High-quality phage assembly from metagenomes with PALACE

Millions of phage genomes have been mined from metagenomic data recently but the genome completeness remains poor because of the limitations of existing phage detection methods, which rely on metagenomic contigs that fragment phage genomes. Here, we present PALACE, a conjugate-graph-based framework for assembling high-quality phage genomes from metagenomes. PALACE incorporates homology-based and deep-learning-based methods to detect phage signals and constructs a conjugate graph from the metagenomic sample. On simulated data, PALACE generates accurate and complete phage genomes, achieving an F₁ score of 0.92-1.00 across simulation settings, outperforming the second-best method by 0.21-0.48. Applying PALACE to 914 gut metagenomic samples from healthy controls and participants with colorectal cancer (CRC) yielded 5,306 high-quality phage genomes, outperforming the second-best benchmark method by 55.98% in median genome completeness. We observed a high degree of functional organization for genes within phage genomes. Phages from participants with CRC exhibited a notable enrichment of metabolic factors, suggesting their adaptation to nutrient availability in the CRC gut environment.