PlasmidLM: A Promptable DNA Language Model via Verifiable-Reward Post-Training

Abstract

Generative DNA models are typically next-token completers: they extend a sequence but offer no native interface for telling the model what to make. PlasmidLM is a promptable DNA language model for plasmids. A designer supplies a human-readable component specification, for example a high-copy E. coli vector with kanamycin resistance and an EGFP reporter, and the model generates the corresponding multi-kilobase construct in a single autoregressive pass. Prompts are unordered sets of named-part tokens at the granularity of biological shorthand, not learned latent codes or rigid grammars. We evaluate outputs along two axes: a sequence is viable if structurally plausible as a plasmid, and faithful if its detected components match the prompt. Their conjunction is the useful-plasmid rate, the primary metric we report. On a held-out 1,000-prompt benchmark, the post-trained model achieves a useful-plasmid rate of 48.5% at single-shot decoding and 89.7% under best-of-4 sampling. Verifiable-reward post-training with GRPO against a 660-entry sequence motif registry improves the useful-plasmid rate across all sampling budgets. We release the 19.3M-parameter model, evaluation suite, and a paired benchmark of prompt-sequence pairs.