Tomer Sidi

BBERT is a nucleotide large language model that accurately assigns taxonomic domains to short metagenomic reads without reference databases, improving assembly and downstream analysis.

The paper trains AI models to predict codon sequences from amino acid sequences in yeast and bacteria, showing they outperform frequency-based methods and reveal codon usage patterns linked to evolutionary pressures.

The paper evaluates different loss functions for improving the accuracy of protein model ranking using the MESHI-score method.

The paper describes methods for estimating model accuracy in protein structure predictions, showing improvements in single model accuracy over previous methods.