Prediction and Dissection of Nitrogenase Activity Based on Protein Sequences using Dual Descriptor

Nitrogenase reduces N2 to ammonia at ambient temperature and pressure, but its activity differs markedly among natural sequences, and conventional wet-laboratory screening suffers from low throughput and long turnaround times. Existing computational methods have partially alleviated this bottleneck, yet most of them rely on protein embeddings obtained from protein language models together with features such as gene expression, and their interpretability analyses mostly stop at the level of feature contribution. Here we present CarmnaDD, a purely sequence-driven framework built on the dual descriptor (DD): two learnable components, the composition weight map (CWM) and the position weight function (PWF), jointly generate a position-resolved description N(k), so that activity can be predicted without any prior knowledge beyond the sequence. On the same dataset used by Carmna, CarmnaDD attains R2 = 0.60 for regression and AUC = 0.94 and F1 = 0.86 for classification, better than our previous Carmna, which integrates multiple features. Interpretability analysis shows that the amplitude of the periodic terms of the PWF decreases as the rank increases, indicating that the positional encoding strategy adapts to the rank; the high-weight 3-mer fragments in the CWM are significantly enriched in polar uncharged residues, which corroborates the SHAP-based conclusion of Carmna; sliding-window scanning identifies fragments significantly associated with activity (FDR < 0.05), and the strongest window displays a D-X-E pattern with conserved flanks and a polymorphic center, sequences carrying R at the central position showing the highest activity; conservation is significantly negatively correlated (ρ = −0.30) with description intensity. Mapping the key windows onto the protein structure shows that four highly correlated windows converge on the nucleotide-binding interface and the conformational-transmission/protein-docking interface of NifH. These results provide a purely sequence-based framework for nitrogenase activity prediction that balances performance and interpretability.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-19
DOI
https://doi.org/10.64898/2026.11.18.767551
Primary Topic
Machine Learning in Bioinformatics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Prediction and Dissection of Nitrogenase Activity Based on Protein Sequences using Dual Descriptor

Bin‐Guang Ma, Xiao-Pan Hu, Gang Wu, Feng‐Biao Guo et al.
bioRxiv (Cold Spring Harbor Laboratory)
Machine Learning in Bioinformatics
preprint

Prediction and Dissection of Nitrogenase Activity Based on Protein Sequences using Dual Descriptor

Bin‐Guang Ma, Xiao-Pan Hu, Gang Wu, Feng‐Biao Guo, Yujia Zhao, Lin-Feng Wu
preprint en

Abstract

Nitrogenase reduces N2 to ammonia at ambient temperature and pressure, but its activity differs markedly among natural sequences, and conventional wet-laboratory screening suffers from low throughput and long turnaround times. Existing computational methods have partially alleviated this bottleneck, yet most of them rely on protein embeddings obtained from protein language models together with features such as gene expression, and their interpretability analyses mostly stop at the level of feature contribution. Here we present CarmnaDD, a purely sequence-driven framework built on the dual descriptor (DD): two learnable components, the composition weight map (CWM) and the position weight function (PWF), jointly generate a position-resolved description N(k), so that activity can be predicted without any prior knowledge beyond the sequence. On the same dataset used by Carmna, CarmnaDD attains R2 = 0.60 for regression and AUC = 0.94 and F1 = 0.86 for classification, better than our previous Carmna, which integrates multiple features. Interpretability analysis shows that the amplitude of the periodic terms of the PWF decreases as the rank increases, indicating that the positional encoding strategy adapts to the rank; the high-weight 3-mer fragments in the CWM are significantly enriched in polar uncharged residues, which corroborates the SHAP-based conclusion of Carmna; sliding-window scanning identifies fragments significantly associated with activity (FDR < 0.05), and the strongest window displays a D-X-E pattern with conserved flanks and a polymorphic center, sequences carrying R at the central position showing the highest activity; conservation is significantly negatively correlated (ρ = −0.30) with description intensity. Mapping the key windows onto the protein structure shows that four highly correlated windows converge on the nucleotide-binding interface and the conformational-transmission/protein-docking interface of NifH. These results provide a purely sequence-based framework for nitrogenase activity prediction that balances performance and interpretability.

bioRxiv (Cold Spring Harbor Laboratory)
Huazhong Agricultural University (CN), Wuhan University (CN)
Machine Learning in Bioinformatics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon