Biology subjects

Pielies Avelli, M.

Publications and source records attributed to Pielies Avelli, M..

5 recordsLinked to original sources

bioRxiv2026

SLIM: A small linear model with STRING embeddings for single-cell genetic perturbation prediction

Predicting cellular responses to genetic perturbations is central to understanding gene function and prioritizing therapeutic targets, but experimental screens cannot exhaustively cover genes, cell types, and perturbation combinations. Recent benchmarks have shown that simple baselines can match or outperform substantially more complex models, suggesting that informative biological priors may be as important as model capacity. Here we present SLIM, a lightweight extension of the bilinear model of Ahlmann-Eltze et al. SLIM represents perturbations with 64-dimensional embeddings derived from the STRING protein network and predicts mean transcriptional responses through a closed-form ridge-regression estimator. It then constructs single-cell populations by retrieving training cells and rescaling each gene to match the predicted mean. We evaluated SLIM against four deep learning models and two simple baselines on four single-gene perturbation datasets and one combinatorial perturbation dataset. Across these within-dataset benchmarks, SLIM achieved competitive mean-response accuracy, ranked first in eight of twelve single-gene dataset-metric comparisons, and produced substantially lower maximum mean discrepancy values than the evaluated alternatives. The model has 640 trainable parameters and fitted each benchmark dataset in under 10 seconds on a CPU. These results show that compact biological representations can support accurate and computationally efficient perturbation prediction. Code is available at https://github.com/RasmussenLab/SLIM. Key PointsO_LISLIM combines a closed-form bilinear predictor with STRING-derived perturbation embeddings. C_LIO_LIAcross five within-dataset benchmarks, SLIM achieved competitive mean-response prediction with only 640 trainable parameters. C_LIO_LISLIM builds cell populations by rescaling retrieved training cells to the predicted mean, so they inherit realistic cell-to-cell variation and gene-gene covariation. C_LIO_LIThe results highlight the importance of perturbation representations and population-construction procedures in low-data benchmarks. C_LIO_LISLIM fits each benchmark dataset in under 10 seconds on a standard CPU. C_LI

bioinformatics↗

bioRxiv2025

Unsupervised learning of multi-omics data enables disease risk prediction in the UK Biobank

The size and complexity of biomedical datasets continue to grow, driving the development of methods that reduce dimensionality while preserving biological signals. Yet, when deep learning is applied to such data, the impact of preprocessing choices and dataset properties on model behavior is often overlooked. Here, we applied our framework Multi-Omics Variational autoEncoder (MOVE) to multiomics data from 452,026 UK Biobank participants, aiming to both evaluate the power of the learned representations for disease risk prediction and critically analyze how non-biological factors, like dataset properties and preprocessing decisions, can shape and influence the results. We show that reducing the dimensionality of the data by a factor of 80 still yields comparable prediction performance across 15 different diseases. We further demonstrate how dataset properties and preprocessing choices impact the model performance, latent representation and downstream results, and our findings strongly underline the need for thorough analysis and understanding of a models behavior before drawing conclusions from its results.

bioinformatics↗

bioRxiv2025

On the use of variational autoencoders for biomedical data integration

Variational Autoencoders (VAEs) are a widely used framework to integrate diverse biomedical data modalities, create representations that capture the underlying structure of the datasets, and obtain insights about the relations between variables. Here we describe how this is achieved from an empirical point of view in our previously developed VAE-based framework MOVE, providing an intuitive perspective on the inner workings of multimodal VAEs in biomedical contexts. We explore how the models emerging dynamics shape their performance and how in silico perturbations can be leveraged to identify potential associations between variables. To do that, we extend our framework to handle perturbations of continuous variables, introduce a new approach to better capture associations between them, and create synthetic datasets to benchmark the proposed methods against well-defined ground truth associations. We finally showcase our findings in real biomedical scenarios, namely a multimodal dataset of inflammatory bowel disease and a dataset containing genetic knockdowns in K562 and RPE1 cells.

bioinformatics↗

bioRxiv2025

Accurate plasmid reconstruction from metagenomics data using assembly-alignment graphs and contrastive learning

Plasmids are extrachromosomal DNA molecules that enable horizontal gene transfer in bacteria, often conferring advantages such as antibiotic resistance. Despite their significance, plasmids are underrepresented in genomic databases due to challenges in assembling them, caused by mosaicism and micro-diversity. Current plasmid assemblers rely on detecting circular paths in single-sample assembly graphs, but face limitations due to graph fragmentation and entanglement, and low coverage. We introduce PlasMAAG (Plasmid and organism Metagenomic binning using Assembly Alignment Graphs), a framework to recover plasmids and organisms from metagenomic samples that leverages an approach that we call "assembly-alignment graphs" alongside common binning features. On synthetic benchmark datasets, PlasMAAG reconstructed 50-121% more near-complete plasmids than competing methods and improved the Matthews Correlation Coefficient of geNomad contig classification by 28-106%. On hospital sewage samples, PlasMAAG outperformed all other methods, reconstructing 33% more plasmid sequences. PlasMAAG enables the study of organism-plasmid associations and intra-plasmid diversity across samples, offering state-of-the-art plasmid reconstruction with reduced computational costs.

bioinformatics↗

bioRxiv2024

Modeling nascent transcription from chromatin landscape and structure

BackgroundDifferent cell types and their associated functionalities emerge from a single genomic sequence when certain regions are expressed while others remain silenced. Modeling gene expression and its potential malfunctioning in different cellular contexts is hence pivotal to understand both development and disease. ResultsWe present the Chromatin Landscape and Structure to Expression Regressor (CLASTER), an epigenetic-based deep neural network that can integrate different data modalities describing the chromatin landscape and its 3D structure. CLASTER effectively translates them into nascent transcription levels measured by EU-seq at a kilobasepair resolution. Our predictions reached a Pearson correlation with targets above r=0.86 at both bin and gene levels, without relying on DNA sequence nor explicitly extracted chromatin features. The model mostly used the information found within 10 kbp of the predicted locus, even when a wide genomic region of 1 Mbp was available. Explicit modeling of long-range interactions using multi-headed attention and high-resolution chromatin contact maps had little impact on model performance, despite the model correctly identifying elements in these inputs influencing nascent transcription. The trained model served then as a platform to predict the transcriptional impact of simulated epigenetic silencing perturbations. ConclusionsOur results point towards a rather local, integrative and combinatorial paradigm of gene regulation, where changes in the chromatin environment surrounding a gene shape its context-specific transcription. We conclude that the predominant locality and limitations of current machine learning approaches might emerge as a genuine signature of genomic organization, having broad implications for future modeling approaches.

bioinformatics↗