AcrNET: predicting anti-CRISPR with deep learning.

Area of Science:

• Biochemistry
• Bioinformatics
• Genetics

Background:

• Anti-CRISPR proteins (Acrs) inhibit bacterial CRISPR-Cas immune systems, presenting therapeutic potential.
• Predicting and discovering Acrs is challenging due to high variability and rapid evolution.
• Current computational methods for Acr prediction exhibit limitations in performance.

Purpose of the Study:

• To develop a novel deep neural network, AcrNET, for enhanced anti-CRISPR protein analysis.
• To overcome data scarcity and improve prediction accuracy for Acrs.
• To enable prediction of detailed anti-CRISPR classes for mechanistic insights.

Main Methods:

• A deep neural network (AcrNET) integrating Transformer protein language model (ESM-1b) features.
• Utilizing evolutionary and local structure features alongside Transformer embeddings.
• Cross-fold and cross-dataset validation to assess performance against state-of-the-art methods.

Main Results:

• AcrNET significantly outperforms existing methods, improving F1 score by at least 15% in cross-dataset tests.
• AcrNET is the first method capable of predicting detailed anti-CRISPR classes.
• Analysis confirms complementary contributions of Transformer, evolutionary, and structural features.
• Further experiments (AlphaFold, motif analysis, docking) validate AcrNET's ability to capture conserved patterns and interactions.

Conclusions:

• AcrNET offers a powerful computational approach for anti-CRISPR protein discovery and analysis.
• The method addresses key challenges in Acr prediction, including data scarcity and protein variability.
• AcrNET's ability to predict Acr classes provides insights into their mechanisms and facilitates applications in gene editing and phage therapy.