This pages lists some achievements for Dylan Lawless across projects and collaborations under the STAR framework. For a curated collection of topics see the profile pages.
Scientific excellence (19)
Population-scale GWAS and epidemiological statistics — 5 years EPFL postdoc
Situation: Understanding genetic susceptibility to severe infection required integrating host genome, pathogen genome, and clinical outcome data across large population cohorts with rigorous epidemiological design and statistical methodology. Task: Lead or co-lead statistical genetics analyses across four major infection programmes (RSV, TB, HBV, chronic inflammatory disease) covering GWAS, burden testing, host-pathogen joint analysis, and multi-omic integration. Action: Designed study protocols and selected appropriate statistical tests across cohorts ranging from 300 birth-cohort infants to 100,000+ biobank genomes (UKBB, Genomics England); applied GWAS case-control designs with ancestry PC covariate adjustment, gene-level burden testing, and gene-set association methods; contributed to Phase II clinical trial design including endpoint strategy, eligibility criteria, and patient trajectory modelling using multi-omic data; mentored 25 researchers (3 PhD, 22 MSc) in statistical methodology and reproducible analytical best practices; delivered four years of annual guest lectures on quantitative methods at EPFL MSc level (BIO-491, personalised health programme). Result: Multiple lead-author publications in AJHG, JACI, JID, Genetic Epidemiology, Journal of Infectious Diseases. CHF 3M+ in collaborative funding. Methods adopted as independent infrastructure by external groups. First reproducible paediatric sepsis susceptibility locus identified (CTNNAL1/ELP1, Lancet eBioMedicine accepted).
Quantitative go/no-go and scenario modelling frameworks for clinical decisions
Situation: Clinical genetics decisions and trial design relied on categorical variant labels without quantified uncertainty, preventing systematic scenario analysis or evidence-based go/no-go decisions. Task: Build quantitative frameworks that replace binary classifications with continuous posterior probabilities and credible intervals, enabling explicit scenario modelling and threshold-based decision-making. Action: Designed three diagnostic scenarios (complete coverage/simple diagnosis; incomplete coverage with unobserved plausible alleles; no observed variants) demonstrating how posterior probability and credible intervals update as evidence changes; modelled both observed true positives and unobserved false negatives jointly in a Beta-Binomial simulation; defined score-positive-total as a non-normalised prior summary enabling ranked candidate prioritisation; derived expected case counts and probability of at least one event across national population sizes with Bayesian mixture adjustment, validated against UK CVID (NFKB1) and CF (CFTR) registries. Separately built the qualifying variant framework enabling pipeline-independent YAML/JSON criteria that reproduce conventional analyses identically while remaining auditable and reusable. Result: Quantitative diagnostic decision framework validated against two national clinical cohorts. Framework adopted in SwissPedHealth Molecular Board workflows directly informing treatment decisions (rituximab, anakinra, tocilizumab) for approximately 1,000 children.
Rare variant association methodology — SKAT-O, pathway VSAT, and novel visualisation
Situation: Standard GWAS cannot detect rare variant effects in clinically ascertained cohorts because individual variants are too rare to reach significance; no method existed for visualising variant-set association test statistics alongside single-variant GWAS results in genomic context. Task: Design and implement a complete multi-tiered rare variant association framework for a paediatric sepsis cohort, and develop a novel visualisation method addressing the coordinate assignment problem for VSAT statistics. Action: Benchmarked SKAT, SKAT-O, burden tests, and ACAT across simulation scenarios; selected SKAT-O with Lee et al. resampling for small-sample calibration; built ProteoMCLustR for unbiased pre-analysis protein pathway construction from STRINGdb (MCL algorithm, 19,566 proteins, confidence thresholds 0.4/0.7/0.9); designed collapsing strategy following Povysil et al.; applied covariate adjustment for ancestry PCs, sex, age, study site, and ICU status across 940 children; separately derived the Archipelago coordinate assignment algorithm assigning genomic positions to VSAT p-values, validated across 1000 Genomes, Pan-UK Biobank (490,640 participants), and UK Biobank WGS. Result: Three significantly enriched protein pathways identified in paediatric sepsis (first such finding in this population). Archipelago published Wiley Genetic Epidemiology 2026. CRAN-released. Used as standard visualisation in subsequent SwissPedHealth analyses.
Likelihood-based evidence ratio framework — finite-sample valid reporting standard
Situation: Clinical trial and biomedical study results are reported heterogeneously across outcome types, making cross-study comparison of evidential support impossible and collapsing distinct inferential questions into P-value proxies. Task: Formalise a likelihood-based evidential reporting unit applicable uniformly across continuous, binary, and time-to-event clinical endpoints, with finite-sample frequentist guarantees. Action: Derived the evidence ratio E(x) = m1(x)/m0(x) from first principles, proved the finite-sample guarantee E[E(X)] ≤ 1 under H0 using results on nonnegative supermartingales and e-values; demonstrated application across seven canonical clinical analysis types including one-sample mean tests, two-sample comparisons, logistic regression, Cox proportional hazards, and accelerated failure time models; formalised as a normative reporting standard (SGA-ERRS-1.0, Zenodo DOI) and implemented in a CRAN R package. Result: First author. Normative standard issued. CRAN implementation. Applicable to all standard statistical tests without new modelling assumptions. Directly supports unified evidence storage and cross-study comparison in clinical development.
Bayesian statistical framework design — from first principles to clinical deployment
Situation: Clinical genomic interpretation relied on heterogeneous, non-comparable evidence checks with no quantitative standard for measuring evidence sufficiency or calibrated uncertainty. Task: Design a principled Bayesian framework for evidence quantification that would operate independently of upstream pipelines, scale to whole-genome analysis, and produce interpretable outputs for clinical and non-specialist audiences. Action: Derived a closed-form Beta-Binomial conjugate model from first principles, choosing the parametrisation deliberately to model evidence rules as Bernoulli trials without assumptions about rule weighting or dependence; implemented as deterministic C algorithm (no simulation required) and cross-platform R package; validated on public GIAB WGS trio; conducted ethics-approved user survey confirming improved clinician understanding. Separately derived a genome-wide prior probability framework integrating Hardy-Weinberg equilibrium across AD, AR, and XL inheritance modes with population allele frequencies and ClinVar classifications, validated against national disease cohorts (NFKB1/CVID UK, CFTR/CF UK registry, SCID-specific genes). Result: Two peer-reviewed Bayesian frameworks (QuantBayes, variant risk estimation), both validated against real clinical cohorts, with 3,000+ combined CRAN downloads per quarter and adoption in live clinical diagnostic workflows.
EPFL translational cohort analyses — population-scale infection genomics
Situation: Understanding why some individuals develop severe responses to infection required integrating host genome, pathogen genome, and clinical outcome data across large population cohorts and international collaborations. Task: Design and execute statistical genetics and multi-omic analyses across four major infection programmes — RSV, tuberculosis, HBV, and chronic inflammatory disease. Action: Developed analytical workflows for GWAS, host-pathogen joint analysis, and multi-omic integration across cohorts up to 5,000 participants and biobank-scale analyses exceeding 100,000 genomes. Released R packages adopted by independent groups. Mentored 3 PhD and 22 MSc students. Contributed to Phase II trial design integrating multi-omic data with patient trajectory modelling. Result: Multiple lead-author publications in JACI, JID, Genetic Epidemiology. CHF 3M+ in collaborative funding programmes. Software adopted as independent infrastructure.
Archipelago — variant set association visualisation method
Situation: Variant set association test (VSAT) statistics lacked genomic coordinates, making visual comparison with single-variant GWAS results impossible. Task: Design a method assigning meaningful genomic positions to VSAT p-values and integrating both scales into a single navigable plot. Action: Developed the Archipelago method. Validated across three independent datasets: 1000 Genomes, Pan-UK Biobank (490,640 participants), and UK Biobank WGS. Released as an R package on CRAN with full customisation. Result: First author. Published Genetic Epidemiology (Wiley, 2026). Applied in SPSS and SwissPedHealth analyses. CRAN DOI 10.32614/CRAN.package.archipelago.
PanelAppRex — disease-gene panel aggregation and AI search
Situation: Disease-gene panel data was fragmented across sources, inconsistently annotated, and inaccessible for programmatic use in diagnostic pipelines. Task: Build a system aggregating, harmonising, and making queryable the full landscape of curated disease-gene panel evidence. Action: Aggregated 58,592 gene-disease associations from Genomics England PanelApp, ClinVar, UniProt, and Ensembl. Applied RAG on the UniProtKB human proteome to compress 6.6 million features fivefold into structured disease-aware panel summaries. Built a natural-language search interface and machine-readable export formats. Benchmarked across 15 published case studies spanning immunology, neurology, and mixed disease areas. Result: Senior/corresponding author. Published Bioinformatics Advances (OUP, 2026). 93–100% benchmark accuracy. Public dataset on Zenodo. Integrated in SwissPedHealth clinical workflows.
Leeds rare immune disease programme — 10+ disease discoveries
Situation: Approximately 500 patients with severe immune-mediated disease at St James's University Hospital lacked a genetic diagnosis, preventing targeted treatment. Task: Contribute genomic discovery and functional validation across multiple disease programmes within the NIHR BioResource rare disease cohort. Action: Led or co-led discovery of germline TET2 deficiency (Blood 2020, IUIS-classified), RAG deficiency in adults (JACI 2018), RAG1/2 variant occurrence modelling (JCI 2019), CRACR2A deficiency (eLife 2021), TNFAIP3 autoinflammation (Frontiers Immunol 2018), CFTR as a modifying cofactor in PID (JACI 2023), and pyrin pathway autoinflammation (Science Translational Medicine 2016). Applied WES/WGS, flow cytometry, B-cell differentiation, iPSC haematopoietic systems, and functional assays. Result: More than 10 peer-reviewed disease discovery and precision treatment findings. Treatments confirmed including rituximab, anakinra, and tocilizumab.
IEI prior probability database — 54,814 ClinVar classifications
Situation: Clinical variant interpretation in inborn errors of immunity lacked a quantitative prior probability baseline, forcing reliance on qualitative expert judgement without uncertainty estimates. Task: Build a systematic prior probability database covering the full IEI gene landscape from population-scale ClinVar data. Action: Processed 54,814 ClinVar variant classifications across 557 IEI genes. Derived gene-level and variant-level prior probabilities under both autosomal dominant and autosomal recessive inheritance models. Validated against NFKB1 (AD) and CFTR (AR) clinical cohorts. Result: Senior author. Under review npj Genomic Medicine. Database publicly released at iei-genetics.github.io.
QuantBayes — Bayesian evidence sufficiency for variant interpretation
Situation: Genomic variant evidence is expressed heterogeneously across pipelines and institutions; no quantitative standard existed for measuring evidence sufficiency with calibrated uncertainty. Task: Build a closed-form Bayesian model producing posterior evidence sufficiency estimates with credible intervals from binary evidence matrices. Action: Designed a Beta-Binomial model, implemented as a deterministic C algorithm (no simulation required), released as standalone macOS/Linux binaries and as an R package on CRAN, validated on the GIAB trio WGS dataset, and preprinted with full methodology. A user survey (DESAT ethics approved) confirmed improved clinician understanding with explicit evidence reporting. Result: Senior/corresponding author. CRAN-released December 2025. Over 3,000 combined downloads per quarter. Implementation at quantbayes.com.
Qualifying variant framework — published open standard
Situation: Variant selection criteria were embedded in pipeline code, making them invisible to reviewers, impossible to audit, and unable to be reused across studies. Task: Propose and publish a unified framework for expressing variant criteria as standalone, versioned, pipeline-independent specifications. Action: Designed the QV framework with YAML/JSON criteria, a reference model covering 16 variant states, and three validation studies (rare disease WES cohort of 940 individuals, HapMap3 GWAS, GIAB WGS trio) each achieving bitwise-identical outputs to conventional hard-coded methods. Released as a public database and peer-reviewed paper. Result: First author. Published OUP Bioinformatics (2026). CRAN-released. Adopted in the SwissPedHealth clinical pipeline serving approximately 1,000 children. NHS Genomic Test Directory panels integrated.
Germline TET2 deficiency — new disease entity in Blood
Situation: A cluster of children with immunodeficiency and lymphoma had no genetic diagnosis despite extensive workup across multiple centres. Task: Lead genomic interpretation and support functional validation to identify the mechanism in an international collaborative study. Action: Identified germline TET2 loss-of-function as the causal mechanism; supported functional validation including B-cell differentiation assays, flow cytometry, and immunofluorescence; coordinated IUIS disease classification submission. Result: Germline TET2 deficiency established as a new disease entity. Published Blood (2020, IF ~20). Added to the IUIS inborn errors of immunity classification. Treatment implications confirmed (rituximab). Joint lead author on 30-centre study.
Paediatric sepsis rare variant pathways — 66 IEI candidates
Situation: Common variants explain little sepsis heritability; rare immune pathway variants had not been systematically tested in a paediatric clinical cohort. Task: Design and execute multi-tiered rare variant analyses — single-case, single-variant, gene-level, and protein pathway — across 940 children with confirmed sepsis. Action: Built automated interpretation pipelines covering SKAT-O, burden testing, and protein pathway clustering. Applied novel pathway grouping to identify enriched immune pathway hits. Result: Identified 66 candidate inborn error of immunity cases and two immune pathway hits spanning 50 genes. Findings inform treatment decisions at the Molecular Board. First author. Under review.
Paediatric sepsis GWAS — first reproducible susceptibility locus
Situation: Sepsis is the leading infectious cause of childhood death; no reproducible genome-wide susceptibility locus had been identified in a paediatric cohort. Task: Lead the statistical genetics analysis within a national multi-centre programme (CHF 930K, three Swiss children's hospitals). Action: Designed and executed a GWAS in 650 blood-culture-confirmed paediatric sepsis cases and 1,395 controls; applied case-control and case-only designs; coordinated genotype data governance across institutions; built analysis pipelines from QC through to fine-mapping. Result: Identified the first reproducible paediatric sepsis susceptibility signal at the CTNNAL1/ELP1 locus. First author. Accepted Lancet eBioMedicine 2025. GWAS summary statistics deposited in EBI GWAS Catalogue (GCST90726424).
Actor-critic reinforcement learning for variant evidence interpretation
Situation: Variant evidence interpretation required sequential decision-making across heterogeneous evidence types with no principled framework for learning optimal interpretation strategies from data. Task: Apply actor-critic reinforcement learning to formalise variant evidence interpretation as a sequential decision problem with learnable reward structure. Action: Designed an actor-critic RL framework mapping variant evidence states to interpretation actions; defined reward signals aligned with clinical validity criteria; preprinted methodology with full technical description. Result: Preprint 2025. Demonstrates rapid adoption of emerging AI/ML methodology into statistical genomics interpretation, directly applicable to automated clinical decision support.
Bayesian reference model for genomic interpretation — 12-state evidence framework
Situation: Clinical genomic interpretation was candidate-centred, starting from what was observed rather than what could have been observed, causing posterior probability to concentrate artificially among visible candidates while ignoring null, unknown, and unresolved hypotheses. Task: Derive a mathematically complete Bayesian reference model for genotype-phenotype interpretation that makes the full denominator of genetic inference explicit. Action: Introduced axiomatic genetic findings (the admissible event set defining the denominator before patient results are observed) and probative genetic findings (observations that change posterior belief, formalised as likelihood ratios); derived 12 evidence states from three primitive dimensions (explanatory truth, evidence resolution, interpretation state); specified the full posterior attribution equation including null hypothesis H0, unknown-model hypothesis H_unknown, and unresolved explanatory mass; defined the expected value of resolving evidence (EVR) as a principled guide for follow-up testing priorities. Implemented as QuantCalc framework. Result: Peer-reviewed preprint. Framework provides a logically complete basis for posterior diagnostic attribution under incomplete genomic information. Directly applicable to variant reclassification, clinical trial endpoint definition, and AI training on genomic evidence.
RSV viral genomics — birth cohort determinants of prolonged infection
Situation: The viral genetic determinants of prolonged RSV infection in infants were unknown, limiting understanding of why some children suffer more severe and extended illness. Task: Identify viral genetic factors associated with prolonged RSV infection in a healthy term birth cohort. Action: Analysed RSV genome sequences from 300+ infants in the Vanderbilt birth cohort. Applied viral genome-wide association methods and gene-level tests to identify viral loci associated with infection duration. Result: First author. Published Journal of Infectious Diseases (2022). Identified specific RSV genomic loci associated with prolonged infection, providing targets for intervention strategy.
RAG1/RAG2 variant occurrence — probabilistic precursor to QuantBayes
Situation: RAG1 and RAG2 deficiency spans a spectrum of rare immune disease; no quantitative basis existed for predicting which variants would be observed clinically versus in the general population. Task: Develop a probabilistic framework for predicting variant occurrence across RAG1 and RAG2 using population genetics and functional annotation. Action: Integrated gnomAD population allele frequencies, Hardy-Weinberg priors, inheritance models, and structural/functional annotation. Applied to approximately 500 severe clinical cases from the NIHR BioResource rare disease cohort. Result: First author. Published Journal of Clinical Immunology (2019). Established the quantitative variant occurrence methodology that became the conceptual precursor to the QuantBayes framework.
Digital innovation (12)
Bayesian statistical framework design — from first principles to clinical deployment
Situation: Clinical genomic interpretation relied on heterogeneous, non-comparable evidence checks with no quantitative standard for measuring evidence sufficiency or calibrated uncertainty. Task: Design a principled Bayesian framework for evidence quantification that would operate independently of upstream pipelines, scale to whole-genome analysis, and produce interpretable outputs for clinical and non-specialist audiences. Action: Derived a closed-form Beta-Binomial conjugate model from first principles, choosing the parametrisation deliberately to model evidence rules as Bernoulli trials without assumptions about rule weighting or dependence; implemented as deterministic C algorithm (no simulation required) and cross-platform R package; validated on public GIAB WGS trio; conducted ethics-approved user survey confirming improved clinician understanding. Separately derived a genome-wide prior probability framework integrating Hardy-Weinberg equilibrium across AD, AR, and XL inheritance modes with population allele frequencies and ClinVar classifications, validated against national disease cohorts (NFKB1/CVID UK, CFTR/CF UK registry, SCID-specific genes). Result: Two peer-reviewed Bayesian frameworks (QuantBayes, variant risk estimation), both validated against real clinical cohorts, with 3,000+ combined CRAN downloads per quarter and adoption in live clinical diagnostic workflows.
Switzerland Omics — scientific product ecosystem from IP to revenue
Situation: Probabilistic genomics methods existed in research code but lacked the product architecture, IP positioning, and brand credibility for adoption outside the originating institution. Task: Build a complete scientific product environment from intellectual property registration through to deployed products and commercial revenue. Action: Established Switzerland Omics. Registered the trademark (Swissreg, Nice classes 9 and 42, valid to 2035). Designed the product taxonomy. Built and shipped QuantBayes Studio, PanelAppRex, Genomic Vault, and Variant Impact. Structured the commercial spin-off and AWS research funding application. Result: Registered trademark valid 2035. Multiple live deployed products. Commercial revenue via QuantBayes Studio. AWS research funding secured 2026. Products in clinical and research use internationally.
Archipelago — variant set association visualisation method
Situation: Variant set association test (VSAT) statistics lacked genomic coordinates, making visual comparison with single-variant GWAS results impossible. Task: Design a method assigning meaningful genomic positions to VSAT p-values and integrating both scales into a single navigable plot. Action: Developed the Archipelago method. Validated across three independent datasets: 1000 Genomes, Pan-UK Biobank (490,640 participants), and UK Biobank WGS. Released as an R package on CRAN with full customisation. Result: First author. Published Genetic Epidemiology (Wiley, 2026). Applied in SPSS and SwissPedHealth analyses. CRAN DOI 10.32614/CRAN.package.archipelago.
Secure LLM deployment — clinical variant interpretation at BioMedIT
Situation: Clinical genetics teams required AI-assisted variant interpretation but could not expose patient data to external cloud APIs under Swiss data protection law. Task: Deploy open-weight LLMs on secure HPC infrastructure for clinical report generation with Molecular Board review. Action: Deployed Llama and DeepSeek R1 on BioMedIT HPC. Built a structured report generation pipeline integrating WGS, RNA-seq, and proteomics outputs, reviewed by geneticists, clinicians, and researchers. Implemented audit-trail logging meeting Swiss data governance requirements. Result: High-throughput clinical variant interpretation with full audit trail, operational at University Children's Hospital Zurich within SwissPedHealth.
PanelAppRex — disease-gene panel aggregation and AI search
Situation: Disease-gene panel data was fragmented across sources, inconsistently annotated, and inaccessible for programmatic use in diagnostic pipelines. Task: Build a system aggregating, harmonising, and making queryable the full landscape of curated disease-gene panel evidence. Action: Aggregated 58,592 gene-disease associations from Genomics England PanelApp, ClinVar, UniProt, and Ensembl. Applied RAG on the UniProtKB human proteome to compress 6.6 million features fivefold into structured disease-aware panel summaries. Built a natural-language search interface and machine-readable export formats. Benchmarked across 15 published case studies spanning immunology, neurology, and mixed disease areas. Result: Senior/corresponding author. Published Bioinformatics Advances (OUP, 2026). 93–100% benchmark accuracy. Public dataset on Zenodo. Integrated in SwissPedHealth clinical workflows.
IEI prior probability database — 54,814 ClinVar classifications
Situation: Clinical variant interpretation in inborn errors of immunity lacked a quantitative prior probability baseline, forcing reliance on qualitative expert judgement without uncertainty estimates. Task: Build a systematic prior probability database covering the full IEI gene landscape from population-scale ClinVar data. Action: Processed 54,814 ClinVar variant classifications across 557 IEI genes. Derived gene-level and variant-level prior probabilities under both autosomal dominant and autosomal recessive inheritance models. Validated against NFKB1 (AD) and CFTR (AR) clinical cohorts. Result: Senior author. Under review npj Genomic Medicine. Database publicly released at iei-genetics.github.io.
QuantBayes — Bayesian evidence sufficiency for variant interpretation
Situation: Genomic variant evidence is expressed heterogeneously across pipelines and institutions; no quantitative standard existed for measuring evidence sufficiency with calibrated uncertainty. Task: Build a closed-form Bayesian model producing posterior evidence sufficiency estimates with credible intervals from binary evidence matrices. Action: Designed a Beta-Binomial model, implemented as a deterministic C algorithm (no simulation required), released as standalone macOS/Linux binaries and as an R package on CRAN, validated on the GIAB trio WGS dataset, and preprinted with full methodology. A user survey (DESAT ethics approved) confirmed improved clinician understanding with explicit evidence reporting. Result: Senior/corresponding author. CRAN-released December 2025. Over 3,000 combined downloads per quarter. Implementation at quantbayes.com.
Qualifying variant framework — published open standard
Situation: Variant selection criteria were embedded in pipeline code, making them invisible to reviewers, impossible to audit, and unable to be reused across studies. Task: Propose and publish a unified framework for expressing variant criteria as standalone, versioned, pipeline-independent specifications. Action: Designed the QV framework with YAML/JSON criteria, a reference model covering 16 variant states, and three validation studies (rare disease WES cohort of 940 individuals, HapMap3 GWAS, GIAB WGS trio) each achieving bitwise-identical outputs to conventional hard-coded methods. Released as a public database and peer-reviewed paper. Result: First author. Published OUP Bioinformatics (2026). CRAN-released. Adopted in the SwissPedHealth clinical pipeline serving approximately 1,000 children. NHS Genomic Test Directory panels integrated.
Actor-critic reinforcement learning for variant evidence interpretation
Situation: Variant evidence interpretation required sequential decision-making across heterogeneous evidence types with no principled framework for learning optimal interpretation strategies from data. Task: Apply actor-critic reinforcement learning to formalise variant evidence interpretation as a sequential decision problem with learnable reward structure. Action: Designed an actor-critic RL framework mapping variant evidence states to interpretation actions; defined reward signals aligned with clinical validity criteria; preprinted methodology with full technical description. Result: Preprint 2025. Demonstrates rapid adoption of emerging AI/ML methodology into statistical genomics interpretation, directly applicable to automated clinical decision support.
Genomic Vault — custodianship and governance platform
Situation: Genome data custody was handled informally, with no auditable access control, governance trail, or payment infrastructure for institutional or individual use. Task: Build a working custodianship and governance platform deployable commercially and institutionally. Action: Built fullstack application with encrypted storage, Drizzle ORM, Supabase row-level security, Stripe payment webhooks, immutable audit logs, and SPHN-compliant RDF metadata. Deployed with operational access controls. Result: Live at genomicvault.switzerlandomics.ch. Working payments, access governance, and immutable audit trail. Designed to SPHN compliance standards.
Microsoft Azure Research Award — USD 100K competitive cloud funding
Situation: Population-scale genomic data science required cloud computing capacity not available through standard academic infrastructure at Leeds. Task: Compete for and secure external research funding to enable cloud-based genomic data science and machine learning work. Action: Wrote and submitted a competitive proposal for data science and ML in predictive genomics. Award granted as USD 100,000 in Azure cloud computing credits. Result: Enabled population-scale genomic analysis and ML pipeline development that contributed to subsequent peer-reviewed publications.
Variant Impact — fullstack molecular informatics platform
Situation: Protein variant interpretation required integration of gnomAD, UniProt, ClinVar, AlphaFold, and RCSB PDB — no single platform combined all with an interactive 3D structure viewer and variant landscape visualisation. Task: Build a fullstack web application demonstrating end-to-end molecular informatics. Action: Built with Next.js, React, TypeScript, and Tailwind. Integrated Molstar 3D protein structure viewer, variant landscape heatmaps, protein feature tracks, gnomAD constraint scores, ClinVar classifications, AlphaFold confidence scores, and RCSB structure data. Deployed on Vercel. Result: Deployed at variantimpact.vercel.app. Demonstrates the complete fullstack molecular informatics stack: molecular databases, interactive 3D visualisation, and variant landscape analysis in a single application.
Stakeholder influence (7)
Population-scale GWAS and epidemiological statistics — 5 years EPFL postdoc
Situation: Understanding genetic susceptibility to severe infection required integrating host genome, pathogen genome, and clinical outcome data across large population cohorts with rigorous epidemiological design and statistical methodology. Task: Lead or co-lead statistical genetics analyses across four major infection programmes (RSV, TB, HBV, chronic inflammatory disease) covering GWAS, burden testing, host-pathogen joint analysis, and multi-omic integration. Action: Designed study protocols and selected appropriate statistical tests across cohorts ranging from 300 birth-cohort infants to 100,000+ biobank genomes (UKBB, Genomics England); applied GWAS case-control designs with ancestry PC covariate adjustment, gene-level burden testing, and gene-set association methods; contributed to Phase II clinical trial design including endpoint strategy, eligibility criteria, and patient trajectory modelling using multi-omic data; mentored 25 researchers (3 PhD, 22 MSc) in statistical methodology and reproducible analytical best practices; delivered four years of annual guest lectures on quantitative methods at EPFL MSc level (BIO-491, personalised health programme). Result: Multiple lead-author publications in AJHG, JACI, JID, Genetic Epidemiology, Journal of Infectious Diseases. CHF 3M+ in collaborative funding. Methods adopted as independent infrastructure by external groups. First reproducible paediatric sepsis susceptibility locus identified (CTNNAL1/ELP1, Lancet eBioMedicine accepted).
Quantitative go/no-go and scenario modelling frameworks for clinical decisions
Situation: Clinical genetics decisions and trial design relied on categorical variant labels without quantified uncertainty, preventing systematic scenario analysis or evidence-based go/no-go decisions. Task: Build quantitative frameworks that replace binary classifications with continuous posterior probabilities and credible intervals, enabling explicit scenario modelling and threshold-based decision-making. Action: Designed three diagnostic scenarios (complete coverage/simple diagnosis; incomplete coverage with unobserved plausible alleles; no observed variants) demonstrating how posterior probability and credible intervals update as evidence changes; modelled both observed true positives and unobserved false negatives jointly in a Beta-Binomial simulation; defined score-positive-total as a non-normalised prior summary enabling ranked candidate prioritisation; derived expected case counts and probability of at least one event across national population sizes with Bayesian mixture adjustment, validated against UK CVID (NFKB1) and CF (CFTR) registries. Separately built the qualifying variant framework enabling pipeline-independent YAML/JSON criteria that reproduce conventional analyses identically while remaining auditable and reusable. Result: Quantitative diagnostic decision framework validated against two national clinical cohorts. Framework adopted in SwissPedHealth Molecular Board workflows directly informing treatment decisions (rituximab, anakinra, tocilizumab) for approximately 1,000 children.
Likelihood-based evidence ratio framework — finite-sample valid reporting standard
Situation: Clinical trial and biomedical study results are reported heterogeneously across outcome types, making cross-study comparison of evidential support impossible and collapsing distinct inferential questions into P-value proxies. Task: Formalise a likelihood-based evidential reporting unit applicable uniformly across continuous, binary, and time-to-event clinical endpoints, with finite-sample frequentist guarantees. Action: Derived the evidence ratio E(x) = m1(x)/m0(x) from first principles, proved the finite-sample guarantee E[E(X)] ≤ 1 under H0 using results on nonnegative supermartingales and e-values; demonstrated application across seven canonical clinical analysis types including one-sample mean tests, two-sample comparisons, logistic regression, Cox proportional hazards, and accelerated failure time models; formalised as a normative reporting standard (SGA-ERRS-1.0, Zenodo DOI) and implemented in a CRAN R package. Result: First author. Normative standard issued. CRAN implementation. Applicable to all standard statistical tests without new modelling assumptions. Directly supports unified evidence storage and cross-study comparison in clinical development.
Swiss Genomics Association — founded neutral standards body
Situation: No institution-neutral standard existed for expressing genomic evidence matrices or reporting statistical evidence ratios across heterogeneous study designs and institutions. Task: Found an independent scientific association capable of issuing and maintaining normative standards for the Swiss and international genomics community. Action: Established the Swiss Genomics Association (2024). Issued the Qualifying Evidence Matrix (QEM, Zenodo DOI 10.5281/zenodo.17936587) and the Evidence Ratio Reporting Standard (ERRS, Zenodo DOI 10.5281/zenodo.18261075). Registered both as versioned standards with permanent DOIs. Built a member network spanning Swiss academic, clinical, and industry bioinformatics. Result: Two normative published standards cited in peer-reviewed work. 35+ members across Swiss institutions. ERRS implemented in the CRAN evidenceratio package.
EPFL translational cohort analyses — population-scale infection genomics
Situation: Understanding why some individuals develop severe responses to infection required integrating host genome, pathogen genome, and clinical outcome data across large population cohorts and international collaborations. Task: Design and execute statistical genetics and multi-omic analyses across four major infection programmes — RSV, tuberculosis, HBV, and chronic inflammatory disease. Action: Developed analytical workflows for GWAS, host-pathogen joint analysis, and multi-omic integration across cohorts up to 5,000 participants and biobank-scale analyses exceeding 100,000 genomes. Released R packages adopted by independent groups. Mentored 3 PhD and 22 MSc students. Contributed to Phase II trial design integrating multi-omic data with patient trajectory modelling. Result: Multiple lead-author publications in JACI, JID, Genetic Epidemiology. CHF 3M+ in collaborative funding programmes. Software adopted as independent infrastructure.
SwissPedHealth — CHF 5M national multi-omics infrastructure
Situation: A CHF 5M national programme required unified multi-omics analytical infrastructure across multiple children's hospitals with heterogeneous data systems, governance structures, and clinical workflows. Task: Design and deliver the computational workflows, database architecture, and reporting systems connecting WGS, RNA-seq, proteomics, metabolomics, and EHR data across UZH, ETHZ, CHUV, HUG, and EPFL. Action: Built reproducible pipelines on BioMedIT secure HPC. Structured clinical research outputs for Molecular Board review. Coordinated technical delivery across 30+ collaborators spanning wet-lab, bioinformatics, governance, ethics, and clinical teams simultaneously. Result: Approximately 1,000 children enrolled. Over 100 TB biomedical data under traceable infrastructure. Multiple peer-reviewed publications. Programme completed on schedule across seven institutions.
EPFL BIO-491 guest lecturer — personalised health master programme
Situation: BIO-491 is a flagship EPFL MSc course on personalised health requiring research-active guest lecturers with current clinical genomics expertise. Task: Deliver annual guest lectures to Life Sciences Engineering Master students over four consecutive years. Action: Designed and delivered lectures on human genomics of infection and immunity, probabilistic variant interpretation, and translational precision medicine, drawing on active research at EPFL and University Children's Hospital Zurich. Result: Four years of annual contribution (2021–2024). Spring semester, 2 hours per week format.
Delivery (11)
Rare variant association methodology — SKAT-O, pathway VSAT, and novel visualisation
Situation: Standard GWAS cannot detect rare variant effects in clinically ascertained cohorts because individual variants are too rare to reach significance; no method existed for visualising variant-set association test statistics alongside single-variant GWAS results in genomic context. Task: Design and implement a complete multi-tiered rare variant association framework for a paediatric sepsis cohort, and develop a novel visualisation method addressing the coordinate assignment problem for VSAT statistics. Action: Benchmarked SKAT, SKAT-O, burden tests, and ACAT across simulation scenarios; selected SKAT-O with Lee et al. resampling for small-sample calibration; built ProteoMCLustR for unbiased pre-analysis protein pathway construction from STRINGdb (MCL algorithm, 19,566 proteins, confidence thresholds 0.4/0.7/0.9); designed collapsing strategy following Povysil et al.; applied covariate adjustment for ancestry PCs, sex, age, study site, and ICU status across 940 children; separately derived the Archipelago coordinate assignment algorithm assigning genomic positions to VSAT p-values, validated across 1000 Genomes, Pan-UK Biobank (490,640 participants), and UK Biobank WGS. Result: Three significantly enriched protein pathways identified in paediatric sepsis (first such finding in this population). Archipelago published Wiley Genetic Epidemiology 2026. CRAN-released. Used as standard visualisation in subsequent SwissPedHealth analyses.
Quantitative Omics Epidemiology Group — founding, direction, and delivery
Situation: Clinical genomics at the children's hospital needed a computational group with a shared methodological foundation that could build trusted infrastructure rather than disposable in-house pipelines. Task: Found and lead a research group delivering validated, reusable, openly published genomics infrastructure within a CHF 5M+ national programme. Action: Founded the Quantitative Omics Epidemiology Group within SwissPedHealth. Defined the core strategy: validate every method using public biobank-scale data before clinical deployment, publish reusable open software rather than internal pipelines, and require independent peer review before clinical use. Coordinated 30+ collaborators across wet-lab, bioinformatics, molecular board, ethics, and clinical teams. Result: Five lead/senior-author publications across OUP, Wiley, Lancet journals within two years. Multiple CRAN packages with 3,000+ combined downloads per quarter. Technologies adopted in clinical diagnostic workflows.
SwissPedHealth — CHF 5M national multi-omics infrastructure
Situation: A CHF 5M national programme required unified multi-omics analytical infrastructure across multiple children's hospitals with heterogeneous data systems, governance structures, and clinical workflows. Task: Design and deliver the computational workflows, database architecture, and reporting systems connecting WGS, RNA-seq, proteomics, metabolomics, and EHR data across UZH, ETHZ, CHUV, HUG, and EPFL. Action: Built reproducible pipelines on BioMedIT secure HPC. Structured clinical research outputs for Molecular Board review. Coordinated technical delivery across 30+ collaborators spanning wet-lab, bioinformatics, governance, ethics, and clinical teams simultaneously. Result: Approximately 1,000 children enrolled. Over 100 TB biomedical data under traceable infrastructure. Multiple peer-reviewed publications. Programme completed on schedule across seven institutions.
Secure LLM deployment — clinical variant interpretation at BioMedIT
Situation: Clinical genetics teams required AI-assisted variant interpretation but could not expose patient data to external cloud APIs under Swiss data protection law. Task: Deploy open-weight LLMs on secure HPC infrastructure for clinical report generation with Molecular Board review. Action: Deployed Llama and DeepSeek R1 on BioMedIT HPC. Built a structured report generation pipeline integrating WGS, RNA-seq, and proteomics outputs, reviewed by geneticists, clinicians, and researchers. Implemented audit-trail logging meeting Swiss data governance requirements. Result: High-throughput clinical variant interpretation with full audit trail, operational at University Children's Hospital Zurich within SwissPedHealth.
Leeds rare immune disease programme — 10+ disease discoveries
Situation: Approximately 500 patients with severe immune-mediated disease at St James's University Hospital lacked a genetic diagnosis, preventing targeted treatment. Task: Contribute genomic discovery and functional validation across multiple disease programmes within the NIHR BioResource rare disease cohort. Action: Led or co-led discovery of germline TET2 deficiency (Blood 2020, IUIS-classified), RAG deficiency in adults (JACI 2018), RAG1/2 variant occurrence modelling (JCI 2019), CRACR2A deficiency (eLife 2021), TNFAIP3 autoinflammation (Frontiers Immunol 2018), CFTR as a modifying cofactor in PID (JACI 2023), and pyrin pathway autoinflammation (Science Translational Medicine 2016). Applied WES/WGS, flow cytometry, B-cell differentiation, iPSC haematopoietic systems, and functional assays. Result: More than 10 peer-reviewed disease discovery and precision treatment findings. Treatments confirmed including rituximab, anakinra, and tocilizumab.
Paediatric sepsis rare variant pathways — 66 IEI candidates
Situation: Common variants explain little sepsis heritability; rare immune pathway variants had not been systematically tested in a paediatric clinical cohort. Task: Design and execute multi-tiered rare variant analyses — single-case, single-variant, gene-level, and protein pathway — across 940 children with confirmed sepsis. Action: Built automated interpretation pipelines covering SKAT-O, burden testing, and protein pathway clustering. Applied novel pathway grouping to identify enriched immune pathway hits. Result: Identified 66 candidate inborn error of immunity cases and two immune pathway hits spanning 50 genes. Findings inform treatment decisions at the Molecular Board. First author. Under review.
Paediatric sepsis GWAS — first reproducible susceptibility locus
Situation: Sepsis is the leading infectious cause of childhood death; no reproducible genome-wide susceptibility locus had been identified in a paediatric cohort. Task: Lead the statistical genetics analysis within a national multi-centre programme (CHF 930K, three Swiss children's hospitals). Action: Designed and executed a GWAS in 650 blood-culture-confirmed paediatric sepsis cases and 1,395 controls; applied case-control and case-only designs; coordinated genotype data governance across institutions; built analysis pipelines from QC through to fine-mapping. Result: Identified the first reproducible paediatric sepsis susceptibility signal at the CTNNAL1/ELP1 locus. First author. Accepted Lancet eBioMedicine 2025. GWAS summary statistics deposited in EBI GWAS Catalogue (GCST90726424).
Genomic Vault — custodianship and governance platform
Situation: Genome data custody was handled informally, with no auditable access control, governance trail, or payment infrastructure for institutional or individual use. Task: Build a working custodianship and governance platform deployable commercially and institutionally. Action: Built fullstack application with encrypted storage, Drizzle ORM, Supabase row-level security, Stripe payment webhooks, immutable audit logs, and SPHN-compliant RDF metadata. Deployed with operational access controls. Result: Live at genomicvault.switzerlandomics.ch. Working payments, access governance, and immutable audit trail. Designed to SPHN compliance standards.
Microsoft Azure Research Award — USD 100K competitive cloud funding
Situation: Population-scale genomic data science required cloud computing capacity not available through standard academic infrastructure at Leeds. Task: Compete for and secure external research funding to enable cloud-based genomic data science and machine learning work. Action: Wrote and submitted a competitive proposal for data science and ML in predictive genomics. Award granted as USD 100,000 in Azure cloud computing credits. Result: Enabled population-scale genomic analysis and ML pipeline development that contributed to subsequent peer-reviewed publications.
Variant Impact — fullstack molecular informatics platform
Situation: Protein variant interpretation required integration of gnomAD, UniProt, ClinVar, AlphaFold, and RCSB PDB — no single platform combined all with an interactive 3D structure viewer and variant landscape visualisation. Task: Build a fullstack web application demonstrating end-to-end molecular informatics. Action: Built with Next.js, React, TypeScript, and Tailwind. Integrated Molstar 3D protein structure viewer, variant landscape heatmaps, protein feature tracks, gnomAD constraint scores, ClinVar classifications, AlphaFold confidence scores, and RCSB structure data. Deployed on Vercel. Result: Deployed at variantimpact.vercel.app. Demonstrates the complete fullstack molecular informatics stack: molecular databases, interactive 3D visualisation, and variant landscape analysis in a single application.
ACM Global Laboratories — regulated bioanalytical delivery
Situation: Global clinical trial sponsors required high-quality bioanalytical assay delivery under GCP-aligned standards. Task: Deliver ELISA, flow cytometry, and metabolomics assays within regulated commercial clinical trial workflows. Action: Executed biomarker and assay workflows under GCP/GLP-adjacent quality standards. Maintained documentation, chain of custody, and reporting to sponsor requirements. Result: Reliable quality-controlled assay delivery in a regulated commercial laboratory environment. Direct experience of the documentation and quality culture required in pharmaceutical analytical operations.
Leadership (4)
Switzerland Omics — scientific product ecosystem from IP to revenue
Situation: Probabilistic genomics methods existed in research code but lacked the product architecture, IP positioning, and brand credibility for adoption outside the originating institution. Task: Build a complete scientific product environment from intellectual property registration through to deployed products and commercial revenue. Action: Established Switzerland Omics. Registered the trademark (Swissreg, Nice classes 9 and 42, valid to 2035). Designed the product taxonomy. Built and shipped QuantBayes Studio, PanelAppRex, Genomic Vault, and Variant Impact. Structured the commercial spin-off and AWS research funding application. Result: Registered trademark valid 2035. Multiple live deployed products. Commercial revenue via QuantBayes Studio. AWS research funding secured 2026. Products in clinical and research use internationally.
Quantitative Omics Epidemiology Group — founding, direction, and delivery
Situation: Clinical genomics at the children's hospital needed a computational group with a shared methodological foundation that could build trusted infrastructure rather than disposable in-house pipelines. Task: Found and lead a research group delivering validated, reusable, openly published genomics infrastructure within a CHF 5M+ national programme. Action: Founded the Quantitative Omics Epidemiology Group within SwissPedHealth. Defined the core strategy: validate every method using public biobank-scale data before clinical deployment, publish reusable open software rather than internal pipelines, and require independent peer review before clinical use. Coordinated 30+ collaborators across wet-lab, bioinformatics, molecular board, ethics, and clinical teams. Result: Five lead/senior-author publications across OUP, Wiley, Lancet journals within two years. Multiple CRAN packages with 3,000+ combined downloads per quarter. Technologies adopted in clinical diagnostic workflows.
Swiss Genomics Association — founded neutral standards body
Situation: No institution-neutral standard existed for expressing genomic evidence matrices or reporting statistical evidence ratios across heterogeneous study designs and institutions. Task: Found an independent scientific association capable of issuing and maintaining normative standards for the Swiss and international genomics community. Action: Established the Swiss Genomics Association (2024). Issued the Qualifying Evidence Matrix (QEM, Zenodo DOI 10.5281/zenodo.17936587) and the Evidence Ratio Reporting Standard (ERRS, Zenodo DOI 10.5281/zenodo.18261075). Registered both as versioned standards with permanent DOIs. Built a member network spanning Swiss academic, clinical, and industry bioinformatics. Result: Two normative published standards cited in peer-reviewed work. 35+ members across Swiss institutions. ERRS implemented in the CRAN evidenceratio package.
EPFL BIO-491 guest lecturer — personalised health master programme
Situation: BIO-491 is a flagship EPFL MSc course on personalised health requiring research-active guest lecturers with current clinical genomics expertise. Task: Deliver annual guest lectures to Life Sciences Engineering Master students over four consecutive years. Action: Designed and delivered lectures on human genomics of infection and immunity, probabilistic variant interpretation, and translational precision medicine, drawing on active research at EPFL and University Children's Hospital Zurich. Result: Four years of annual contribution (2021–2024). Spring semester, 2 hours per week format.
