Background

Bacteriophage ΦX174 is a prototypical single-stranded DNA virus that infects Escherichia coli, with a genome consisting of 5,386 nucleotides and 11 proteins. Its complete genome sequence was first decoded in the 1970s, and in the 2000s, it became the first genome to be chemically synthesized. Recently, it has regained attention as a model system in synthetic biology, as genome language models have used ΦX174 to design infectious phages.

However, the ability to read or generate nucleotide sequences does not necessarily equate to understanding the biological consequences of mutations. Existing variant effect predictors (VEPs) rely on evolutionary conservation, protein structure, and statistical rules learned from large-scale sequence data. In systems with limited sequence data and tightly interacting proteins, such as viruses, their performance has not been sufficiently validated.

Key Findings

Weijun Hui, Shanghua Li, and Ben Lehner conducted a comprehensive mutagenesis analysis by substituting each nucleotide in the ΦX174 genome with the other three possible nucleotides and each amino acid in the proteins with the remaining 19 types. The resulting mutant library exceeded 44,000 variants. The researchers cultured these variants with susceptible E. coli for 80 minutes. During one or two infection cycles, they distinguished between proliferating and extinct variants using DNA sequencing to calculate fitness.

More than half of all single-nucleotide variants and over 60% of amino acid substitutions reduced phage fitness. A small number of variants were found to enhance the replication of ΦX174, which was already considered optimized for laboratory conditions. Among harmful amino acid variants, 47.6% were located at interaction interfaces between phage proteins or between phage and host proteins. Variants that disrupted the core regions of proteins accounted for 25.4%. The remaining approximately one-quarter could not be explained by known structural or interaction data.

The researchers compared the experimental results with several state-of-the-art protein language models, including ESM-1v, ESM-2, ESM3, ProGen2, Tranception-L, GEMME, and SaProt, as well as structure- and evolution-based predictors. The predictive power of these models remained generally moderate. A simple structural metric, relative solvent accessibility calculated in protein complexes, achieved a median Spearman correlation of 0.41, outperforming the 0.26 correlation from monomer structural features. Complex AI models did not consistently surpass the basic structural information of interaction interfaces.

Implications and Outlook

These results highlight a significant gap between the ability to generate biologically plausible sequences and the ability to causally predict the phenotypic effects of specific mutations. In particular, focusing solely on protein folding may overlook the disruption of intermolecular interactions, which are a primary cause of mutational damage. Mutation interpretation models must incorporate not only single-protein sequences but also complex structures, host factors, and phenotype data across infection stages.

This dataset could serve as a genome-wide benchmark for evaluating viral mutation prediction and phage design models. However, the target is a small model phage, and fitness was measured in an 80-minute competitive culture using E. coli. The effects of mutations may change with different host species or culture conditions. Additionally, the study is a preprint that has not undergone peer review, and this should be considered in interpretation.