Related Experiment Video
Updated: Jun 12, 2025
09:01
Phage-Mediated Genetic Manipulation of the Lyme Disease Spirochete Borrelia burgdorferi
Published on: September 28, 2022
2.0K
Transformer model generated bacteriophage genomes are compositionally distinct from natural sequences.
1Johns Hopkins University Applied Physics Laboratory, 11000 Johns Hopkins Road, 20723 Maryland, Laurel, MD 20723, USA.
NAR Genomics and Bioinformatics
|September 19, 2024
Summary
The megaDNA model creates synthetic viral genomes but lacks realistic genomic composition. Genome composition analysis can reliably detect these AI-generated sequences, highlighting areas for improvement in generative models.
Area of Science:
- Genomics and Bioinformatics
- Computational Biology
- Artificial Intelligence in Biology
Background:
- Language models are emerging as powerful tools in genomics research.
- The megaDNA model is the first publicly available generative model for synthetic viral genomes.
- Evaluating the biological realism and detectability of AI-generated genomes is crucial.
Purpose of the Study:
- To assess megaDNA's ability to generate synthetic bacteriophage genomes with realistic compositional biases.
- To determine if transformer-generated viral genomes can be algorithmically detected.
- To evaluate a framework for assessing generative genomic models.
Main Methods:
- Comparative analysis of compositional metrics between 4969 natural and 1002 synthetic bacteriophage genomes.
- Utilized rank-sum tests and principal component analyses to compare genomic features.
- Trained a neural network on compositional metrics to detect transformer-generated sequences.
Main Results:
- Synthetic genomes exhibited realistic lengths, with 58% classified as viral by geNomad.
- Significant differences in compositional metrics were observed between natural and synthetic genomes.
- A neural network achieved high detection accuracy (93.0% sensitivity, 97.9% specificity) for AI-generated sequences.
Conclusions:
- The megaDNA model currently does not produce bacteriophage genomes with authentic compositional biases.
- Genome composition serves as a reliable indicator for identifying megaDNA-generated sequences.
- The developed evaluation framework is applicable to various genomic sequence generative models.
Related Concept Videos
01:30Lytic Cycle of Bacteriophages
70.5K
Bacteriophages, also known as phages, are specialized viruses that infect bacteria. A key characteristic of phages is their distinctive “head-tail” morphology. A phage begins the infection process (i.e., lytic cycle) by attaching to the outside of a bacterial cell. Attachment is accomplished via proteins in the phage tail that bind to specific receptor proteins on the outer surface of the bacterium. The tail injects the phage’s DNA genome into the bacterial cytoplasm. In the...
70.5K
00:43Lysogenic Cycle of Bacteriophages
62.0K
In contrast to the lytic cycle, phages infecting bacteria via the lysogenic cycle do not immediately kill their host cell. Instead, they combine their genome with the host genome, allowing the bacteria to replicate the phage DNA along with the bacterial genome. The incorporated copy of the phage genome is called the prophage. Some prophages can re-activate and enter the lytic cycle. This often occurs in response to a perturbation, such as DNA damage, but can also transpire in the absence of...
62.0K
02:53CRISPR and crRNAs
16.9K
Bacteria and archaea are susceptible to viral infections just like eukaryotes; therefore, they have developed a unique adaptive immune system to protect themselves. Clustered regularly interspaced short palindromic repeats and CRISPR-associated proteins (CRISPR-Cas) are present in more than 45% of known bacteria and 90% of known archaea.
The CRISPR-Cas system stores a copy of foreign DNA in the host genome and uses it to identify the foreign DNA upon reinfection. CRISPR-Cas has three different...
The CRISPR-Cas system stores a copy of foreign DNA in the host genome and uses it to identify the foreign DNA upon reinfection. CRISPR-Cas has three different...
16.9K