SENIOR DATA SCIENTIST · MITRA BIO · LONDON
I'm Harry Pink, a data scientist and computational biologist. My PhD inferred the gene regulatory networks behind lettuce's defence against fungal disease. Today I build machine-learning models that read skin-ageing phenotypes from DNA methylation.
Jul 2025 – present
Machine-learning models that predict skin-ageing phenotypes directly from DNA methylation. Lead analyst on clinical-research collaborations, including the Candela split-face laser study (Scientific Reports, 2026).
Sep 2023 – Jul 2025
Built core parts of the customer-facing duet +modC / evoC Nextflow pipeline, from raw FASTQ to biological insight. Led the pysam tooling for simultaneous 5mC and 5hmC quantification from BAM files. Drove productisation, performance work and large cuts to storage footprint.
Oct 2022 – Mar 2023 · contract
Designed the database and data structures behind Bauer Media's SMS "Competitions" channel across its UK and EU businesses; complex PostgreSQL for metrics, data quality, lineage and BI.
2019 – 2023
Causal gene regulatory network inference from time-course RNA-seq to find master regulators of lettuce defence against Botrytis cinerea; validated hub genes in planta. QTL mapping of quantitative resistance loci (Theoretical and Applied Genetics, 2022). Also: data analyst for Metaverse Holdings (2021–22) and graduate teaching assistant in R / tidyverse (2020–21).
2016 – 2019
Graduated top of year. A summer research project in the Hochegger lab became my first paper, on centrosome dynamics and genome stability (Cell Reports, 2020).
[02] PhD · Denby Lab · University of York · 2019–2023
My PhD used machine learning on time-series gene expression to find the transcription factors that run lettuce's defence against fungal disease, and which are worth breeding for. Scroll to follow the signal from infection to hub to target.
01 / 04 · time series
I RNA-sequenced infected lettuce leaves every 3 hours. The hub regulators switched on before any lesion was visible.
in the data · 14 timepoints from 9 to 48 h, so you can see which genes move first instead of guessing.
02 / 04 · inference
Random forests learned which transcription factors predict each gene's expression across four datasets, drawing the regulator→gene arrows.
in the data · Ranking each TF by feature importance gave a network of 10,947 regulator→gene links.
03 / 04 · hubs → targets
A few hubs control hundreds of genes, and their targets match known biology: defence hormones, cell death, antifungal compounds.
in the data · Sorting regulators by how many genes they control cut thousands of candidates to a handful worth testing.
04 / 04 · validation
LsNAC53 dampens defence and switches on its predicted targets; LsBOS1 boosts resistance.
in the data · The network made the predictions; experiments in the plant confirmed them.
[03] biomodal → Mitra Bio · 2023–now
DNA methylation is one of the switches that decides whether a stretch of genome can be read. Scroll to watch a promoter lose its methyl marks, its chromatin open, and its gene get transcribed — the biology underneath the data I now model.
01 / 04 · silenced
Methyl marks (5mC) on a promoter pack the DNA tightly. The gene stays silent.
in the data · Each CpG becomes a number between 0 and 1, with millions of them per sample.
02 / 04 · demethylation
TET enzymes oxidise the marks away, via 5hmC, back to a plain C.
in the data · At biomodal I wrote the tooling that separates 5mC from 5hmC within the same sequencing reads.
03 / 04 · chromatin opens
Nucleosomes slide apart and transcription factors can bind.
in the data · A single CpG is noisy, but thousands shifting together give a model something to learn.
04 / 04 · transcription
RNA polymerase reads the gene, and the mRNA heads out to make protein.
in the data · These shifts track age. At Mitra Bio I turn them into measurements of skin ageing.
Shown in sequence for clarity. In real cells these steps are coupled, and the order varies from locus to locus.
[04] Mitra Bio · Senior Data Scientist · Jul 2025–now
One promoter is one switch. Across thousands of CpG sites, methylation drifts with age consistently enough to learn. I build machine-learning models that turn that pattern into quantitative skin-ageing phenotypes for skincare and therapeutic R&D.
Lead analyst on a split-face clinical study with Candela: 1940-nm non-ablative fractional laser shifted the skin methylome away from its age-associated drift at most responsive loci.
Scientific Reports · 2026 ↗[05] Reading room
Pull a paper off the pile to read it here.
[06] Toolkit
[07] Get in touch
© Harry Pink · London ·