Data scientist & lab automation · PhD, Pharmacology & Physiology, UCL
I automate and analyse high-content screening for CFTR drug discovery, from the microscope to the final report. During my PhD I built:
- lab automation and systems integration that connects automated high-content confocal imaging systems directly to the analysis pipeline. Finished plates are transferred automatically, and the instrument's acquisition metadata sets the QC settings and records which wells were acquired. Wells lost to instrument faults are removed without manual editing, and Excel, GraphPad Prism and PowerPoint reports are produced at the end. This cut hands-on time per plate from 3 hours to 20 minutes.
- deep-learning segmentation (Cellpose) of single cells, scaled on high-performance computing (HPC) clusters
- statistical QC, assay validation (Z′-factor, SSMD) and dose-response modelling
The results inform treatment choices for patients with rare CFTR mutations.
🌐 Portfolio & CV · 💼 LinkedIn
| Project | What it does | Stack |
|---|---|---|
| hcs-lab-automation | Lab automation and systems integration, from instrument to report with no manual steps. It watches the imaging system's output, detects finished plates and transfers them with checksum verification, acquisition metadata included. That metadata selects the instrument-specific QC profile, and the tool reconciles the plate map, the acquisition record and the images so that unacquired or faulty wells are removed automatically. It builds the run and executes the analysis locally, on HPC clusters (SSH, job scheduler) or in the cloud (AWS S3 + Batch, Docker, GPU), with a Z′ QC gate that flags or rejects failing plates. It then generates an Excel workbook, a GraphPad Prism file, figures, a PowerPoint deck and a summary. It also sends email/Teams alerts, logs every run to SQLite with statistical process control trend charts, and exports ELN/LIMS packages (RO-Crate), with a full audit trail. | Python · instrument integration · Prism · python-pptx · CI |
| hcs-lab-agent | Agentic AI for the lab. A multi-agent system (Claude API) of QC-reviewer, planner and plate-designer agents. They reason over the run history through tools to propose the next, most informative plate, with structured hand-offs, self-correction and a full audit trace. Tested in CI with a scripted model. | Python · Claude API · agentic AI · tool use |
| assay-qc-optimisation | Chooses image-analysis parameters by assay quality: a constrained grid search scored by robust and paired SSMD and by Z′-factor, with annotated heatmaps and simulation-validated tests. | Python · NumPy · statistics |
| dose-response-fitting | Four-parameter Hill fitting with asymptotic and bootstrap EC50 confidence intervals, plus an extra-sum-of-squares F-test for potency shifts. Validated by simulation (CI coverage, type I error). | Python · SciPy · statistics |
| single-cell-outlier-qc | Statistical QC that removes dead and detached cells from imaging data, using per-condition mean + k·SD or fixed thresholds. Includes a sensitivity analysis and validation on simulated ground truth, with tests run in CI. | Python · pandas · NumPy · pytest |
CFTR-SCAN, Vergani Lab, UCL (co-author on the associated paper)
CFTR-SCAN is a single-cell pipeline for the mCherry-YFP-CFTR assay, built by the lab over several years. My contributions:
- Assay-quality analysis: SSMD and Z′-factor optimisation to define membrane zones and red-fluorescence thresholds
- Standard Cellpose version of the pipeline, built as a fallback when the custom-trained model underperformed
- HPC cluster deployment, which I set up end to end
- Figure and analysis packages for the GoSlo-SR-5-6, pHuji and WNK studies
Python MATLAB R · Cellpose scikit-image pandas · Lab automation Instrument integration High-content confocal imaging GraphPad Prism · Assay development & validation (Z′, SSMD, S/B) · High-performance computing (HPC) Cloud computing (AWS, Docker) Agentic AI & multi-agent systems · ELN / LIMS · Patch-clamp