Dataset and Results for Generative Plasmid Design

Description

This archive contains all datasets, generated sequences, processed outputs, analysis notebooks, and experimental results required to fully reproduce the computational and wet-lab findings presented in the associated manuscript. The contents are organised into modular folders reflecting each stage of the workflow: dataset curation, model training, generative sampling, bioinformatic quality control, candidate selection, and experimental validation.

Folder Overview

all_fastas/

Provides the complete curated E. coli plasmid datasets (ft15k and ft35k) used as model training corpora. Information of these are in the readme.txt file.

datasets/

Contains all training resources used to fine-tune the DNA language models.

GFP_cassette.fasta

training_metadata/ includes all accompanying metadata tables (Addgene annotations, PlasmidScope records, and filtering outputs).

A detailed description of dataset construction is provided in  readme.txt.

figure_notebooks/

Jupyter notebooks and final PDF outputs used to generate every figure in the manuscript, including Figures 1–5 and all supplementary figures.These notebooks contain the exact code and plotting steps used for statistical analyses, distributions, similarity metrics, and visualisation of generated and experimental outcomes.

generated/

All sequences produced by the base, ft15k, and ft35k models under both ATG and GFP prompts.Each generation set includes:

the raw sequence file from the sampling run,

accompanying metadata tables describing length, GC content, and prompt context.This provides full transparency of all 6,000 sampled plasmids used in downstream QC.

qc_results/

Comprehensive bioinformatic assessments for all generated plasmids.This includes:

ORI and ARG detection summaries,

pass/fail realism filters,

aggregated repeat analysis,

similarity scores to training sets,

QC summaries for each model and prompt.These results correspond directly to the analyses shown in Figures 2–3.

synthetic_candidates/

Mutation calls and sequence-level analyses for plasmids selected for wet-lab synthesis (719, 808, 881, 891) and the positive control.Each plasmid folder contains full mutation call TSVs from long-read sequencing of individual colonies, enabling validation of backbone stability and mutation patterns.

training_logs/

Training logs for the 15k and 35k fine-tuning runs, providing information on optimisation behaviour, loss curves, and convergence characteristics.

wetlab_results/

Processed experimental measurements supporting Figure 5 and supplementary analyses.Includes:

plate_reader_data/ (GFP fluorescence and growth curves for all tested colonies),

sequencing/ (raw sequencing outputs structured by plasmid).These files document the viability, expression levels, and mutational outcomes of all experimentally validated plasmids.

Purpose

Together, these files constitute a complete reproducibility package:

all training data,

all generated sequences,

all QC outputs,

all analysis notebooks,

all experimental readouts.

This archive enables full reconstruction of the computational pipeline, biological filtering and wet-lab characterisation described in the manuscript.