#2 · Genomics

Transformer and deep learning tools benchmarked against established genomics methods

The dominant thread is evaluation of deep learning models against established baselines across genomic and proteomic tasks. Metaxa compares against Kraken 2 and MetageNN on long Nanopore reads, with genus-level gains on out-of-sample taxa. Tiberius is extended to lineage-specific ab initio gene models across six eukaryotic clades and benchmarked on 33 species, while OrionGeno reports outperformance over existing annotation methods. DDA-BERT improves peptide identification over existing rescoring tools, and metagenomic-deepFRI adds structural templates to protein function annotation. OGR pairs a 1.25B-parameter MoE foundation model trained on 422 rice genomes with an open benchmark suite. Not every item shares this emphasis: TaxVAMB addresses metagenome binning through a semisupervised variational autoencoder. What distinguishes this set is explicit testing of generalization across clades, out-of-sample taxa, and species, alongside released benchmark infrastructure.