The MODNAP benchmark set for modeling ribose-based modified nucleic acid and nucleic acid-protein complexes: a case study with AlphaFold 3.

AlphaFold 3 achieves over 80% success in modeling 1039 nucleic acid and nucleic acid-protein structures with modified nucleotides, but accuracy varies for modifications not in training data.

• Why it matters: Understanding the structural diversity and interactions of modified nucleic acids is crucial for advances in biology and biotechnology, yet their modeling performance remains unassessed.
• What they did: The authors created MODNAP, a curated dataset of experimentally determined structures with modified nucleotides, and benchmarked AlphaFold 3 across diverse structures, modifications, and functions.
• The result: While AlphaFold 3 performs well overall, its variable accuracy and frequent intramolecular geometry failures highlight the need for improved methods, with MODNAP serving as a key resource for development.