Leonhard Kohleick
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Leonhard Kohleick has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Claudia Langenberg
Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Berlin, Germany
3 shared publications - Maik Pietzner
Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Berlin, Germany
3 shared publications - Alice Williamson
Precision Healthcare University Research Institute, Queen Mary University of London, London, UK
2 shared publications - Benjamin Wild
Digital Health, Berlin Institute of Health at Charité - Universitätsmedizin Berlin, Germany
2 shared publications - Carl Beuchel
Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Berlin, Germany
2 shared publications - Martijn Zoodsma
Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Berlin, Germany
2 shared publications
Latest Publications
- Open Access
Domain-informed effector gene assignment exposes training-data dependence of drug target prioritisation from GWAS
About 90% of trait-associated lead variants lie outside protein-coding regions, so nominating the effector gene at genome-wide association study (GWAS) loci remains the principal bottleneck in translating associations into biological insight. Since no experimentally resolved ground truth exists at scale, every method rests on quasi-true-positive sets dominated by a single line of evidence. Here, we benchmarked established prioritisation methods across three independent evaluation sets, trained five models each informed by a different biological domain, and annotated lead variants for six common diseases. Performance did not transfer across domains (AUPRC 0.91 to 0.46 for the same method), and we traced this to how differently models weight the distance between variant and gene (p-123 ), not to how they were built. Curation assumptions propagated into predictions and can become self-confirming, such as druggability in the Open Targets-informed model. The five models shared only 17.1% to 31.8% of effector gene assignments, converging on established disease biology but diverging where GWAS are expected to yield novel insight. Recovery of 340 approved drug targets differed markedly between models, and no way of combining them, by averaging scores (38.8%) or training on all curations jointly (36.5%), approached the 52.1% recovered across all five together. Our workflow matched Open Targets L2G across three annotation sets without prior biological knowledge and exceeded it where curated knowledge already linked gene and disease. We provide the Domain-informed EFfector gene Assignment framework (DEFA; https://github.com/lkohleick/DEFA), which reports rather than hides the curation assumption behind each gene assignment so that it can be examined and tested.
Open AccessDomain-informed effector gene assignment exposes training-data dependence of drug target prioritisation from GWAS
About 90% of trait-associated lead variants lie outside protein-coding regions, so nominating the effector gene at genome-wide association study (GWAS) loci remains the principal bottleneck in translating associations into biological insight. Since no experimentally resolved ground truth exists at scale, every method rests on quasi-true-positive sets dominated by a single line of evidence. Here, we benchmarked established prioritisation methods across three independent evaluation sets, trained five models each informed by a different biological domain, and annotated lead variants for six common diseases. Performance did not transfer across domains (AUPRC 0.91 to 0.46 for the same method), and we traced this to how differently models weight the distance between variant and gene (p-123 ), not to how they were built. Curation assumptions propagated into predictions and can become self-confirming, such as druggability in the Open Targets-informed model. The five models shared only 17.1% to 31.8% of effector gene assignments, converging on established disease biology but diverging where GWAS are expected to yield novel insight. Recovery of 340 approved drug targets differed markedly between models, and no way of combining them, by averaging scores (38.8%) or training on all curations jointly (36.5%), approached the 52.1% recovered across all five together. Our workflow matched Open Targets L2G across three annotation sets without prior biological knowledge and exceeded it where curated knowledge already linked gene and disease. We provide the Domain-informed EFfector gene Assignment framework (DEFA; https://github.com/lkohleick/DEFA), which reports rather than hides the curation assumption behind each gene assignment so that it can be examined and tested.
- Open Access
An interpretable omnigenic neural network architecture for the human genome
Genetic prediction of complex phenotypes typically relies on additive linear models, which scale well but cannot capture non-additive effects or deeply integrate molecular and clinical data. Domain-specific neural networks have driven advances in images, text, and other modalities, but genome-scale neural networks remain challenging because genotypes are sparse and high-dimensional, effective sample sizes are limited, and generic architectures lack interpretability. Here, we introduce the omnigenic neural network, a biologically structured architecture inspired by the omnigenic model of complex traits. The model learns hierarchical representations of biological processes, accommodates multimodal inputs, supports transfer learning, and enables multitask prediction. Models trained in the UK Biobank and evaluated in the All of Us cohort for ischemic heart disease, type 2 diabetes, and schizophrenia outperformed published PGS Catalog and PRS-CSx scores. A multitask model trained across 36 cardiovascular endpoints further outperformed corresponding single-phenotype models and baselines. The architecture provides systems-level interpretability by quantifying the contributions of biological processes, which were consistent with established disease mechanisms. It also captures non-linear interactions between variants. Analysis of these interactions using Integrated Hessians revealed patterns concordant with previously reported epistatic associations. Together, these findings establish the omnigenic neural network as a flexible framework for interpretable, multimodal, and multitask genomic prediction.
Open AccessAn interpretable omnigenic neural network architecture for the human genome
Genetic prediction of complex phenotypes typically relies on additive linear models, which scale well but cannot capture non-additive effects or deeply integrate molecular and clinical data. Domain-specific neural networks have driven advances in images, text, and other modalities, but genome-scale neural networks remain challenging because genotypes are sparse and high-dimensional, effective sample sizes are limited, and generic architectures lack interpretability. Here, we introduce the omnigenic neural network, a biologically structured architecture inspired by the omnigenic model of complex traits. The model learns hierarchical representations of biological processes, accommodates multimodal inputs, supports transfer learning, and enables multitask prediction. Models trained in the UK Biobank and evaluated in the All of Us cohort for ischemic heart disease, type 2 diabetes, and schizophrenia outperformed published PGS Catalog and PRS-CSx scores. A multitask model trained across 36 cardiovascular endpoints further outperformed corresponding single-phenotype models and baselines. The architecture provides systems-level interpretability by quantifying the contributions of biological processes, which were consistent with established disease mechanisms. It also captures non-linear interactions between variants. Analysis of these interactions using Integrated Hessians revealed patterns concordant with previously reported epistatic associations. Together, these findings establish the omnigenic neural network as a flexible framework for interpretable, multimodal, and multitask genomic prediction.
- Open Access
- 24 July 2026
- Published by Springer Nature in Nature Genetics
- Vol. 58 (8) , 1866-1876
Nanoparticle-enriched mass spectrometry proteomics in British South Asians identifies links between genetic variants, plasma protein levels and disease risk
Understanding genetic variation associated with differences in plasma protein levels can elucidate human disease mechanisms. Here we demonstrate how untargeted nanoparticle-enriched mass spectrometry (MS)-based plasma proteomics delivers quantitatively and qualitatively different insights compared to two affinity-based assays in a sample of ~1,400 British South Asian individuals. We identify >1,200 significant locus–protein associations ( P < 8.7 × 10 −12 ; n = 895 cis -protein quantitative trait loci (pQTLs)), more than half of which have not been reported previously. Cross-platform comparison demonstrated that multiple platforms are required to capture the full spectrum of pQTLs of blood proteins. We combine proteogenomic results with evidence from multiple biological domains to suggest a potential role of 21 proteins in the pathology of 44 diseases, including a previously uncharacterized role of immunoglobulin λ variable 3-21 in the development of Graves’ disease. Our results demonstrate the potential of MS-based blood proteomics in non-European ancestries for pQTL discovery and the need to consolidate proteogenomic evidence to confidently assign proteins to disease pathology.
- 24 July 2026
- Published by Springer Nature in Nature Genetics
- Vol. 58 (8) , 1866-1876
Open AccessNanoparticle-enriched mass spectrometry proteomics in British South Asians identifies links between genetic variants, plasma protein levels and disease risk
Understanding genetic variation associated with differences in plasma protein levels can elucidate human disease mechanisms. Here we demonstrate how untargeted nanoparticle-enriched mass spectrometry (MS)-based plasma proteomics delivers quantitatively and qualitatively different insights compared to two affinity-based assays in a sample of ~1,400 British South Asian individuals. We identify >1,200 significant locus–protein associations ( P < 8.7 × 10 −12 ; n = 895 cis -protein quantitative trait loci (pQTLs)), more than half of which have not been reported previously. Cross-platform comparison demonstrated that multiple platforms are required to capture the full spectrum of pQTLs of blood proteins. We combine proteogenomic results with evidence from multiple biological domains to suggest a potential role of 21 proteins in the pathology of 44 diseases, including a previously uncharacterized role of immunoglobulin λ variable 3-21 in the development of Graves’ disease. Our results demonstrate the potential of MS-based blood proteomics in non-European ancestries for pQTL discovery and the need to consolidate proteogenomic evidence to confidently assign proteins to disease pathology.
- Open Access
- 3 October 2025
- Published by Springer Nature in Nature Genetics
- Vol. 57 (10) , 2445-2455
A genetic map of human metabolism across the allele frequency spectrum
Genetic studies of human metabolism have been limited in scale and allelic breadth. Here we provide a data-driven map of the genetic regulation of circulating small molecules and lipoprotein characteristics (249 traits) measured using proton nuclear magnetic resonance spectroscopy across the allele frequency spectrum in ~450,000 individuals. Trans-ancestral meta-analyses identify 29,824 locus–metabolite associations mapping to 753 regions with effects largely consistent between men and women and large ancestral groups represented in UK Biobank. We observe and classify extreme genetic pleiotropy, identify regulators of lipid metabolism, and assign effector genes at >100 loci through rare-to-common allelic series. We propose roles for genes less established in metabolic control (for example, SIDT2), genes characterized by phenotypic heterogeneity (for example, APOA1) and genes with specific disease relevance (for example, VEGFA). Our study demonstrates the value of broad, large-scale metabolomic phenotyping to identify and characterize regulators of human metabolism.
- 3 October 2025
- Published by Springer Nature in Nature Genetics
- Vol. 57 (10) , 2445-2455
Open AccessA genetic map of human metabolism across the allele frequency spectrum
Genetic studies of human metabolism have been limited in scale and allelic breadth. Here we provide a data-driven map of the genetic regulation of circulating small molecules and lipoprotein characteristics (249 traits) measured using proton nuclear magnetic resonance spectroscopy across the allele frequency spectrum in ~450,000 individuals. Trans-ancestral meta-analyses identify 29,824 locus–metabolite associations mapping to 753 regions with effects largely consistent between men and women and large ancestral groups represented in UK Biobank. We observe and classify extreme genetic pleiotropy, identify regulators of lipid metabolism, and assign effector genes at >100 loci through rare-to-common allelic series. We propose roles for genes less established in metabolic control (for example, SIDT2), genes characterized by phenotypic heterogeneity (for example, APOA1) and genes with specific disease relevance (for example, VEGFA). Our study demonstrates the value of broad, large-scale metabolomic phenotyping to identify and characterize regulators of human metabolism.