Harri J Lähdesmäki

LGTM couples a neural topic model with additive basis-function Gaussian processes, learning microbial co-abundance topics and how covariates relate to their trajectories in longitudinal gut microbiome data.

This paper improves patient phenotype prediction using scRNA-seq data by incorporating hierarchical cell type information into multiple instance learning, achieving better performance and interpretability.

DeconV is a probabilistic cell-type deconvolution method for bulk RNA-Seq data using scRNA-Seq references, offering comparable accuracy to Scaden with added interpretability through confidence intervals.

TSignal is a transformer model that predicts signal peptides and their cleavage sites, showing improved accuracy across various organisms and peptide types.

LuxHMM segments the genome into regions and infers differential methylation using a hidden Markov model and Bayesian regression, showing competitive performance on real and simulated data.

ChromDMM clusters heterogeneous epigenetic data by improving profile shifting, strand orientation, and smoothness, outperforming existing methods on simulated and ENCODE data.

TCRconv predicts T-cell receptor-epitope recognition using a deep learning model and provides insights into T-cell dynamics in COVID-19 patients.

The paper evaluates statistical and probabilistic methods for classifying cancer types using cell-free DNA methylation data, finding that some methods outperform GLMnet and show varying performance across sequencing depths and cancer types.

The paper evaluates the type 1 error rate in bisulfite sequencing analysis of umbilical cord blood samples and finds that permutation-based significance analysis reduces it, with minimal differential methylation associated with clinical variables except sex.

PairGP models longitudinal gene expression data from paired multi-condition studies to identify condition groups with different dynamics.

This paper introduces a scalable method using Variational Autoencoders to estimate microbial co-occurrence networks from high-dimensional microbiome data.

The paper introduces a probabilistic model for enhancer prediction in the human genome using chromatin feature patterns, showing improved performance over existing methods.

This paper introduces a personalized method for analyzing time-course gene expression data in heterogeneous diseases, showing it can identify disease-relevant pathways more effectively than non-personalized approaches.

LuxUS improves differential DNA methylation analysis by incorporating spatial correlation and mixed models, enhancing statistical power and detecting biologically significant cytosines.

LuxHS improves DNA methylation analysis by incorporating a sparsity promoting prior to account for deviations from spatial correlation patterns in BS-seq data.

The paper develops a data-driven spatial model to study yeast colony growth dynamics, showing that metabolic coupling is central to colony formation.

The paper identifies DNA methylation differences in peripheral blood of Alzheimer’s disease discordant twin pairs, linking these to brain gene expression and neuronal functions.

LuxRep improves methylation level and differentially methylated site estimation by incorporating low bisulfite conversion rate technical replicates in BS-seq analysis.

SCHiRM detects gene dependencies in scRNA-seq data using a hierarchical regression model, outperforming standard techniques in parameter inference and variable selection.

LonGP is a Gaussian process regression model for longitudinal studies that handles time-varying covariates and non-linear effects, showing performance on simulated and real -omics data.