CheckM2

CheckM2 is a machine-learning tool for rapid, accurate quality assessment of metagenome-assembled genomes (MAGs), providing predictions of genome completeness and contamination prior to downstream analysis. By learning quality-associated patterns from large reference genome collections, it improves accuracy and speed relative to traditional marker-based approaches and remains reliable for phylogenetically novel lineages and reduced genomes (e.g., Patescibacteria and the DPANN superphylum).

The software uses two distinct completeness models:

• General model (gradient boosting) — designed to generalize well to organisms that are poorly represented in reference databases (approximately novel at the order/class/phylum level).
• Specific model (neural network) — optimized for higher accuracy when genomes are closely related to the training references (approximately within known species/genus/family).

CheckM2 automatically selects the most appropriate completeness model for each input genome using a cosine similarity criterion, while also allowing users to force a particular model or output predictions from both. Contamination is estimated using a single gradient-boosting model applied across all taxa. Its reference database can be rapidly updated with new high-quality genomes, including taxa represented by only one genome, supporting robust performance as reference collections expand.