[レビュー] CRISPR/Cas9における標的部位活性, 非標的部位リスク, および修復結果の深層学習による予測：現状とマルチスケールな視点 : crisp

　CRISPR/Cas9システムは、ゲノム編集における革新的なツールとして登場し、精密な遺伝子工学を実現する上で極めて重要な役割を果たしています。標的部位活性を最大化しつつ非標的部位活性（オフターゲット編集）を最小限に抑えることは、CRISPR/Cas9を信頼性の高い実験および治療応用へと展開する上で不可欠です。従来のオフターゲット編集の検出方法は、労力とコストがかかり、拡張性に限界がありましたが、人工知能の登場により、検出コストは大幅に削減され、スループットも大幅に向上しました。

　CRISPR/Cas9分野における初期の浅層学習（shallow learning）モデルは、基本的な分類タスクには有効でしたが、特徴表現が限定的で汎化性能が低いという課題がありました。その後、アルゴリズムと計算能力の進歩に伴い、深層学習アーキテクチャがオフターゲット編集の予測精度を飛躍的に向上させました。

　しかし、依然として重要な盲点が残っています。現在のモデルのほとんどは主に配列レベルで動作しており、ゲノム編集の下流における機能的影響を見落としているのです。

　今回、中国の研究チームによるレビューにおいて、AIを活用したCRISPR/Cas9予測手法の現状が概観され、分子、細胞、組織の3つの次元を統合した将来を見据えた「3層フレームワーク」が提案されています。このフレームワークは、ヌクレオチドレベルの編集をタンパク質の変化、細胞機能の変化、組織特異的な応答と関連付けることで、配列に基づく予測と表現型の結果との間のギャップを埋め、CRISPR/Cas9技術の精度と臨床応用可能性を高めることを目指しています。

[構成] (ARTICLE IN PRESSの版（一段組）で参考文献238件を含む69頁

1. Introduction
2. On-Target Activity
2.1 On-Target Efficacy in CRISPR-Cas9 System
2.2 Experimental Approaches in On-Target Task
2.2.1 Current Biological Experimental Approaches for Assessing On-Target Activity
2.2.2 Public Databases and Libraries for On-Target sgRNA Profiling
2.3 Computational Approaches for Assessing On-Target Activity
2.4 Non-Deep Learning Approaches for On-Target Prediction
2.5 Deep Learning Approaches for On-Target Prediction
2.5.1 CNN
2.5.2 RNN
2.5.3 CNN-RNN
2.5.4 Transformer
3. Off-Target Effect
3.1 Overview of Off-target
3.1.1 Off-target in CRISPR-Cas9 System
3.2 From experimental approaches to in silico predction in off-target detection
3.2.1 Cell-Based Experimental Methods
3.2.2 Cell-Free Experimental Methods
3.2.3 Detection of Cas9 Binding
3.2.4 Detection of Cas9-Induced DSB
3.2.5 Detection of Cas9 Repair Products
3.3 Current in silico Prediction Methods for CRISPR-Cas9 Off-Target Activity
3.3.1 Rule-Based Approaches for Off-Target Prediction in CRISPR-Cas9 Systems
3.3.2 Energy-Based Methods
3.3.3 Shallow learning-Based Models for CRISPR-Cas9 Off-Target Prediction
3.4 Deep Learning
3.4.1 CNNs for Predicting Off-Target Effects in CRISPR-Cas9 Systems
3.4.2 RNNs for CRISPR-Cas9 Off-Target Effect Modeling
3.4.3 Hybrid CNN–RNN Architectures for CRISPR-Cas9 Off-Target Prediction
3.4.4 Transformer-Based Architectures for CRISPR-Cas9 Off-Target Prediction
3.5 Forward-Looking Solution to Overcome Existing Limitations
4.Outcome
4.1 Predicting Repair Outcomes Enhances Control in CRISPR Editing
4.2 Understanding DNA Repair Pattern Regularity Improves Outcome Prediction and sgRNA Design
4.3 Data Limitations and Feature-Driven Modeling of CRISPR Repair Outcomes
5. Feature
5.1 Difference in Sequence Features Between On-target and Off-target Prediction Tasks
5.2 Feature Engineering Strategies for Modeling CRISPR-Cas9 Activity
5.2.1 Epigenetic Features
5.2.2 Contextual Information
5.2.3 Structural Features of sgRNA
5.2.4 Application of DNA/RNA-Pre-trained Language Models in CRISPR-Cas9 Sequence Modeling
5.2.5 Interaction Features and Thermodynamic Contributions
6. Discussion
6.1 Structural blind spots in current CRISPR prediction models
6.2 Data imbalance, labeling inconsistency, and limited representativeness
6. 3 Root causes of current modeling limitations: technical and data bottlenecks
6.4 A multi-scale analytical perspective for CRISPR/Cas9 prediction
6.5 Future directions: multi-omics integration and low-data learning
6.6 Outlook toward biologically grounded and clinically translatable CRISPR modeling.

[図表リスト]

Figure 1. An overview of CRISPR-Cas9 editing pipeline.
Figure 2. The timeline of CRISPR-Cas9 studies.
Figure 3. Mechanistic Cascade of CRISPR-Cas9 Editing from Sequence to System.
Table 1. Representative Deep Learning Models for CRISPR-Cas9 On-Target Activity Prediction
Table 2. Experimental Methods for Assessing CRISPR-Cas9 On-Target Editing Efficiency
Table 3. Public Databases and Screening Resources for CRISPR-Cas9 On-Target Activity Analysis
Table 4. Public Databases and Computational Resources for CRISPR-Cas9 Off-Target Site Prediction and Analysis
Table 5. Experimental Methods for Detecting Cas9-Induced DNA Double-Strand Breaks: Principles, Advantages, and Limitations
Table 6. Computational Models for CRISPR-Cas9 Off-Target Prediction: Algorithms, Input Features, and Training Datasets
Table 7. Computational Models for Predicting Indel Outcomes After CRISPR-Cas9 Editing
Table 8.Feature Representations, Model Architectures, and Training Data Used in CRISPR-Cas9 On-Target Prediction Models

[出典]

Review "Deep learning–driven prediction of on-target activity, off-target risk, and repair outcomes in CRISPR/Cas9: current landscape and multi-scale perspectives" Du W, Zhang T [..] Hu H, Chen L, Liu C. J Transl Med 2026-04-28. https://doi.org/10.1186/s12967-026-08175-1 [所属] Jinan University (中国), Foshan University, The First Affiliated Hospital/Southern University of Science and Technology, Lund University (スエーデン), Southern Medical University