis and evaluation of individual and integrated modalities, thereby empowering decision-making before downstream investigations.

bioinformatics↗

bioRxiv2020

Single-cell sequencing of human iPSC-derived cerebellar organoids shows recapitulation of cerebellar development

Current protocols for producing cerebellar neurons from human pluripotent stem cells (hPSCs) are reliant on animal co-culture and mostly exist as monolayers, which have limited capability to recapitulate the complex arrangement of the brain. We developed a method to differentiate hPSCs into cerebellar organoids that display hallmarks of in vivo cerebellar development. Single-cell profiling followed by comparison to an atlas of the developing murine cerebellum revealed transcriptionally-discrete populations encompassing all major cerebellar cell types. Matrigel encapsulation altered organoid growth dynamics, resulting in differential regulation of cell cycle, migration and cell-death pathways. However, this was at the expense of reproducibility. Furthermore, we showed the contribution of basement membrane signalling to both cellular composition of the organoids and developmentally-relevant gene expression programmes. This model system has exciting implications for studying cerebellar development and disease most notably by providing xeno-free conditions, representing a more biologically relevant and therapeutically tractable culture setting.

neuroscience↗