Molecular Diagnostics · Microbial Genomics · Computational Biology
I help biotechnology and diagnostics teams select and evaluate diagnostic targets, support molecular assay development, and build reproducible genomics workflows — combining experimental experience with computational analysis.
I am a molecular biologist with a microbiology background and industry experience in molecular diagnostics, assay development and biotechnology.
Alongside that work, I have built independent computational biology projects: reproducible bioinformatics workflows for microbial genomics, metagenomics and diagnostic applications, extended more recently to human clinical genomics — variant calling, ACMG/AMP classification, pharmacogenomics and somatic tumour–normal analysis.
Having worked on both sides, I can connect what an assay has to do in practice with what the sequence data can and cannot show. The case studies below are self-directed projects, with code on GitHub, that show how I work.
I work with biotechnology, diagnostics and life-science teams on problems where molecular-biology insight and reproducible computational analysis both matter.
Computational support for diagnostic assay design and pathogen genomics.
Reproducible, benchmarked analysis workflows for human genomic data.
My work is computational and advisory. I do not provide clinical interpretation of patient results or regulated diagnostic reporting. Engagements are scoped project by project — get in touch to discuss yours.
Independent projects that show the methods behind the services above. Each links to its code on GitHub.
Benchmark of taxonomic profiling and antimicrobial-resistance gene detection on public Nanopore shotgun data from a defined gut mock community (Zymo D6331, PRJNA1181840). Experiments were registered before they were run, tuned on a development read pool, frozen with a git tag, and scored once on a separate evaluation pool; every arm and deviation is reported, including the target that was missed. Research use only.
Kraken2 and Sylph were assessed against the community’s theoretical composition, including sequencing-depth, confidence-threshold and abundance comparisons. Resistance-gene reads were then tested for correct assignment to their bacterial carriers. Species detection was strong, but species attribution missed its registered target. The sequenced kit lot could not be confirmed, limiting interpretation against the theoretical reference composition.
Reproduced the metagenomic analysis of a published wastewater study (Taylor et al., 2025; PRJNA904380) using public sequencing reads, reported methods and documented assumptions. Nine Day 1 liquid samples were used for development; the criteria and scoring script were then frozen before 18 held-out liquid samples were evaluated once. The study examined read counts for five resistance genes and compared detection calls with the authors’ published qPCR findings. Research use only.
All registered reproduction criteria were met, but ermB counts differed substantially in three effluent samples and the published ARG/16S scale could not be reproduced. Exploratory analyses supported competition between closely related references as a factor in KMA read assignment, although the exact cause of the difference from the authors remains unresolved. Gene-level agreement was stronger than subtype-level agreement; the qPCR comparison was descriptive.
Reanalysed published 16S rRNA data from humanised mice to compare gut microbiome recovery after ciprofloxacin under standard and low-MAC diets. Built a QIIME 2 workflow using DADA2 and SILVA, with a prespecified analysis plan, cage-level summaries and sensitivity analyses. Reproduced the primary result in nf-core/ampliseq with matched settings.
Low-MAC cages were closer to their own pre-antibiotic composition during days 8–12 by Bray–Curtis distance. Baseline samples for the two diets were sequenced in different runs, which limits attributing this pattern to diet.
Original data: Ng KM, Aranda-Díaz A, Tropini C, et al. Recovery of the gut microbiota after antibiotics depends on host diet, community context, and environmental reservoirs. Cell Host & Microbe. 2019;26(5):650–665.e4. doi:10.1016/j.chom.2019.10.011
Sequence data: Stanford Digital Repository, purl.stanford.edu/rm696bw4911, licensed CC BY-NC 3.0.
Analysis and figures: Krzysztof Gizynski, independently reanalysing the published dataset of Ng et al. (2019).
End-to-end Nextflow DSL2 pipeline characterising antimicrobial resistance across 283 clinical strains. Integrates AMRFinderPlus resistance profiling with a custom alignment-based mutation caller (penA, gyrA, parC, mtrR, porB) and promoter-variant detection, core-genome phylogeny via Parsnp, and an interactive Dash dashboard linking resistance phenotypes to phylogenetic context for exploratory analysis and antimicrobial-resistance surveillance.
Scalable pipeline processing 375 publicly available S. pyogenes genomes with integrated resistance profiling, MLST, and emm typing. Identified 22.9% AMR prevalence across the cohort, with ST28 as the dominant sequence type (22.4%) and emm5 the most prevalent emm type (47.5%). Designed for rapid integration of newly sequenced clinical isolates into ongoing surveillance.
Reproducible target-selection workflow for a multiplex qPCR/TaqMan assay for C. trachomatis. Genome QC and tiering are performed before target discovery, ensuring that consensus sequences and inclusivity assessments are built only from high-confidence assemblies. A provenance audit traced 149 GenBank-only genomes to a single study of engineered laboratory strains; these were excluded from the diagnostic reference set, which was rebuilt from 125 complete RefSeq assemblies that were independently verified and tiered.
Four chromosomal loci (mutL, aroB, PSD2, and group_306) were independently evaluated as candidate targets alongside the multicopy diagnostic plasmid, which serves as the higher-sensitivity target. The proposed duplex pairs one chromosomal locus with the plasmid target, combining genomic redundancy with sensitivity.
The workflow integrates pan-genome analysis, sequence-conservation and inclusivity screening, exclusivity BLAST, quantitative Tm-gap assessment, and three-way structural QC of primer–probe sets, producing a computationally screened candidate panel for downstream experimental validation.
Comprehensive phylogenomics pipeline spanning the full analysis chain for 50 K. pneumoniae genomes. Assembly QC (QUAST), AMR and resistance gene profiling (AMRFinderPlus), virulence and typing (Kleborate), SNP-based variant calling (Snippy), recombination masking (Gubbins), and maximum-likelihood phylogeny (IQ-TREE). Results delivered via a three-tab Dash dashboard. ST11 identified as the dominant lineage, consistent with global carbapenem-resistant K. pneumoniae trends.
Reads-to-VCF pipeline scoped to the CAPN3 and DMD loci, independently benchmarked against Genome in a Bottle's HG002 reference sample. Raw reads are QC'd (FastQC/MultiQC) and adapter/quality-trimmed (fastp) before alignment. Two variant callers — GATK HaplotypeCaller and DeepVariant — are run independently and cross-checked against each other, then the concordant call set is annotated with Ensembl VEP transcript consequence and gnomAD population frequency, producing evidence ready for the downstream ACMG classifier.
A scoped ACMG/AMP (Richards et al. 2015) variant classification engine, now spanning three genes and two inheritance mechanisms: CAPN3 (limb-girdle muscular dystrophy, autosomal recessive) and DMD (Duchenne/Becker muscular dystrophy, X-linked) form the original scope, extended to BRCA1 (hereditary breast/ovarian cancer syndrome, autosomal dominant with incomplete penetrance) as a deliberate generalization test against the real ClinGen ENIGMA BRCA1/BRCA2 VCEP v1.2 specification. Applies twelve ACMG criteria across population frequency, protein effect, computational evidence, functional-assay data, and structural variation — including a dedicated CNV deletion/duplication scoring module for DMD's largely exon-deletion mutational spectrum — combined by two independent systems (classic Table 5 and Tavtigian et al. 2020's Bayesian points) cross-checked against each other, plus a case-level layer reasoning about whether a patient's variant(s) actually explain their disease: biallelic X-linked (XX) inheritance scenarios for DMD, and a new "risk-conferring" case status reflecting BRCA1's dominant, incompletely penetrant mechanism rather than reusing the deterministic status the other two genes use.
A reproducible pharmacogenomics interpretation engine translating selected genomic variants into gene-specific allele/diplotype calls, functional phenotypes, and guideline-linked drug recommendations — built around explicit uncertainty rather than resolving genuinely ambiguous phasing by assumption. Five genes were chosen deliberately because they don't share one interpretation model: linked-variant dosage tables, activity-score summation, transporter biology, and direct multi-locus resolution. Deliberately scopes out CYP2D6 (structural-variant calling, unsolved even by specialist tools) rather than quietly shipping an incomplete answer. A standalone companion to the ACMG/AMP variant classifier below — same engineering discipline, applied to drug response instead of disease causation.
Tumour-normal somatic variant calling pipeline built around GATK Mutect2, benchmarked against a real published truth set for the COLO829/COLO829BL melanoma reference cell line pair. Reads are QC'd and aligned before contamination estimation and an interval-sharded Mutect2 call set — rebuilt as a genome-scale scatter/gather architecture after an unsharded run hit a real memory ceiling — is benchmarked with som.py and extended with CNVkit copy-number calling. A final module annotates PASS calls with SnpEff and queries CIViC's live clinical-evidence API, carrying a called variant through to real oncology evidence rather than stopping at a filtered VCF. Run against real ENA-extracted sequencing reads, the pipeline independently recovered three of COLO829's own documented driver mutations — BRAF V600E, a CDKN2A frameshift, and a UV-signature TERT promoter mutation — with zero false positives against the benchmark truth set.
If you have a diagnostics, genomics or bioinformatics problem you would like to discuss, or want to talk through any of the case studies, I'd be glad to hear from you. A short description of the question and your timeline is enough to start.
How I handle your information: Privacy Notice