Thomas Hillerton
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Thomas Hillerton has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Erik L. L. Sonnhammer
Science for Life Laboratory, Department of Biochemistry and Biophysics, Stockholm University, Solna, Sweden
4 shared publications - David Martínez-Enguita
Linköping University
2 shared publications - Deniz Seçilmiş
Department of Biochemistry and Biophysics, Stockholm University, Science for Life Laboratory , Box 1031, 17121 Solna , Sweden
2 shared publications - Julia Åkesson
Linköpings universitet, Bioinformatik
2 shared publications - Mateusz Garbulowski
Department of Biochemistry and Biophysics, Stockholm University, Science for Life Laboratory , Solna 17121
2 shared publications - Mika Gustafsson
Bioinformatics, Department of Physics, Chemistry and Biology, Linköping University, Linköping, Sweden
2 shared publications
Latest Publications
- Open Access
System-level health profiling from blood DNA methylation with explainable deep learning
Genome-scale DNA methylation (DNAm) profiles capture organismal physiology, but most predictive models lack transparency and multi-level applicability. Here we develop an explainable framework that quantifies respiratory, cardiovascular, and metabolic status as bounded health scores (0–1) derived from sex-specific clinical reference ranges and disease penalties, and then predicts these scores from whole-blood DNAm. Using Generation Scotland and case-control samples (n = 14,496 individuals), we screened 39 covariates for disease relevance and DNAm predictability, yielding system- relevant panels that were aggregated into scores. We compressed DNAm profiles with a protein-interaction-guided autoencoder, and trained health predictors on 128- dimensional embeddings using fully connected networks. On held-out samples, models reproduced the composite scores with strong rank agreement (Spearman ρ = 0.87, R 2 = 0.71 for respiratory health; ρ = 0.82, R 2 = 0.66 for cardiovascular; ρ = 0.81, R 2 = 0.64 for metabolic) and recover expected population structure in a generally healthy cohort, with clear separation between “single-system low” and “multi-system low” phenotypes, and graded coupling across systems without redundancy. Further, the top features retrieved from each explainable predictor aligned with system biology: airway epithelial repair, hypoxia and inflammatory trafficking for respiratory; endothelial remodeling and cardiomyocyte programs for cardiovascular; glucose-lipid metabolism and metaflammation for metabolic. These results show that DNAm embeddings can yield accurate, transparent, and system-aware health profiling from blood, providing actionable summaries while revealing the molecular processes the models use to infer multi-system status. This approach positions DNAm embeddings plus interpretable penalty targets as a practical bridge from epigenomic signal to system-level triage and is extensible for evaluation in larger, more diverse cohorts.
Open AccessSystem-level health profiling from blood DNA methylation with explainable deep learning
Genome-scale DNA methylation (DNAm) profiles capture organismal physiology, but most predictive models lack transparency and multi-level applicability. Here we develop an explainable framework that quantifies respiratory, cardiovascular, and metabolic status as bounded health scores (0–1) derived from sex-specific clinical reference ranges and disease penalties, and then predicts these scores from whole-blood DNAm. Using Generation Scotland and case-control samples (n = 14,496 individuals), we screened 39 covariates for disease relevance and DNAm predictability, yielding system- relevant panels that were aggregated into scores. We compressed DNAm profiles with a protein-interaction-guided autoencoder, and trained health predictors on 128- dimensional embeddings using fully connected networks. On held-out samples, models reproduced the composite scores with strong rank agreement (Spearman ρ = 0.87, R 2 = 0.71 for respiratory health; ρ = 0.82, R 2 = 0.66 for cardiovascular; ρ = 0.81, R 2 = 0.64 for metabolic) and recover expected population structure in a generally healthy cohort, with clear separation between “single-system low” and “multi-system low” phenotypes, and graded coupling across systems without redundancy. Further, the top features retrieved from each explainable predictor aligned with system biology: airway epithelial repair, hypoxia and inflammatory trafficking for respiratory; endothelial remodeling and cardiomyocyte programs for cardiovascular; glucose-lipid metabolism and metaflammation for metabolic. These results show that DNAm embeddings can yield accurate, transparent, and system-aware health profiling from blood, providing actionable summaries while revealing the molecular processes the models use to infer multi-system status. This approach positions DNAm embeddings plus interpretable penalty targets as a practical bridge from epigenomic signal to system-level triage and is extensible for evaluation in larger, more diverse cohorts.
- Open Access
Protein-protein interaction priors shape biologically coherent latent spaces for causally concordant cross-omic translation
Deep learning models routinely compress omics into low-dimensional codes, yet many equally accurate embeddings fail to reflect how cells are wired, which limits explanation and causal reasoning. We present a simple, architecture-agnostic approach to make latent spaces biologically legible: a protein-protein interaction (PPI) prior that softly steers autoencoder units to recruit genes that are proximal on the interactome while discouraging redundant reuse. Applied to large DNA methylation (∼155k) and RNAseq (∼993k) compendia and to knowledge-driven (STRING), structure-predicted (RoseTTAFold2-PPI), and union interactomes, this objective reorganizes methylation latents into compact, non-overlapping network neighborhoods without sacrificing reconstruction accuracy. The resulting units map cleanly onto biological processes such as cell-cycle control, immune signaling, proteostasis, mitochondrial metabolism, and RNA handling, with a limited, hub-enriched overlap that plausibly bridges modules. We then asked whether this structured geometry transfers downstream. Using paired TCGA cohorts spanning 23 cancers, omic translators trained on these embeddings, especially a shared-latent bidirectional model, outperformed full-matrix baselines in biologically concordant directions (methylation to transcription, genomics to methylation and transcription) and, crucially, inherited the mechanistic imprint of the upstream encoder. Analytical sensitivity mapping showed that translators fed PPI-guided embeddings preferentially learned known cancer drivers and enriched hallmark pathways, whereas accuracy-matched models trained on non-constrained embeddings did not. Thus, the prior not only regularizes but passes forward a functional coordinate system that makes subsequent predictors mechanistically aware. By keeping biology in the loss rather than hard-wiring it into the network, our approach scales to very large cohorts, preserves flexibility for understudied genes, and yields latents that are both performant and interpretable. More broadly, it outlines a practical route to mechanism-anchored representation learning that propagates explanatory structure into downstream tasks, advancing explainable AI for multi-omic analysis and clinical decision support.
Open AccessProtein-protein interaction priors shape biologically coherent latent spaces for causally concordant cross-omic translation
Deep learning models routinely compress omics into low-dimensional codes, yet many equally accurate embeddings fail to reflect how cells are wired, which limits explanation and causal reasoning. We present a simple, architecture-agnostic approach to make latent spaces biologically legible: a protein-protein interaction (PPI) prior that softly steers autoencoder units to recruit genes that are proximal on the interactome while discouraging redundant reuse. Applied to large DNA methylation (∼155k) and RNAseq (∼993k) compendia and to knowledge-driven (STRING), structure-predicted (RoseTTAFold2-PPI), and union interactomes, this objective reorganizes methylation latents into compact, non-overlapping network neighborhoods without sacrificing reconstruction accuracy. The resulting units map cleanly onto biological processes such as cell-cycle control, immune signaling, proteostasis, mitochondrial metabolism, and RNA handling, with a limited, hub-enriched overlap that plausibly bridges modules. We then asked whether this structured geometry transfers downstream. Using paired TCGA cohorts spanning 23 cancers, omic translators trained on these embeddings, especially a shared-latent bidirectional model, outperformed full-matrix baselines in biologically concordant directions (methylation to transcription, genomics to methylation and transcription) and, crucially, inherited the mechanistic imprint of the upstream encoder. Analytical sensitivity mapping showed that translators fed PPI-guided embeddings preferentially learned known cancer drivers and enriched hallmark pathways, whereas accuracy-matched models trained on non-constrained embeddings did not. Thus, the prior not only regularizes but passes forward a functional coordinate system that makes subsequent predictors mechanistically aware. By keeping biology in the loss rather than hard-wiring it into the network, our approach scales to very large cohorts, preserves flexibility for understudied genes, and yields latents that are both performant and interpretable. More broadly, it outlines a practical route to mechanism-anchored representation learning that propagates explanatory structure into downstream tasks, advancing explainable AI for multi-omic analysis and clinical decision support.
- Open Access
- 24 March 2025
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 41 (5)
Topology-based metrics for finding the optimal sparsity in gene regulatory network inference
Motivation: Gene regulatory network (GRN) inference is a complex task aiming to unravel regulatory interactions between genes in a cell. A major shortcoming of most GRN inference methods is that they do not attempt to find the optimal sparsity, i.e. the single best GRN, which is important when applying GRN inference in a real situation. Instead, the sparsity tends to be controlled by an arbitrarily set hyperparameter. Results: In this paper, two new methods for predicting the optimal sparsity of GRNs are formulated and benchmarked on simulated perturbation-based gene expression data using four GRN inference methods: LASSO, Zscore, LSCON, and GENIE3. Both sparsity prediction methods are defined using the hypothesis that the topology of real GRNs is scale-free, and are evaluated based on their ability to predict the sparsity of the true GRN. The results show that the new topology-based approaches reliably predict a sparsity close to the true one. This ability is valuable for real-world applications where a single GRN is inferred from real data. In such situations, it is vital to be able to infer a GRN with the correct sparsity. Availability and implementation: https://bitbucket.org/sonnhammergrni/powerlaw_sparsity/ and https://codeocean.com/capsule/4393635/.
- 24 March 2025
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 41 (5)
Open AccessTopology-based metrics for finding the optimal sparsity in gene regulatory network inference
Motivation: Gene regulatory network (GRN) inference is a complex task aiming to unravel regulatory interactions between genes in a cell. A major shortcoming of most GRN inference methods is that they do not attempt to find the optimal sparsity, i.e. the single best GRN, which is important when applying GRN inference in a real situation. Instead, the sparsity tends to be controlled by an arbitrarily set hyperparameter. Results: In this paper, two new methods for predicting the optimal sparsity of GRNs are formulated and benchmarked on simulated perturbation-based gene expression data using four GRN inference methods: LASSO, Zscore, LSCON, and GENIE3. Both sparsity prediction methods are defined using the hypothesis that the topology of real GRNs is scale-free, and are evaluated based on their ability to predict the sparsity of the true GRN. The results show that the new topology-based approaches reliably predict a sparsity close to the true one. This ability is valuable for real-world applications where a single GRN is inferred from real data. In such situations, it is vital to be able to infer a GRN with the correct sparsity. Availability and implementation: https://bitbucket.org/sonnhammergrni/powerlaw_sparsity/ and https://codeocean.com/capsule/4393635/.
- Open Access
- 2 July 2024
- Published by Oxford University Press (OUP) in NAR Genomics and Bioinformatics
- Vol. 6 (3) , lqae121
GeneSPIDER2: large scale GRN simulation and benchmarking with perturbed single-cell data
Single-cell data is increasingly used for gene regulatory network (GRN) inference, and benchmarks for this have been developed based on simulated data. However, existing single-cell simulators cannot model the effects of gene perturbations. A further challenge lies in generating large-scale GRNs that often struggle with computational and stability issues. We present GeneSPIDER2, an update of the GeneSPIDER MATLAB toolbox for GRN benchmarking, inference, and analysis. Several software modules have improved capabilities and performance, and new functionalities have been added. A major improvement is the ability to generate large GRNs with biologically realistic topological properties in terms of scale-free degree distribution and modularity. Another major addition is a simulation of single-cell data, which is becoming increasingly popular as input for GRN inference. Specifically, we introduced the unique feature to generate single-cell data based on genetic perturbations. Finally, the simulated single-cell data was compared to real single-cell Perturb-seq data from two cell lines, showing that the synthetic and real data exhibit similar properties.
- 2 July 2024
- Published by Oxford University Press (OUP) in NAR Genomics and Bioinformatics
- Vol. 6 (3) , lqae121
Open AccessGeneSPIDER2: large scale GRN simulation and benchmarking with perturbed single-cell data
Single-cell data is increasingly used for gene regulatory network (GRN) inference, and benchmarks for this have been developed based on simulated data. However, existing single-cell simulators cannot model the effects of gene perturbations. A further challenge lies in generating large-scale GRNs that often struggle with computational and stability issues. We present GeneSPIDER2, an update of the GeneSPIDER MATLAB toolbox for GRN benchmarking, inference, and analysis. Several software modules have improved capabilities and performance, and new functionalities have been added. A major improvement is the ability to generate large GRNs with biologically realistic topological properties in terms of scale-free degree distribution and modularity. Another major addition is a simulation of single-cell data, which is becoming increasingly popular as input for GRN inference. Specifically, we introduced the unique feature to generate single-cell data based on genetic perturbations. Finally, the simulated single-cell data was compared to real single-cell Perturb-seq data from two cell lines, showing that the synthetic and real data exhibit similar properties.
- Open Access
- 24 May 2022
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 50 (W1) , W398-W404
GRNbenchmark - a web server for benchmarking directed gene regulatory network inference methods
Accurate inference of gene regulatory networks (GRN) is an essential component of systems biology, and there is a constant development of new inference methods. The most common approach to assess accuracy for publications is to benchmark the new method against a selection of existing algorithms. This often leads to a very limited comparison, potentially biasing the results, which may stem from tuning the benchmark's properties or incorrect application of other methods. These issues can be avoided by a web server with a broad range of data properties and inference algorithms, that makes it easy to perform comprehensive benchmarking of new methods, and provides a more objective assessment. Here we present https://GRNbenchmark.org/ - a new web server for benchmarking GRN inference methods, which provides the user with a set of benchmarks with several datasets, each spanning a range of properties including multiple noise levels. As soon as the web server has performed the benchmarking, the accuracy results are made privately available to the user via interactive summary plots and underlying curves. The user can then download these results for any purpose, and decide whether or not to make them public to share with the community.
- 24 May 2022
- Published by Oxford University Press (OUP) in Nucleic Acids Research
- Vol. 50 (W1) , W398-W404
Open AccessGRNbenchmark - a web server for benchmarking directed gene regulatory network inference methods
Accurate inference of gene regulatory networks (GRN) is an essential component of systems biology, and there is a constant development of new inference methods. The most common approach to assess accuracy for publications is to benchmark the new method against a selection of existing algorithms. This often leads to a very limited comparison, potentially biasing the results, which may stem from tuning the benchmark's properties or incorrect application of other methods. These issues can be avoided by a web server with a broad range of data properties and inference algorithms, that makes it easy to perform comprehensive benchmarking of new methods, and provides a more objective assessment. Here we present https://GRNbenchmark.org/ - a new web server for benchmarking GRN inference methods, which provides the user with a set of benchmarks with several datasets, each spanning a range of properties including multiple noise levels. As soon as the web server has performed the benchmarking, the accuracy results are made privately available to the user via interactive summary plots and underlying curves. The user can then download these results for any purpose, and decide whether or not to make them public to share with the community.
- Open Access
- 17 February 2022
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 38 (8) , 2263-2268
Fast and accurate gene regulatory network inference by normalized least squares regression
Inferring an accurate gene regulatory network (GRN) has long been a key goal in the field of systems biology. To do this, it is important to find a suitable balance between the maximum number of true positive and the minimum number of false-positive interactions. Another key feature is that the inference method can handle the large size of modern experimental data, meaning the method needs to be both fast and accurate. The Least Squares Cut-Off (LSCO) method can fulfill both these criteria, however as it is based on least squares it is vulnerable to known issues of amplifying extreme values, small or large. In GRN this manifests itself with genes that are erroneously hyper-connected to a large fraction of all genes due to extremely low value fold changes. We developed a GRN inference method called Least Squares Cut-Off with Normalization (LSCON) that tackles this problem. LSCON extends the LSCO algorithm by regularization to avoid hyper-connected genes and thereby reduce false positives. The regularization used is based on normalization, which removes effects of extreme values on the fit. We benchmarked LSCON and compared it to Genie3, LASSO, LSCO and Ridge regression, in terms of accuracy, speed and tendency to predict hyper-connected genes. The results show that LSCON achieves better or equal accuracy compared to LASSO, the best existing method, especially for data with extreme values. Thanks to the speed of least squares regression, LSCON does this an order of magnitude faster than LASSO. Data: https://bitbucket.org/sonnhammergrni/lscon; Code: https://bitbucket.org/sonnhammergrni/genespider. Supplementary data are available at Bioinformatics online.
- 17 February 2022
- Published by Oxford University Press (OUP) in Bioinformatics
- Vol. 38 (8) , 2263-2268
Open AccessFast and accurate gene regulatory network inference by normalized least squares regression
Inferring an accurate gene regulatory network (GRN) has long been a key goal in the field of systems biology. To do this, it is important to find a suitable balance between the maximum number of true positive and the minimum number of false-positive interactions. Another key feature is that the inference method can handle the large size of modern experimental data, meaning the method needs to be both fast and accurate. The Least Squares Cut-Off (LSCO) method can fulfill both these criteria, however as it is based on least squares it is vulnerable to known issues of amplifying extreme values, small or large. In GRN this manifests itself with genes that are erroneously hyper-connected to a large fraction of all genes due to extremely low value fold changes. We developed a GRN inference method called Least Squares Cut-Off with Normalization (LSCON) that tackles this problem. LSCON extends the LSCO algorithm by regularization to avoid hyper-connected genes and thereby reduce false positives. The regularization used is based on normalization, which removes effects of extreme values on the fit. We benchmarked LSCON and compared it to Genie3, LASSO, LSCO and Ridge regression, in terms of accuracy, speed and tendency to predict hyper-connected genes. The results show that LSCON achieves better or equal accuracy compared to LASSO, the best existing method, especially for data with extreme values. Thanks to the speed of least squares regression, LSCON does this an order of magnitude faster than LASSO. Data: https://bitbucket.org/sonnhammergrni/lscon; Code: https://bitbucket.org/sonnhammergrni/genespider. Supplementary data are available at Bioinformatics online.