The abstract is read at the publisher; the summary is JClub's.
CarmnaDD predicts nitrogenase activity from nitrogenase protein sequences alone, achieving R2 = 0.60 for regression and AUC = 0.94 with F1 = 0.86 for classification.
- Why it matters: Activity varies strongly across natural sequences, and wet-lab screening is slow and low throughput. Existing models improve prediction but often depend on embeddings and other features, limiting interpretability beyond feature contributions.
- What they did: The purely sequence-driven CarmnaDD uses the dual descriptor with learnable composition weight map (CWM) and position weight function (PWF) to generate a position-resolved description N(k). It is evaluated on the same dataset used by Carmna, alongside interpretability analyses including SHAP, sliding-window scanning (FDR < 0.05), and structural mapping to NifH interfaces.
- The result: Positional encoding adapts with rank, key CWM 3-mer fragments are enriched in polar uncharged residues, and sliding windows link activity to a D-X-E motif with R at the center showing the highest activity. Four correlated windows converge on NifH nucleotide-binding and conformational-transmission/protein-docking interfaces, with conservation negatively correlated with description intensity (rho = -0.30).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.