Software & Pipelines
Open-source code behind my research: metagenomics workflows built for HPC clusters, machine-learning models for antibiotic discovery, and interactive apps for exploring large screening datasets. Everything is on GitHub.
Workflows for turning raw shotgun reads into genomes, taxonomic profiles and resistome/mobilome quantification, written to run reproducibly on SLURM clusters.
A modular metagenomic binning framework that recovers high-quality metagenome-assembled genomes using ensemble binning followed by fast consensus refinement. Supports direct binning of single assemblies or optional group-based co-assembly of replicates and conditions, with separate Illumina and Nanopore paths.
End-to-end analysis for a culture-enriched hybrid metagenomics study of the AMR footprint of livestock manure in anaerobic digestate: read QC, MetaPhlAn profiling, binning, resistance-gene and mobile-element annotation with RGI, AMRFinder, BacMet and mobileOG, read-mapping quantification (CPM), and publication figures. Co-first author on the resulting paper.
Scripts for genome-resolved profiling of an expanded swine gut isolate collection: genome annotation, feature integration, read mapping, machine-learning classification of health and disease signatures with XGBoost, and network-based guild analysis. Built for modular use on the Grex HPC cluster. Companion to my first-author paper accepted at Microbiome.
Analysis code and interactive apps from CIMPLE-Seq, a pooled CRISPRi screening approach that maps chemical–genetic interactions across the essential genome of Burkholderia cenocepacia K56-2.
An interactive Shiny application for exploring chemical–genetic interaction profiles of essential-gene knockdown mutants screened against ~5,000 compounds. Includes CGIP normalisation, k-means clustering, t-SNE embedding, Tanimoto similarity search, and filtering by compound, gene target or mechanism of action.
A Python/Plotly Dash implementation of the interaction-profile explorer for the B. cenocepacia K56-2 essential genome, packaged for cloud deployment.
Machine-learning analysis of pooled CRISPRi-seq data: random-forest prediction of mutant depletion, kernel PCA with k-means cluster labelling, and sgRNA design scripts used to build the rationally designed knockdown library.
Models that learn from high-throughput screening and genomic data to prioritise antibacterial compounds and essential targets.
Trained weights, raw screening data and benchmarks for a directed message-passing neural network (D-MPNN) trained on a high-throughput antibacterial screen. Classification and regression models were evaluated on random and scaffold splits with RDKit descriptors and Morgan fingerprints, and the model measurably increased the hit rate of follow-up screening.
Domain-search tooling for identifying putative essential protein domains from high-density transposon mutagenesis data, supporting target prioritisation for antibacterial discovery.
Active pipelines from my current collaborations. Code is in private repositories until the associated manuscripts are submitted.
HPC-scale docking of a 791,883-compound natural-product library (COCONUT, ZINC15, Super Natural 3, UNPD) against 20 archaea-specific Methanobrevibacter smithii targets prioritised from genome-scale metabolic models, with automated ligand and receptor preparation, SLURM array orchestration, and OpenMM molecular-dynamics refinement of top hits.
Curated genome-scale metabolic models for M. smithii and nine further rumen methanogens, merged from CarveMe, ModelSEED and RAVEN drafts, harmonised to the BiGG namespace, gap-filled, GPR-annotated and quality-scored with memote. Used to define essential, archaea-specific reactions for methane mitigation, and extended to community-level rumen GEMs that couple methanogens with fermenting bacteria.
End-to-end workflow for the Topigs Norsvin collaboration: metagenome assembly and DAS Tool binning, functional profiling against an in-house isolate and MAG reference, CAZyme and metabolic-gene-cluster feature matrices, and interpretable classifiers (random forest, gradient boosting, FT-Transformer, TabNet) with SHAP explanations and leave-one-farm-out validation.