A Deep Learning Framework for Species-Specific Detection of ProtGPT2-Generated Synthetic DNA in Escherichia coli

The use of generative artificial intelligence in synthetic biology raises the need for strong methods to trace the origin of DNA sequences. In this study, we introduce a speciesspecific deep learning model that can classify DNA sequences produced by the protein language model ProtGPT2 in contrast to natural Escherichia coli K-12 genomic DNA. We produce 3,796 synthetic coding sequences by conditioning ProtGPT2 on natural N-terminal protein prompts and reversing translation to generate outputs with E. coli-optimized codon usage. We train a lightweight 1D convolutional neural network (CNN) to classify sequence origin, with naturally occurring sequences controlling for species identification and DNA source characteristics. The model achieved 100.0% accuracy and 1.000 AUC (Area Under the Curve) on a held-out test set, with 99.97% ± 0.05% accuracy via 5-fold cross validation. Using SHAP (SHapley Additive exPlanations) analysis, we show that the discriminative signal originates from effects on frequency of codon pairs and hexamers—statistical manufacts stemming from autoregressive language models. This work provides a framework for synthetic DNA forensics with implications for biosecurity and regulatory oversight in the era of AI-powered genome design.