PlasmidGPT

Plasmids are built by reuse. The same origins of replication, selectable markers, promoters and terminators recur across the tens of thousands of constructs deposited in Addgene, recombined into new vectors one junction at a time by hand. That reuse is what makes plasmid design laborious, and also what makes plasmid DNA statistically regular: parts appear in characteristic orders and in characteristic company. PlasmidGPT takes that regularity literally, tokenizing raw plasmid DNA with a byte-pair encoder and training a decoder-only transformer on next-token prediction, so that "what usually follows a kanamycin resistance cassette" becomes a question answerable by sampling rather than by reading a parts registry.

Bin Shao developed PlasmidGPT in the Department of Molecular and Cellular Biology at Harvard University, releasing code, tokenizer and a pretrained checkpoint alongside a bioRxiv preprint in October 2024. An expanded, peer-reviewed version with collaborators at the Beijing Institute of Technology appeared in Science Advances in 2026, adding head-to-head benchmarking against dedicated attribution tools, evaluation on natural plasmids from metagenomes, and wet-lab validation of two generated constructs.

The single frozen checkpoint serves two jobs. Generation is the visible half: given a starting sequence or a vector-type prompt, the model continues it into a complete plasmid. Annotation is the quieter and arguably more useful half — mean-pooled hidden states become a fixed-length embedding of any plasmid, and a single-layer classifier on top of that embedding recovers properties the model was never trained to predict. PlasmidLM, a later promptable designer from University College London, uses PlasmidGPT as its base model.

#Key Features

• Conditional generation from a seed: Supplying a starting sequence, such as a constitutively expressed YFP cassette, biases the continuation toward parts that accompany it in real vectors — the bom site, KanR and a compatible replication origin appear far more often than under random 4-nucleotide seeds.
• Vector-type prompting: Fine-tuning on special tokens for the ten most common vector types lets a user ask for a mammalian expression, bacterial expression, lentiviral, CRISPR or AAV plasmid, and the generated part distributions track the corresponding real ones.
• Attribute prediction from frozen embeddings: A single-layer network over the embeddings predicts lab of origin, vector type, host species, growth strain and selectable marker, reaching above 90% top-1 accuracy on selectable markers and improving on a one-hot CNN baseline by 18% for species.
• Generalization to natural plasmids: Embeddings of environmental plasmids from the IMG/PR database support host taxonomy prediction at phylum and genus level, matching a CNN trained directly on those sequences and outperforming HOTSPOT, despite no natural plasmid appearing in pretraining.
• Wet-lab validated designs: Two generated sequences, 3,553 bp and 4,120 bp, were assembled by de novo gene synthesis and expressed YFP stably in Escherichia coli.

#Technical Details

The architecture is a 12-layer decoder-only transformer with an embedding dimension of 512, 8 attention heads, a context of 2,048 tokens and roughly 110 million parameters. Its BPE tokenizer was trained on the plasmid corpus itself, producing units from 1 to 3,123 bp long with a mean of 54 bp, which compresses a typical plasmid to a median of 223 tokens; sequences exceeding the context window are truncated. Training used 153,208 Addgene plasmids longer than 2,000 bp with a February 2023 cutoff, expanded tenfold by circular-rotation augmentation, optimized with Adam at a learning rate of 2 × 10⁻⁴ for 15 epochs. On lab-of-origin prediction with cluster-based cross-validation the model reaches 74% top-1 and 88% top-10 accuracy; evaluated under the protocol used by prior attribution work, it reports 76.6% top-1 against 74.1% for a triplet metric-learning network and 90.3% top-10 against 89.2%.

#Applications

The generation side suits early-stage construct design, where a researcher wants several plausible backbones around a payload before committing to synthesis. The annotation side is broader: a laboratory receiving an unlabeled plasmid can infer vector type, likely host and marker from sequence alone; attribution and biosecurity work can use lab-of-origin prediction; and metagenomics groups can assign hosts to plasmid contigs from environmental samples. Colab notebooks run both halves in a browser without local installation.

#Impact

PlasmidGPT extended sequence-level language modeling from genes and genomes to the engineered vectors that molecular biology runs on, and showed that pretraining on a curated engineering repository yields representations that transfer to wild sequences. It has since been adopted as a base model for downstream plasmid designers. The limitations are real: generation recombines existing modules rather than inventing new function, the fixed token vocabulary bounds the diversity of parts the model can express, the Addgene corpus is biased toward heavily reused backbones, and only two generated constructs have been tested experimentally. The code is MIT-licensed, but the released weights carry a CC BY-NC 4.0 license that excludes commercial use.