DeepCas12a: A hybrid deep learning framework for accurate Cas12a efficiency prediction from sequence and epigenetic information

DeepCas12a: A hybrid deep learning framework for accurate Cas12a efficiency prediction from sequence and epigenetic information | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article DeepCas12a: A hybrid deep learning framework for accurate Cas12a efficiency prediction from sequence and epigenetic information Yiming Shi, Junkai Yin, Shurui Ning, Jinling Yuan, Degang Yang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8736163/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract CRISPR-Cas12a (Cpf1) offers distinct advantages for genome editing due to its flexible, T-rich PAM recognition. However, variable cleavage efficiency—modulated by sequence context and epigenetic features—remains a challenge, with existing predictors limited in accuracy and interpretability. Here, we present DeepCas12a, a hybrid deep learning framework integrating Convolutional Neural Networks (CNNs) and a Vision Transformer (ViT) encoder to capture both local sequence motifs and long-range dependencies. The model fuses DNA sequence data with epigenetic profiles (DNA methylation and chromatin accessibility) in an end-to-end architecture. Benchmarked on an independent test set, DeepCas12a outperformed state-of-the-art predictors, achieving an Average Precision of 0.783, an AUC of 0.868, and a Spearman correlation of 0.630. Furthermore, interpretability analysis via saliency maps confirms the model captures biologically relevant features, including PAM specificity and seed region sensitivity, facilitating rational guide RNA design. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Since its initial application in 2012[ 1 ], CRISPR genome editing has demonstrated transformative potential across agriculture, medicine, and basic research. Among these, the type II CRISPR-Cas9 system has emerged as the dominant tool due to its high editing efficiency. However, limitations such as strict NGG protospacer adjacent motif (PAM) dependence and off-target effects restrict its utility in complex genomic contexts[ 2 ]. Consequently, the type V CRISPR-Cas12a system has gained traction as an alternative, offering distinct mechanistic advantages that address the core limitations of Cas9[ 3 ]. Unlike Cas9, Cas12a recognizes a T-rich PAM (TTTV), significantly expanding the accessible genomic target space[ 4 – 6 ]. Furthermore, its reliance on a single crRNA simplifies construct design. Additionally, Cas12a exhibits collateral cleavage activity triggered upon target binding, enabling single-molecule nucleic acid detection[ 7 , 8 ] —a feature exploited for rapid pathogen diagnostics, including SARS-CoV-2 detection[ 9 ]. Despite these advantages, Cas12a orthologs display significant heterogeneity in cleavage efficiency, with variation exceeding an order of magnitude across different guide RNAs[ 10 ]. Current efficiency prediction tools, such as DeepCpf1, have achieved only moderate predictive performance[ 11 ]. This performance gap stems from the mechanistic complexity of Cas12a, where efficiency is modulated by crRNA secondary structure[ 12 ], local epigenetic landscapes, and sequence features including PAM-proximal GC content[ 13 ]. While robust models exist for Cas9, direct transfer learning to Cas12a results in a substantial performance decline (AUC decrease > 10%)[ 14 ], necessitating the development of dedicated Cas12a predictors. To address these challenges, we propose DeepCas12a, a deep learning framework utilizing a dual-stream hybrid architecture (Fig. 1 ). This model integrates local feature extraction via Convolutional Neural Networks (CNNs) and global sequence modeling via a Transformer-based attention mechanism. Specifically, parallel convolutional layers capture motif preferences in PAM-proximal regions, while a Transformer encoder models long-range interactions—such as mismatch tolerance—within the crRNA–target DNA duplex. Benchmarking results demonstrate that DeepCas12a significantly outperforms existing tools in prediction accuracy on independent test sets. Furthermore, we provide interpretable saliency maps to elucidate the mechanistic basis of Cas12a activity. This framework offers practical guidance for therapeutic gene editing design—exemplified by targets for sickle cell disease—thereby reducing experimental screening costs. 2. Material and Methods 2 .1 Data Preparation CRISPR-Cas12a cleavage efficiency data were retrieved from Kim et al. (2018)[11]. This dataset (HT1) quantifies Cas12a editing efficiency across target sites in HEK293T cells via high-throughput sequencing. The complete dataset (N=16,292) was randomly partitioned into a training set (n=15,000) and an independent test set (n=1,292). For binary classification, labels were assigned based on background-corrected indel frequencies: samples exhibiting a >2-fold change over controls and an absolute indel frequency >50% were labeled positive (1); all others were labeled negative (0). This threshold aligns with high-efficacy classification standards used in prior CRISPR studies. To model epigenetic effects, DNA methylation (RRBS) and chromatin accessibility (DNase-seq) data for HEK293T cells were obtained from the ENCODE project[15]. Target sequences (34 bp) were aligned to the human reference genome (hg38) using Bowtie 2[16]. We identified target sites overlapping with RRBS-covered regions or DNase narrow peaks to generate binary epigenetic feature maps (presence=1, absence=0)[17]. Input sequences were encoded as 6-channel tensors. Channels 1–4 represented nucleotide identity via one-hot encoding (A, C, G, T). Channel 5 encoded DNA methylation status, and Channel 6 encoded chromatin accessibility, as shown in Figure 2. This representation captures both local sequence context and the epigenetic landscape, reflecting the multifactorial regulation of Cas12a activity. 2.2 Model Architecture DeepCas12a utilizes a hybrid neural network architecture to predict CRISPR-Cas12a cleavage efficiency. The framework comprises three primary modules: a Convolutional Neural Network (CNN) for local feature extraction, a Vision Transformer encoder (ViT) for global context modeling, and a Multi-Layer Perceptron (MLP) classifier. Input tensors (6×34) are processed by the CNN module to extract local motif definitions and epigenetic patterns. Convolutional feature maps are tokenized into embeddings, augmented with positional encodings, and input into the vision transformer module. A multi-head self-attention mechanism models long-range dependencies, capturing global regulatory interactions within the sequence. This design adapts the vision transformer strategy [18] to genomic data, leveraging the efficacy of attention mechanisms in sequence modeling demonstrated by architectures such as DNABERT [19]. Finally, the global class token ([CLS]) is fed into the MLP to output a probability score representing binary cleavage efficiency. 2.3 Experimental Setup and Training Parameters DeepCas12a was implemented in Python using the PyTorch framework. Training was accelerated via CUDA 12.4 on a workstation equipped with dual NVIDIA GeForce RTX 3090 GPUs (24 GB VRAM). Hyperparameters were systematically optimized to maximize validation performance. We employed Bayesian optimization to tune the learning rate, attention/global dropout rates, network depth, embedding dimension, MLP expansion ratio, and the number of attention heads. The final hyperparameter configuration is detailed in Table 1. To mitigate overfitting, we applied early stopping and assessed model robustness via k-fold cross-validation on the training data. Table 1 Selected model hyperparameters: Hyperparameter Value attention_drop 0.271 dropout 0.307 depth 10 embed_dim 256 learning rate 0.00027 mlp_ratio 2.45 num_heads 12 DeepCas12a was benchmarked against four representative Cas12a efficiency predictors, as shown in Table 2: DeepCpf1[11]: A CNN-based end-to-end predictor utilizing sgRNA sequences; CRISPR-DT[20]: An SVM classifier integrating sequence features with manually engineered biological characteristics; C-SVR[21]: A hybrid architecture combining CNN-based feature extraction with Support Vector Regression (SVR); Cas12a_predictor[22]: A Random Forest model incorporating sgRNA sequences and biological features. Table 2 Summary of mainstream Cas12a efficiency prediction models, their references, and implementation sources: Model Name Reference Method Type Code Availability DeepCpf1 1 Kim et al., 2018 CNN GitHub CRISPR-DT 2 Zhu et al., 2019 Feature Engineering + SVM Paper Reproduction C-SVR 3 Ameen et al., 2021 CNN + SVR Paper Reproduction Cas12a_predictor 4 O’Brien et al., 2023 Feature Engineering + RandomForest GitHub Source: Model results are obtained from either the original publications or official code repositories using the same dataset and evaluation pipeline for consistency. 1 Kim et al. (2018), Nature Biotechnology. Deep learning improves prediction of CRISPR–Cpf1 guide RNA activity.[11] 2 Zhu et al. (2019), Bioinformatics. CRISPR-DT: designing gRNAs for the CRISPR-Cpf1 system with improved target efficiency and specificity.[20] 3 Ameen et al. (2021), Alexandria Engineering Journal. C-SVR Crispr: Prediction of CRISPR/Cas12 guideRNA activity using deep learning models.[21] 4 O’Brien et al. (2023), PLOS ONE. Predicting CRISPR-Cas12a guide efficiency for targeting using machine learning.[22] 3. Results 3.1 Model performance on holdout test set We evaluated DeepCas12a on the independent test set described in Section 2.1. The model achieved an Average Precision (AP) of 0.783 and an Area Under the ROC Curve (AUC) of 0.868. To assess quantitative predictive accuracy, we computed Spearman (ρ) and Pearson (r) correlation coefficients between predicted probabilities and experimental cleavage scores, yielding 0.630 and 0.631, respectively. These correlations indicate that DeepCas12a captures both the rank order and magnitude of cleavage efficiency, supporting its utility for quantitative prediction. 3.2 Comparison with current sgRNA efficiency prediction models To evaluate DeepCas12a, we benchmarked it against four state-of-the-art Cas12a efficiency predictors: DeepCpf1, CRISPR-DT, C-SVR, and Cas12a_predictor. DeepCpf1 and C-SVR rely on sequence-derived deep learning features, whereas CRISPR-DT and Cas12a_predictor utilize handcrafted biological features within traditional machine learning frameworks. Baseline models were evaluated on the independent test set using the authors' public implementations and default hyperparameters to ensure fair comparison. In binary classification tasks, DeepCas12a achieved the highest performance metrics (AP = 0.783, AUC = 0.868; Fig. 3c–d). It outperformed the second-best model, DeepCpf1 (AP = 0.755, AUC = 0.858), while other baselines showed significantly lower discrimination. To assess quantitative ranking accuracy, we computed Spearman rank (ρ) and Pearson linear (r) correlations between predicted and experimental efficiencies. DeepCas12a yielded the highest correlation coefficients (ρ = 0.630, r = 0.631). These values exceed those of DeepCpf1 (ρ = 0.612, r = 0.594) and other baselines (Fig. 3a–b), indicating a more accurate global ordering of target efficacy. In summary, DeepCas12a demonstrates consistent improvements over existing state-of-the-art methods in both classification and regression tasks. 3.3 Generalization evaluation on independent test sets To assess generalization beyond the primary test set, we evaluated DeepCas12a on extra two independent datasets, HT2 (n = 2,963) and HT3 (n = 1,251), which were excluded from model training. These datasets were also derived from the Kim et al. (2018) study[11] DeepCas12a demonstrated superior correlation and classification performance across both datasets compared to baseline models (Fig. 4). Specifically, it achieved the highest rank correlations with experimental cleavage efficiencies: Spearman coefficients (ρ) were 0.538 (HT2) and 0.313 (HT3), while Pearson coefficients (r) were 0.540 (HT2) and 0.294 (HT3) (Fig. 4a–b). These results suggest that the model generalizes effectively to unseen data. In classification tasks, DeepCas12a consistently showed the highest discrimination (Fig. 4c–f). On HT2, it achieved an AUC of 0.823 and Average Precision (AP) of 0.697, marginally surpassing DeepCpf1 (AUC = 0.819) and significantly outperforming C-SVR (AUC = 0.751). A similar trend was observed on HT3, where DeepCas12a maintained leading performance (AUC = 0.726, AP = 0.360). These results demonstrate the model's robustness to dataset shifts. 3.4 Ablation study We performed ablation studies on the independent test set to quantify the contributions of specific feature types and architectural components. Five variants were evaluated: (1) DeepCas12a (Full); (2) Seq-Only (excluding epigenetic features); (3) CNN-Only (excluding the Transformer); (4) 50 bp context; and (5) 20 bp guide-only. The full model achieved the highest correlation with experimental cleavage efficiencies (ρ = 0.630, r = 0.631; Fig. 5a–b). Removing epigenetic features resulted in a minor performance decline (ρ = 0.621, r = 0.621), suggesting that chromatin accessibility and methylation data provide limited but complementary information. Regarding input length, the 34 bp window (4 bp upstream, PAM, 23 bp protospacer, 3 bp downstream) proved optimal, outperforming 50 bp and 20 bp configurations. Extending the window to 50 bp likely introduced noise, while restricting it to the 20 bp guide sequence failed to capture essential flanking signals. DeepCas12a also led in classification metrics (AUC = 0.868, AP = 0.783; Fig. 5c–d). Removal of the Transformer encoder caused a distinct decrease in performance, supporting the utility of modeling long-range dependencies. These results indicate that sequence context, epigenetic features, and the hybrid CNN-Transformer architecture collectively contribute to predictive accuracy. 3.5 Interpretability analysis via saliency maps We visualized multi-feature saliency heatmaps for representative high-efficiency (positive) and low-efficiency (negative) targets to elucidate feature contributions (Fig. 6a–b). Red regions indicate positive attribution (promoting cleavage), while blue regions indicate negative attribution (inhibiting cleavage). (1) PAM Recognition: In high-efficiency targets (Fig. 6a), the PAM core (positions 5–7) exhibits contiguous positive attribution for Thymine, whereas low-efficiency targets (Fig. 6b) display inhibitory signals at these sites. This confirms the model successfully learned the preference for a T-rich PAM. At position 8, the model assigns positive weights to non-T bases (A/C/G) and negative weights to T, reflecting the biological requirement for a V (A/C/G) nucleotide at the terminal PAM position[23]. (2) Seed Region Sensitivity: High-efficiency targets display concentrated positive saliency within the protospacer "seed" region (positions 9–14). This aligns with experimentally validated seed regions where mismatches drastically reduce Cas12a activity[24, 25]. Conversely, low-efficiency targets show dispersed, weak saliency, indicative of ineffective guide–target pairing. Attribution magnitudes diminish beyond position 25, consistent with the minimal impact of distal nucleotides on cleavage. (3) Epigenetic Modulation: The DNase accessibility channel predominantly shows positive attribution (red) in high-efficiency targets—particularly PAM-proximal regions—suggesting open chromatin enhances accessibility. In contrast, the DNA methylation (RRBS) channel appears predominantly negative (blue), capturing the repressive effect of local methylation. These patterns indicate that DeepCas12a has learned biologically relevant rules, including PAM specificity, seed region sensitivity, distal tolerance, and the inhibitory influence of DNA methylation. 4. Discussion DeepCas12a demonstrates improved accuracy in predicting CRISPR-Cas12a cleavage efficiency. It outperforms existing baselines (DeepCpf1, CRISPR-DT, C-SVR, Cas12a_predictor) in both classification (AP, AUC) and ranking (ρ, r) metrics. The integration of local convolutional features with global attention mechanisms enables robust generalization, effectively capturing multi-level regulatory signals. By integrating epigenetic features (DNA methylation, chromatin accessibility), the architecture maintains stability where sequence-only models may falter, highlighting the value of multimodal integration in functional genomics. Furthermore, saliency analysis confirms that DeepCas12a independently identifies critical functional regions, including the PAM (positions 5–8) and the seed region (positions 9–14). Gradient distributions align with established biological constraints, such as the requirement for a terminal "V" in the TTTV PAM and mismatch tolerance in distal regions. Practically, researchers can prioritize targets with high predicted efficacy and favorable saliency profiles to improve experimental validation rates. This approach is particularly valuable for therapeutic design, where identifying high-activity sites minimizes resource expenditure. Additionally, this "interpretation–validation" loop facilitates hypothesis generation—such as verifying causal links between PAM-proximal features and activity—accelerating precision functional genomics. Despite these advances, limitations persist. First, complex cellular contexts—including cell-type specific chromatin states and transcription factor binding—may limit generalization to unseen tissues. Second, while the hybrid architecture improves representation, reliance on static high-throughput datasets risks overfitting to library-specific biases. Future work should incorporate diverse datasets from primary cells and clinical samples to evaluate cross-condition adaptability. Moreover, current reliance on bulk ENCODE data lacks temporal resolution; integrating single-cell multi-omics (e.g., scATAC-seq) could refine epigenetic modeling. The Vision transformer architecture also positions the framework for future generative design and causal inference tasks. Generative extensions (e.g., diffusion models) could enable the de novo design of optimal guide RNAs under functional constraints[ 26 ]. Furthermore, attention mechanisms offer pathways to uncover regulatory factors, facilitating a shift from correlation discovery to causal mechanism elucidation[ 27 , 28 ]. In addition, the combination of pre-training and self-supervised learning strategies is expected to enhance the model’s generalization ability on rare sequences or novel enzyme variants, promoting a positive feedback loop between high-throughput virtual screening and experimental validation and accelerating innovation in CRISPR applications. In summary, DeepCas12a provides an accurate, interpretable solution for Cas12a efficiency prediction, establishing a foundation for intelligent gene editing design and mechanistic research. 5. Conclusion In this study, we present DeepCas12a, a multimodal deep learning framework integrating sequence and epigenetic contexts to predict CRISPR-Cas12a cleavage efficiency. By synergizing local convolutional feature extraction with global attention mechanisms, the model achieves state-of-the-art performance, robustness, and biological interpretability. DeepCas12a offers an effective tool for high-throughput target screening and rational guide RNA design, facilitating applications in precision gene editing and therapeutic development. Future integration of diverse multi-omics datasets and generative modeling strategies will further expand the utility of deep learning in CRISPR functional genomics. Declarations Ethics, Consent to Participate, and Consent to Publish declarations: not applicable. Funding Declaration This study was supported by National Natural Science Foundation of China [62002265]; Tongji University "Medicine + X" Cross Research Program[2025080107]. Data availability The datasets analyzed during the current study are available on the Github at https://github.com/bm2-lab-submission/DeepCas12a. Code availability The model and model usage are available on the Github at https://github.com/bm2-lab-submission/DeepCas12a. Author Contribution Y.S., J.K.Y., S.N., J.L.Y., D.Y. and G.C. conceived and designed the study. Y.S. designed and trained the model. S.N. implemented the benchmarking. Y.S. and J.L.Y. analyzed the results and generated the figures. Y.S., J.K.Y., D.Y. and G.C. wrote the manuscript. G.C. and D.Y. supervised the entire project. All authors read and approved the final manuscript. References M. Jinek, K. Chylinski, I. Fonfara, et al., A Programmable Dual-RNA-Guided DNA Endonuclease in Adaptive Bacterial Immunity, Science 337(6096) (2012) 816–821. https://doi.org/10.1126/science.1225829. D. Kim, S. Kim, S. Kim, et al., Genome-wide target specificities of CRISPR-Cas9 nucleases revealed by multiplex Digenome-seq, Genome Research 26(3) (2016) 406–415. https://doi.org/10.1101/gr.199588.115. D.C. Swarts, M. Jinek, Mechanistic Insights into the cis- and trans-Acting DNase Activities of Cas12a, Molecular Cell 73(3) (2019) 589–+. https://doi.org/10.1016/j.molcel.2018.11.021. B. Zetsche, J.S. Gootenberg, O.O. Abudayyeh, et al., Cpf1 Is a Single RNA-Guided Endonuclease of a Class 2 CRISPR-Cas System, Cell 163(3) (2015) 759–771. https://doi.org/10.1016/j.cell.2015.09.038. S. Stella, P. Alcon, G. Montoya, Structure of the Cpf1 endonuclease R-loop complex after target DNA cleavage, Nature 546(7659) (2017) 559–+. https://doi.org/10.1038/nature22398. E. Toth, E. Varga, P.I. Kulcsar, et al., Improved LbCas12a variants with altered PAM specificities further broaden the genome targeting range of Cas12a nucleases, Nucleic Acids Research 48(7) (2020) 3722–3733. https://doi.org/10.1093/nar/gkaa110. J.S. Chen, E. Ma, L.B. Harrington, et al., CRISPR-Cas12a target binding unleashes indiscriminate single-stranded DNase activity, Science 360(6387) (2018) 436–+. https://doi.org/10.1126/science.aar6245. S.-Y. Li, Q.-X. Cheng, J.-K. Liu, et al., CRISPR-Cas12a has both cis- and trans-cleavage activities on single-stranded DNA, Cell Research 28(4) (2018) 491–493. https://doi.org/10.1038/s41422-018-0022-x. J.P. Broughton, X. Deng, G. Yu, et al., CRISPR-Cas12-based detection of SARS-CoV-2, Nature Biotechnology 38(7) (2020) 870–+. https://doi.org/10.1038/s41587-020-0513-4. D.Y. Kim, J.M. Lee, S.B. Moon, et al., Efficient CRISPR editing with a hypercompact Cas12f1 and engineered guide RNAs delivered by adeno-associated virus, Nature Biotechnology 40(1) (2022) 94–+. https://doi.org/10.1038/s41587-021-01009-z. H.K. Kim, S. Min, M. Song, et al., Deep learning improves prediction of CRISPR-Cpf1 guide RNA activity, Nature Biotechnology 36(3) (2018) 239–+. https://doi.org/10.1038/nbt.4061. D.C. Swarts, J. van der Oost, M. Jinek, Structural Basis for Guide RNA Processing and Seed-Dependent DNA Targeting by CRISPR-Cas12a, Molecular Cell 66(2) (2017) 221–+. https://doi.org/10.1016/j.molcel.2017.03.016. T. Yamano, B. Zetsche, R. Ishitani, et al., Structural Basis for the Canonical and Non-canonical PAM Recognition by CRISPR-Cpf1, Molecular Cell 67(4) (2017) 633–+. https://doi.org/10.1016/j.molcel.2017.06.035. P.C. DeWeirdt, K.R. Sanson, A.K. Sangree, et al., Optimization of AsCas12a for combinatorial genetic screens in human cells, Nature Biotechnology 39(1) (2021) 94–104. https://doi.org/10.1038/s41587-020-0600-6. I. Dunham, A. Kundaje, S.F. Aldred, et al., An integrated encyclopedia of DNA elements in the human genome, Nature 489(7414) (2012) 57–74. https://doi.org/10.1038/nature11247. B. Langmead, S.L. Salzberg, Fast gapped-read alignment with Bowtie 2, Nature Methods 9(4) (2012) 357–U54. https://doi.org/10.1038/nmeth.1923. G. Chuai, H. Ma, J. Yan, et al., DeepCRISPR: optimized CRISPR guide RNA design by deep learning, Genome Biology 19 (2018), 80. https://doi.org/10.1186/s13059-018-1459-4. A. Dosovitskiy, L. Beyer, A. Kolesnikov, et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, Arxiv (2021). https://doi.org/arXiv:2010.11929. Y. Ji, Z. Zhou, H. Liu, et al., DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome, Bioinformatics 37(15) (2021) 2112–2120. https://doi.org/10.1093/bioinformatics/btab083. H. Zhu, C. Liang, CRISPR-DT: designing gRNAs for the CRISPR-Cpf1 system with improved target efficiency and specificity, bioRxiv (2018). https://doi.org/10.1101/269910. Z.S. Ameen, M. Ozsoz, A.S. Mubarak, et al., C-SVR Crispr: Prediction of CRISPR/Cas12 guideRNA activity using deep learning models, Alexandria Engineering Journal 60(4) (2021) 3501–3508. https://doi.org/10.1016/j.aej.2021.02.007. A. O'Brien, D.C. Bauer, G. Burgio, Predicting CRISPR-Cas12a guide efficiency for targeting using machine learning, PloS one 18(10) (2023) e0292924–e0292924. https://doi.org/10.1371/journal.pone.0292924. H. Kim, M. Song, J. Lee, et al., In vivo high-throughput profiling of CRISPR-Cpf1 activity, Nature Methods 14(2) (2017) 153–159. https://doi.org/10.1038/nmeth.4104. T. Yamano, H. Nishimasu, B. Zetsche, et al., Crystal Structure of Cpf1 in Complex with Guide RNA and Target DNA, Cell 165(4) (2016) 949–962. https://doi.org/10.1016/j.cell.2016.04.003. X. Liao, Q. Liu, G. Chuai, PrimeNet: rational design of Prime editing pegRNAs by deep learning, Briefings in Bioinformatics 26(3) (2025), bbaf293. https://doi.org/10.1093/bib/bbaf293. X. Kong, R. Jiao, H. Lin, et al., Peptide design through binding interface mimicry with PepMimic, Nature Biomedical Engineering (2025). https://doi.org/10.1038/s41551-025-01507-4. L. Jiao, Y. Wang, X. Liu, et al., Causal Inference Meets Deep Learning: A Comprehensive Survey, Research 7 (2024), 0467. https://doi.org/10.34133/research.0467. J. Pearl, The seven tools of causal inference, with reflections on machine learning, Communications of the ACM 62(3) (2019) 54–60. https://doi.org/10.1145/3241036. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 16 Mar, 2026 Reviews received at journal 02 Mar, 2026 Reviewers agreed at journal 01 Mar, 2026 Reviews received at journal 01 Mar, 2026 Reviewers agreed at journal 05 Feb, 2026 Reviewers agreed at journal 05 Feb, 2026 Reviewers agreed at journal 04 Feb, 2026 Reviewers invited by journal 04 Feb, 2026 Editor assigned by journal 04 Feb, 2026 Editor invited by journal 03 Feb, 2026 Submission checks completed at journal 02 Feb, 2026 First submitted to journal 02 Feb, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8736163","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":586516463,"identity":"d3dbc455-9aba-4041-a8bf-240512813183","order_by":0,"name":"Yiming Shi","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Yiming","middleName":"","lastName":"Shi","suffix":""},{"id":586516464,"identity":"4aa8cf1f-2285-407e-87e6-e1e5a1271dca","order_by":1,"name":"Junkai Yin","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Junkai","middleName":"","lastName":"Yin","suffix":""},{"id":586516465,"identity":"e42d46e1-f0b4-41b1-b580-44974b922d35","order_by":2,"name":"Shurui Ning","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Shurui","middleName":"","lastName":"Ning","suffix":""},{"id":586516466,"identity":"08f44fe7-86a5-4ee2-af8c-8738a557c0d5","order_by":3,"name":"Jinling Yuan","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Jinling","middleName":"","lastName":"Yuan","suffix":""},{"id":586516467,"identity":"1af69705-c88c-4803-8cb6-e8b3333a5acb","order_by":4,"name":"Degang Yang","email":"","orcid":"","institution":"Tongji University","correspondingAuthor":false,"prefix":"","firstName":"Degang","middleName":"","lastName":"Yang","suffix":""},{"id":586516468,"identity":"38df30ee-39a2-4ca1-a97c-3496653621e8","order_by":5,"name":"Guohui Chuai","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABK0lEQVRIie3QvWqDQADA8RPhXA7nC4K+wgVBMpQ+y0kgLkkRugQ6JFDQxXRO3sI+QU8O6nJ0tlPt0NLR4NKA0J4SmuWkGQv1P/hx+rs7BWBo6A8GjbXGuqv2SMmX3d1gchxRZCIGTqQMdRf9RmxMT0QrK91Pfp71EIh4yfbR5ZWZb95SSmDwYGz4xyQEtllQrQ5V3xKRbCum1yORewUlaJGgp9lEbswdFVS3tqpVGOFoqftpMYOS4EWC5x6RRI5QqCMFwUHFG7Ly05f3lpAAHcmqn8wJB0su54QtoVQSt5SEkj6CRJglIvd3otsYGydCePIn4/FOvN5aCuLE8X15iG78u/wRPn82zDHixK1xc+GY+TSrFUQZtDDA8qytzwQA6Pvq7HeHhoaG/kHfyFdu5gnI3rAAAAAASUVORK5CYII=","orcid":"","institution":"Tongji University","correspondingAuthor":true,"prefix":"","firstName":"Guohui","middleName":"","lastName":"Chuai","suffix":""}],"badges":[],"createdAt":"2026-01-30 02:23:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8736163/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8736163/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102297678,"identity":"6479acdb-38fa-44f8-8082-bf5612a5687d","added_by":"auto","created_at":"2026-02-10 10:28:44","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":111866,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eArchitecture of the DeepCas12a Model Integrating Convolutional and Vision Transformer Modules for Predicting CRISPR-Cas12a Cleavage Efficiency.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/daa84b6d028abd75dbf1d3fe.jpg"},{"id":102243928,"identity":"3348201b-0fb4-493d-ae13-778281f7e690","added_by":"auto","created_at":"2026-02-09 17:31:34","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":70474,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOne-hot Encoding and Epigenetic Channel Representation of the 6×34 Input Tensor Structure. \u003c/strong\u003eChannels 1–4 represent nucleotide identity (A/T/C/G), channel 5 indicates DNA methylation (RRBS), and channel 6 indicates chromatin accessibility (DNase-seq).\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/f9dc1744ff36241144c9320f.jpg"},{"id":102243933,"identity":"df4db10d-f69b-4a15-8905-ba271ef01ebf","added_by":"auto","created_at":"2026-02-09 17:31:34","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":136538,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBenchmarking DeepCas12a against Existing gRNA Efficiency Prediction Models.\u003c/strong\u003e \u003cstrong\u003e(a)\u003c/strong\u003e Spearman correlation coefficients and \u003cstrong\u003e(b) \u003c/strong\u003ePearson correlation coefficients between predicted and observed cleavage efficiency \u003cstrong\u003e(c)\u003c/strong\u003ePrecision–recall and \u003cstrong\u003e(d) \u003c/strong\u003eROC curves comparing DeepCas12a with DeepCpf1, C-SVR, CRISPR-DT, and Cas12a_predictor. DeepCas12a achieves the highest AP and AUC values.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/8e1bf4a4bff2e613fd932250.jpg"},{"id":102243930,"identity":"96d7fbcd-ea73-4a40-a5bf-26deabd71e0a","added_by":"auto","created_at":"2026-02-09 17:31:34","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":148802,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGeneralization Performance of DeepCas12a on Independent Test Datasets. (a-b)\u003c/strong\u003e Spearman and Pearson correlation coefficients on HT2 and HT3 datasets, showing that DeepCas12a achieves the highest consistency with experimental data.\u003cstrong\u003e(c-f)\u003c/strong\u003eROC and precision–recall curves on HT2 and HT3 test sets, where DeepCas12a outperforms existing models, demonstrating superior generalization ability.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/bd610248bd648d993153a820.jpg"},{"id":102297429,"identity":"15d4f7f1-a9fb-4cb3-bf5d-c537a3b3e2c0","added_by":"auto","created_at":"2026-02-10 10:27:24","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":92677,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAblation Study of DeepCas12a on the HT1-2 Test Set. (a-b)\u003c/strong\u003e Spearman and Pearson correlation coefficients of different model variants, showing that the full DeepCas12a integrating both sequence and epigenetic features achieves the best correlation with experimental data.\u003cstrong\u003e(c-d)\u003c/strong\u003e ROC and precision–recall curves of ablation variants, indicating that removing epigenetic information, shortening/lengthening the input window, or excluding the transformer encoder reduces predictive performance.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/acbabc2a0b9087b1bb50c6a0.jpg"},{"id":102243934,"identity":"9275ca9e-a138-49b1-9c37-a4ef81e32a70","added_by":"auto","created_at":"2026-02-09 17:31:34","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":56230,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInterpretable Modeling of Sequence and Epigenetic Features in Predictions.\u003c/strong\u003e (a) Multi-feature saliency heatmap for positive sample. (b) Multi-feature saliency heatmap for negative sample.\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/692f634ad3cbaa635da8e757.jpg"},{"id":102397428,"identity":"1b2d1fd1-5852-49a8-9394-f8059b5ef054","added_by":"auto","created_at":"2026-02-11 10:16:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1396929,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8736163/v1/2ae4d0cb-739f-4bef-856a-a7694a8ebfe2.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eDeepCas12a: A hybrid deep learning framework for accurate Cas12a efficiency prediction from sequence and epigenetic information\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eSince its initial application in 2012[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], CRISPR genome editing has demonstrated transformative potential across agriculture, medicine, and basic research. Among these, the type II CRISPR-Cas9 system has emerged as the dominant tool due to its high editing efficiency. However, limitations such as strict NGG protospacer adjacent motif (PAM) dependence and off-target effects restrict its utility in complex genomic contexts[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Consequently, the type V CRISPR-Cas12a system has gained traction as an alternative, offering distinct mechanistic advantages that address the core limitations of Cas9[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eUnlike Cas9, Cas12a recognizes a T-rich PAM (TTTV), significantly expanding the accessible genomic target space[\u003cspan additionalcitationids=\"CR5\" citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Furthermore, its reliance on a single crRNA simplifies construct design. Additionally, Cas12a exhibits collateral cleavage activity triggered upon target binding, enabling single-molecule nucleic acid detection[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e] \u0026mdash;a feature exploited for rapid pathogen diagnostics, including SARS-CoV-2 detection[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDespite these advantages, Cas12a orthologs display significant heterogeneity in cleavage efficiency, with variation exceeding an order of magnitude across different guide RNAs[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Current efficiency prediction tools, such as DeepCpf1, have achieved only moderate predictive performance[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. This performance gap stems from the mechanistic complexity of Cas12a, where efficiency is modulated by crRNA secondary structure[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], local epigenetic landscapes, and sequence features including PAM-proximal GC content[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. While robust models exist for Cas9, direct transfer learning to Cas12a results in a substantial performance decline (AUC decrease\u0026thinsp;\u0026gt;\u0026thinsp;10%)[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], necessitating the development of dedicated Cas12a predictors.\u003c/p\u003e \u003cp\u003eTo address these challenges, we propose DeepCas12a, a deep learning framework utilizing a dual-stream hybrid architecture (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This model integrates local feature extraction via Convolutional Neural Networks (CNNs) and global sequence modeling via a Transformer-based attention mechanism. Specifically, parallel convolutional layers capture motif preferences in PAM-proximal regions, while a Transformer encoder models long-range interactions\u0026mdash;such as mismatch tolerance\u0026mdash;within the crRNA\u0026ndash;target DNA duplex. Benchmarking results demonstrate that DeepCas12a significantly outperforms existing tools in prediction accuracy on independent test sets. Furthermore, we provide interpretable saliency maps to elucidate the mechanistic basis of Cas12a activity. This framework offers practical guidance for therapeutic gene editing design\u0026mdash;exemplified by targets for sickle cell disease\u0026mdash;thereby reducing experimental screening costs.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"2. Material and Methods","content":"\u003ch2\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003cstrong\u003e.1 Data Preparation\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eCRISPR-Cas12a cleavage efficiency data were retrieved from Kim et al. (2018)[11]. This dataset (HT1) quantifies Cas12a editing efficiency across target sites in HEK293T cells via high-throughput sequencing. The complete dataset (N=16,292) was randomly partitioned into a training set (n=15,000) and an independent test set (n=1,292).\u003c/p\u003e\n\u003cp\u003eFor binary classification, labels were assigned based on background-corrected indel frequencies: samples exhibiting a \u0026gt;2-fold change over controls and an absolute indel frequency \u0026gt;50% were labeled positive (1); all others were labeled negative (0). This threshold aligns with high-efficacy classification standards used in prior CRISPR studies.\u003c/p\u003e\n\u003cp\u003eTo model epigenetic effects, DNA methylation (RRBS) and chromatin accessibility (DNase-seq) data for HEK293T cells were obtained from the ENCODE project[15]. Target sequences (34 bp) were aligned to the human reference genome (hg38) using Bowtie 2[16]. We identified target sites overlapping with RRBS-covered regions or DNase narrow peaks to generate binary epigenetic feature maps (presence=1, absence=0)[17]. Input sequences were encoded as 6-channel tensors. Channels 1\u0026ndash;4 represented nucleotide identity via one-hot encoding (A, C, G, T). Channel 5 encoded DNA methylation status, and Channel 6 encoded chromatin accessibility, as shown in Figure 2.\u003c/p\u003e\n\u003cp\u003eThis representation captures both local sequence context and the epigenetic landscape, reflecting the multifactorial regulation of Cas12a activity.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003e2.2 Model Architecture\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eDeepCas12a utilizes a hybrid neural network architecture to predict CRISPR-Cas12a cleavage efficiency. The framework comprises three primary modules: a Convolutional Neural Network (CNN) for local feature extraction, a Vision Transformer encoder (ViT) for global context modeling, and a Multi-Layer Perceptron (MLP) classifier.\u003c/p\u003e\n\u003cp\u003eInput tensors (6\u0026times;34) are processed by the CNN module to extract local motif definitions and epigenetic patterns.\u003c/p\u003e\n\u003cp\u003eConvolutional feature maps are tokenized into embeddings, augmented with positional encodings, and input into the vision transformer module. A multi-head self-attention mechanism models long-range dependencies, capturing global regulatory interactions within the sequence. This design adapts the vision transformer strategy [18] to genomic data, leveraging the efficacy of attention mechanisms in sequence modeling demonstrated by architectures such as DNABERT [19].\u003c/p\u003e\n\u003cp\u003eFinally, the global class token ([CLS]) is fed into the MLP to output a probability score representing binary cleavage efficiency.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003e2.3 Experimental Setup and Training Parameters\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eDeepCas12a was implemented in Python using the PyTorch framework. Training was accelerated via CUDA 12.4 on a workstation equipped with dual NVIDIA GeForce RTX 3090 GPUs (24 GB VRAM).\u003c/p\u003e\n\u003cp\u003eHyperparameters were systematically optimized to maximize validation performance. We employed Bayesian optimization to tune the learning rate, attention/global dropout rates, network depth, embedding dimension, MLP expansion ratio, and the number of attention heads.\u003c/p\u003e\n\u003cp\u003eThe final hyperparameter configuration is detailed in Table 1. To mitigate overfitting, we applied early stopping and assessed model robustness via k-fold cross-validation on the training data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eSelected model hyperparameters:\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eHyperparameter\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eValue\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eattention_drop\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e0.271\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003edropout\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e0.307\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003edepth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eembed_dim\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e256\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003elearning rate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e0.00027\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003emlp_ratio\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp;2.45\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003enum_heads\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003e\u0026nbsp;12\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eDeepCas12a was benchmarked against four representative Cas12a efficiency predictors, as shown in Table 2:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eDeepCpf1[11]: A CNN-based end-to-end predictor utilizing sgRNA sequences;\u003c/li\u003e\n \u003cli\u003eCRISPR-DT[20]: An SVM classifier integrating sequence features with manually engineered biological characteristics;\u003c/li\u003e\n \u003cli\u003eC-SVR[21]: A hybrid architecture combining CNN-based feature extraction with Support Vector Regression (SVR);\u003c/li\u003e\n \u003cli\u003eCas12a_predictor[22]: A Random Forest model incorporating sgRNA sequences and biological features.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u0026nbsp;\u003c/strong\u003eSummary of mainstream Cas12a efficiency prediction models, their references, and implementation sources:\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"580\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 159px;\"\u003e\n \u003cp\u003eModel Name\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 107px;\"\u003e\n \u003cp\u003eReference\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 166px;\"\u003e\n \u003cp\u003eMethod Type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 149px;\"\u003e\n \u003cp\u003eCode Availability\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eDeepCpf1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003csup\u003e1\u003c/sup\u003eKim et al., 2018\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003eCNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 149px;\"\u003e\n \u003cp\u003eGitHub\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eCRISPR-DT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003csup\u003e2\u003c/sup\u003eZhu et al., 2019\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003eFeature Engineering + SVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 149px;\"\u003e\n \u003cp\u003ePaper Reproduction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eC-SVR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003csup\u003e3\u003c/sup\u003eAmeen et al., 2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003eCNN + SVR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 149px;\"\u003e\n \u003cp\u003ePaper Reproduction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 159px;\"\u003e\n \u003cp\u003eCas12a_predictor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 107px;\"\u003e\n \u003cp\u003e\u003csup\u003e4\u003c/sup\u003eO\u0026rsquo;Brien et al., 2023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003eFeature Engineering + RandomForest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 149px;\"\u003e\n \u003cp\u003eGitHub\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eSource: Model results are obtained from either the original publications or official code repositories using the same dataset and evaluation pipeline for consistency.\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e1\u003c/sup\u003eKim et al. (2018), Nature Biotechnology. Deep learning improves prediction of CRISPR\u0026ndash;Cpf1 guide RNA activity.[11]\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e2\u003c/sup\u003eZhu et al. (2019), Bioinformatics. CRISPR-DT: designing gRNAs for the CRISPR-Cpf1 system with improved target efficiency and specificity.[20]\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e3\u003c/sup\u003eAmeen et al. (2021), Alexandria Engineering Journal. C-SVR Crispr: Prediction of CRISPR/Cas12 guideRNA activity using deep learning models.[21]\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003csup\u003e4\u003c/sup\u003eO\u0026rsquo;Brien et al. (2023), PLOS ONE. Predicting CRISPR-Cas12a guide efficiency for targeting using machine learning.[22]\u0026nbsp;\u003c/p\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003e3.1 Model performance on holdout test set\u003c/h2\u003e\n \u003cp\u003eWe evaluated DeepCas12a on the independent test set described in Section 2.1. The model achieved an Average Precision (AP) of 0.783 and an Area Under the ROC Curve (AUC) of 0.868.\u003c/p\u003e\n \u003cp\u003eTo assess quantitative predictive accuracy, we computed Spearman (ρ) and Pearson (r) correlation coefficients between predicted probabilities and experimental cleavage scores, yielding 0.630 and 0.631, respectively. These correlations indicate that DeepCas12a captures both the rank order and magnitude of cleavage efficiency, supporting its utility for quantitative prediction.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\"\u003e\n \u003ch2\u003e3.2 Comparison with current sgRNA efficiency prediction models\u003c/h2\u003e\n \u003cp\u003eTo evaluate DeepCas12a, we benchmarked it against four state-of-the-art Cas12a efficiency predictors: DeepCpf1, CRISPR-DT, C-SVR, and Cas12a_predictor. DeepCpf1 and C-SVR rely on sequence-derived deep learning features, whereas CRISPR-DT and Cas12a_predictor utilize handcrafted biological features within traditional machine learning frameworks. Baseline models were evaluated on the independent test set using the authors' public implementations and default hyperparameters to ensure fair comparison.\u003c/p\u003e\n \u003cp\u003eIn binary classification tasks, DeepCas12a achieved the highest performance metrics (AP = 0.783, AUC = 0.868; Fig.\u0026nbsp;3c–d). It outperformed the second-best model, DeepCpf1 (AP = 0.755, AUC = 0.858), while other baselines showed significantly lower discrimination.\u003c/p\u003e\n \u003cp\u003eTo assess quantitative ranking accuracy, we computed Spearman rank (ρ) and Pearson linear (r) correlations between predicted and experimental efficiencies. DeepCas12a yielded the highest correlation coefficients (ρ = 0.630, r = 0.631). These values exceed those of DeepCpf1 (ρ = 0.612, r = 0.594) and other baselines (Fig.\u0026nbsp;3a–b), indicating a more accurate global ordering of target efficacy.\u003c/p\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cp\u003eIn summary, DeepCas12a demonstrates consistent improvements over existing state-of-the-art methods in both classification and regression tasks.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\"\u003e\n \u003ch2\u003e3.3 Generalization evaluation on independent test sets\u003c/h2\u003e\n \u003cp\u003eTo assess generalization beyond the primary test set, we evaluated DeepCas12a on extra two independent datasets, HT2 (n = 2,963) and HT3 (n = 1,251), which were excluded from model training. These datasets were also derived from the Kim et al. (2018) study[11]\u003c/p\u003e\n \u003cp\u003eDeepCas12a demonstrated superior correlation and classification performance across both datasets compared to baseline models (Fig. 4). Specifically, it achieved the highest rank correlations with experimental cleavage efficiencies: Spearman coefficients (ρ) were 0.538 (HT2) and 0.313 (HT3), while Pearson coefficients (r) were 0.540 (HT2) and 0.294 (HT3) (Fig. 4a–b). These results suggest that the model generalizes effectively to unseen data.\u003c/p\u003e\n \u003cp\u003eIn classification tasks, DeepCas12a consistently showed the highest discrimination (Fig. 4c–f). On HT2, it achieved an AUC of 0.823 and Average Precision (AP) of 0.697, marginally surpassing DeepCpf1 (AUC = 0.819) and significantly outperforming C-SVR (AUC = 0.751). A similar trend was observed on HT3, where DeepCas12a maintained leading performance (AUC = 0.726, AP = 0.360). These results demonstrate the model's robustness to dataset shifts.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\"\u003e\n \u003ch2\u003e3.4 Ablation study\u003c/h2\u003e\n \u003cp\u003eWe performed ablation studies on the independent test set to quantify the contributions of specific feature types and architectural components. Five variants were evaluated: (1) DeepCas12a (Full); (2) Seq-Only (excluding epigenetic features); (3) CNN-Only (excluding the Transformer); (4) 50 bp context; and (5) 20 bp guide-only.\u003c/p\u003e\n \u003cp\u003eThe full model achieved the highest correlation with experimental cleavage efficiencies (ρ = 0.630, r = 0.631; Fig.\u0026nbsp;5a–b). Removing epigenetic features resulted in a minor performance decline (ρ = 0.621, r = 0.621), suggesting that chromatin accessibility and methylation data provide limited but complementary information.\u003c/p\u003e\n \u003cp\u003eRegarding input length, the 34 bp window (4 bp upstream, PAM, 23 bp protospacer, 3 bp downstream) proved optimal, outperforming 50 bp and 20 bp configurations. Extending the window to 50 bp likely introduced noise, while restricting it to the 20 bp guide sequence failed to capture essential flanking signals.\u003c/p\u003e\n \u003cp\u003eDeepCas12a also led in classification metrics (AUC = 0.868, AP = 0.783; Fig. 5c–d). Removal of the Transformer encoder caused a distinct decrease in performance, supporting the utility of modeling long-range dependencies. These results indicate that sequence context, epigenetic features, and the hybrid CNN-Transformer architecture collectively contribute to predictive accuracy.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\"\u003e\n \u003ch2\u003e3.5 Interpretability analysis via saliency maps\u003c/h2\u003e\n \u003cp\u003eWe visualized multi-feature saliency heatmaps for representative high-efficiency (positive) and low-efficiency (negative) targets to elucidate feature contributions (Fig. 6a–b). Red regions indicate positive attribution (promoting cleavage), while blue regions indicate negative attribution (inhibiting cleavage).\u003c/p\u003e\n \u003cp\u003e(1) PAM Recognition: In high-efficiency targets (Fig. 6a), the PAM core (positions 5–7) exhibits contiguous positive attribution for Thymine, whereas low-efficiency targets (Fig. 6b) display inhibitory signals at these sites. This confirms the model successfully learned the preference for a T-rich PAM. At position 8, the model assigns positive weights to non-T bases (A/C/G) and negative weights to T, reflecting the biological requirement for a V (A/C/G) nucleotide at the terminal PAM position[23].\u003c/p\u003e\n \u003cp\u003e(2) Seed Region Sensitivity: High-efficiency targets display concentrated positive saliency within the protospacer \"seed\" region (positions 9–14). This aligns with experimentally validated seed regions where mismatches drastically reduce Cas12a activity[24, 25]. Conversely, low-efficiency targets show dispersed, weak saliency, indicative of ineffective guide–target pairing. Attribution magnitudes diminish beyond position 25, consistent with the minimal impact of distal nucleotides on cleavage.\u003c/p\u003e\n \u003cp\u003e(3) Epigenetic Modulation: The DNase accessibility channel predominantly shows positive attribution (red) in high-efficiency targets—particularly PAM-proximal regions—suggesting open chromatin enhances accessibility. In contrast, the DNA methylation (RRBS) channel appears predominantly negative (blue), capturing the repressive effect of local methylation.\u003c/p\u003e\n \u003cp\u003eThese patterns indicate that DeepCas12a has learned biologically relevant rules, including PAM specificity, seed region sensitivity, distal tolerance, and the inhibitory influence of DNA methylation.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eDeepCas12a demonstrates improved accuracy in predicting CRISPR-Cas12a cleavage efficiency. It outperforms existing baselines (DeepCpf1, CRISPR-DT, C-SVR, Cas12a_predictor) in both classification (AP, AUC) and ranking (ρ, r) metrics. The integration of local convolutional features with global attention mechanisms enables robust generalization, effectively capturing multi-level regulatory signals. By integrating epigenetic features (DNA methylation, chromatin accessibility), the architecture maintains stability where sequence-only models may falter, highlighting the value of multimodal integration in functional genomics.\u003c/p\u003e \u003cp\u003eFurthermore, saliency analysis confirms that DeepCas12a independently identifies critical functional regions, including the PAM (positions 5\u0026ndash;8) and the seed region (positions 9\u0026ndash;14). Gradient distributions align with established biological constraints, such as the requirement for a terminal \"V\" in the TTTV PAM and mismatch tolerance in distal regions. Practically, researchers can prioritize targets with high predicted efficacy and favorable saliency profiles to improve experimental validation rates. This approach is particularly valuable for therapeutic design, where identifying high-activity sites minimizes resource expenditure. Additionally, this \"interpretation\u0026ndash;validation\" loop facilitates hypothesis generation\u0026mdash;such as verifying causal links between PAM-proximal features and activity\u0026mdash;accelerating precision functional genomics.\u003c/p\u003e \u003cp\u003eDespite these advances, limitations persist. First, complex cellular contexts\u0026mdash;including cell-type specific chromatin states and transcription factor binding\u0026mdash;may limit generalization to unseen tissues. Second, while the hybrid architecture improves representation, reliance on static high-throughput datasets risks overfitting to library-specific biases. Future work should incorporate diverse datasets from primary cells and clinical samples to evaluate cross-condition adaptability. Moreover, current reliance on bulk ENCODE data lacks temporal resolution; integrating single-cell multi-omics (e.g., scATAC-seq) could refine epigenetic modeling.\u003c/p\u003e \u003cp\u003eThe Vision transformer architecture also positions the framework for future generative design and causal inference tasks. Generative extensions (e.g., diffusion models) could enable the de novo design of optimal guide RNAs under functional constraints[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Furthermore, attention mechanisms offer pathways to uncover regulatory factors, facilitating a shift from correlation discovery to causal mechanism elucidation[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. In addition, the combination of pre-training and self-supervised learning strategies is expected to enhance the model\u0026rsquo;s generalization ability on rare sequences or novel enzyme variants, promoting a positive feedback loop between high-throughput virtual screening and experimental validation and accelerating innovation in CRISPR applications.\u003c/p\u003e \u003cp\u003eIn summary, DeepCas12a provides an accurate, interpretable solution for Cas12a efficiency prediction, establishing a foundation for intelligent gene editing design and mechanistic research.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this study, we present DeepCas12a, a multimodal deep learning framework integrating sequence and epigenetic contexts to predict CRISPR-Cas12a cleavage efficiency. By synergizing local convolutional feature extraction with global attention mechanisms, the model achieves state-of-the-art performance, robustness, and biological interpretability.\u003c/p\u003e\n\u003cp\u003eDeepCas12a offers an effective tool for high-throughput target screening and rational guide RNA design, facilitating applications in precision gene editing and therapeutic development. Future integration of diverse multi-omics datasets and generative modeling strategies will further expand the utility of deep learning in CRISPR functional genomics.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics, Consent to Participate, and Consent to Publish declarations: not applicable.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eFunding Declaration\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThis study was supported by National Natural Science Foundation of China [62002265]; Tongji University \u0026quot;Medicine + X\u0026quot; Cross Research Program[2025080107].\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe datasets analyzed during the current study are available on the Github at https://github.com/bm2-lab-submission/DeepCas12a.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe model and model usage are available on the Github at https://github.com/bm2-lab-submission/DeepCas12a.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eY.S., J.K.Y., S.N., J.L.Y., D.Y. and G.C. conceived and designed the study. Y.S. designed and trained the model. S.N. implemented the benchmarking. Y.S. and J.L.Y. analyzed the results and generated the figures. Y.S., J.K.Y., D.Y. and G.C. wrote the manuscript. G.C. and D.Y. supervised the entire project. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eM. Jinek, K. Chylinski, I. Fonfara, et al., A Programmable Dual-RNA-Guided DNA Endonuclease in Adaptive Bacterial Immunity, Science 337(6096) (2012) 816\u0026ndash;821. https://doi.org/10.1126/science.1225829.\u003c/li\u003e\n\u003cli\u003eD. Kim, S. Kim, S. Kim, et al., Genome-wide target specificities of CRISPR-Cas9 nucleases revealed by multiplex Digenome-seq, Genome Research 26(3) (2016) 406\u0026ndash;415. https://doi.org/10.1101/gr.199588.115.\u003c/li\u003e\n\u003cli\u003eD.C. Swarts, M. Jinek, Mechanistic Insights into the cis- and trans-Acting DNase Activities of Cas12a, Molecular Cell 73(3) (2019) 589\u0026ndash;+. https://doi.org/10.1016/j.molcel.2018.11.021.\u003c/li\u003e\n\u003cli\u003eB. Zetsche, J.S. Gootenberg, O.O. Abudayyeh, et al., Cpf1 Is a Single RNA-Guided Endonuclease of a Class 2 CRISPR-Cas System, Cell 163(3) (2015) 759\u0026ndash;771. https://doi.org/10.1016/j.cell.2015.09.038.\u003c/li\u003e\n\u003cli\u003eS. Stella, P. Alcon, G. Montoya, Structure of the Cpf1 endonuclease R-loop complex after target DNA cleavage, Nature 546(7659) (2017) 559\u0026ndash;+. https://doi.org/10.1038/nature22398.\u003c/li\u003e\n\u003cli\u003eE. Toth, E. Varga, P.I. Kulcsar, et al., Improved LbCas12a variants with altered PAM specificities further broaden the genome targeting range of Cas12a nucleases, Nucleic Acids Research 48(7) (2020) 3722\u0026ndash;3733. https://doi.org/10.1093/nar/gkaa110.\u003c/li\u003e\n\u003cli\u003eJ.S. Chen, E. Ma, L.B. Harrington, et al., CRISPR-Cas12a target binding unleashes indiscriminate single-stranded DNase activity, Science 360(6387) (2018) 436\u0026ndash;+. https://doi.org/10.1126/science.aar6245.\u003c/li\u003e\n\u003cli\u003eS.-Y. Li, Q.-X. Cheng, J.-K. Liu, et al., CRISPR-Cas12a has both cis- and trans-cleavage activities on single-stranded DNA, Cell Research 28(4) (2018) 491\u0026ndash;493. https://doi.org/10.1038/s41422-018-0022-x.\u003c/li\u003e\n\u003cli\u003eJ.P. Broughton, X. Deng, G. Yu, et al., CRISPR-Cas12-based detection of SARS-CoV-2, Nature Biotechnology 38(7) (2020) 870\u0026ndash;+. https://doi.org/10.1038/s41587-020-0513-4.\u003c/li\u003e\n\u003cli\u003eD.Y. Kim, J.M. Lee, S.B. Moon, et al., Efficient CRISPR editing with a hypercompact Cas12f1 and engineered guide RNAs delivered by adeno-associated virus, Nature Biotechnology 40(1) (2022) 94\u0026ndash;+. https://doi.org/10.1038/s41587-021-01009-z.\u003c/li\u003e\n\u003cli\u003eH.K. Kim, S. Min, M. Song, et al., Deep learning improves prediction of CRISPR-Cpf1 guide RNA activity, Nature Biotechnology 36(3) (2018) 239\u0026ndash;+. https://doi.org/10.1038/nbt.4061.\u003c/li\u003e\n\u003cli\u003eD.C. Swarts, J. van der Oost, M. Jinek, Structural Basis for Guide RNA Processing and Seed-Dependent DNA Targeting by CRISPR-Cas12a, Molecular Cell 66(2) (2017) 221\u0026ndash;+. https://doi.org/10.1016/j.molcel.2017.03.016.\u003c/li\u003e\n\u003cli\u003eT. Yamano, B. Zetsche, R. Ishitani, et al., Structural Basis for the Canonical and Non-canonical PAM Recognition by CRISPR-Cpf1, Molecular Cell 67(4) (2017) 633\u0026ndash;+. https://doi.org/10.1016/j.molcel.2017.06.035.\u003c/li\u003e\n\u003cli\u003eP.C. DeWeirdt, K.R. Sanson, A.K. Sangree, et al., Optimization of AsCas12a for combinatorial genetic screens in human cells, Nature Biotechnology 39(1) (2021) 94\u0026ndash;104. https://doi.org/10.1038/s41587-020-0600-6.\u003c/li\u003e\n\u003cli\u003eI. Dunham, A. Kundaje, S.F. Aldred, et al., An integrated encyclopedia of DNA elements in the human genome, Nature 489(7414) (2012) 57\u0026ndash;74. https://doi.org/10.1038/nature11247.\u003c/li\u003e\n\u003cli\u003eB. Langmead, S.L. Salzberg, Fast gapped-read alignment with Bowtie 2, Nature Methods 9(4) (2012) 357\u0026ndash;U54. https://doi.org/10.1038/nmeth.1923.\u003c/li\u003e\n\u003cli\u003eG. Chuai, H. Ma, J. Yan, et al., DeepCRISPR: optimized CRISPR guide RNA design by deep learning, Genome Biology 19 (2018), 80. https://doi.org/10.1186/s13059-018-1459-4.\u003c/li\u003e\n\u003cli\u003eA. Dosovitskiy, L. Beyer, A. Kolesnikov, et al., An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, Arxiv (2021). https://doi.org/arXiv:2010.11929.\u003c/li\u003e\n\u003cli\u003eY. Ji, Z. Zhou, H. Liu, et al., DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome, Bioinformatics 37(15) (2021) 2112\u0026ndash;2120. https://doi.org/10.1093/bioinformatics/btab083.\u003c/li\u003e\n\u003cli\u003eH. Zhu, C. Liang, CRISPR-DT: designing gRNAs for the CRISPR-Cpf1 system with improved target efficiency and specificity, bioRxiv (2018). https://doi.org/10.1101/269910.\u003c/li\u003e\n\u003cli\u003eZ.S. Ameen, M. Ozsoz, A.S. Mubarak, et al., C-SVR Crispr: Prediction of CRISPR/Cas12 guideRNA activity using deep learning models, Alexandria Engineering Journal 60(4) (2021) 3501\u0026ndash;3508. https://doi.org/10.1016/j.aej.2021.02.007.\u003c/li\u003e\n\u003cli\u003eA. O\u0026apos;Brien, D.C. Bauer, G. Burgio, Predicting CRISPR-Cas12a guide efficiency for targeting using machine learning, PloS one 18(10) (2023) e0292924\u0026ndash;e0292924. https://doi.org/10.1371/journal.pone.0292924.\u003c/li\u003e\n\u003cli\u003eH. Kim, M. Song, J. Lee, et al., In vivo high-throughput profiling of CRISPR-Cpf1 activity, Nature Methods 14(2) (2017) 153\u0026ndash;159. https://doi.org/10.1038/nmeth.4104.\u003c/li\u003e\n\u003cli\u003eT. Yamano, H. Nishimasu, B. Zetsche, et al., Crystal Structure of Cpf1 in Complex with Guide RNA and Target DNA, Cell 165(4) (2016) 949\u0026ndash;962. https://doi.org/10.1016/j.cell.2016.04.003.\u003c/li\u003e\n\u003cli\u003eX. Liao, Q. Liu, G. Chuai, PrimeNet: rational design of Prime editing pegRNAs by deep learning, Briefings in Bioinformatics 26(3) (2025), bbaf293. https://doi.org/10.1093/bib/bbaf293.\u003c/li\u003e\n\u003cli\u003eX. Kong, R. Jiao, H. Lin, et al., Peptide design through binding interface mimicry with PepMimic, Nature Biomedical Engineering (2025). https://doi.org/10.1038/s41551-025-01507-4.\u003c/li\u003e\n\u003cli\u003eL. Jiao, Y. Wang, X. Liu, et al., Causal Inference Meets Deep Learning: A Comprehensive Survey, Research 7 (2024), 0467. https://doi.org/10.34133/research.0467.\u003c/li\u003e\n\u003cli\u003eJ. Pearl, The seven tools of causal inference, with reflections on machine learning, Communications of the ACM 62(3) (2019) 54\u0026ndash;60. https://doi.org/10.1145/3241036.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-genomics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gics","sideBox":"Learn more about [BMC Genomics](http://bmcgenomics.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/gics","title":"BMC Genomics","twitterHandle":"#BMCGenomics","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8736163/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8736163/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCRISPR-Cas12a (Cpf1) offers distinct advantages for genome editing due to its flexible, T-rich PAM recognition. However, variable cleavage efficiency\u0026mdash;modulated by sequence context and epigenetic features\u0026mdash;remains a challenge, with existing predictors limited in accuracy and interpretability. Here, we present DeepCas12a, a hybrid deep learning framework integrating Convolutional Neural Networks (CNNs) and a Vision Transformer (ViT) encoder to capture both local sequence motifs and long-range dependencies. The model fuses DNA sequence data with epigenetic profiles (DNA methylation and chromatin accessibility) in an end-to-end architecture. Benchmarked on an independent test set, DeepCas12a outperformed state-of-the-art predictors, achieving an Average Precision of 0.783, an AUC of 0.868, and a Spearman correlation of 0.630. Furthermore, interpretability analysis via saliency maps confirms the model captures biologically relevant features, including PAM specificity and seed region sensitivity, facilitating rational guide RNA design.\u003c/p\u003e","manuscriptTitle":"DeepCas12a: A hybrid deep learning framework for accurate Cas12a efficiency prediction from sequence and epigenetic information","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-09 17:31:29","doi":"10.21203/rs.3.rs-8736163/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-16T20:47:40+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-02T14:31:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"42209076413304365658670656983971161065","date":"2026-03-02T04:14:45+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-01T17:29:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"245270172551634932150032089274174898546","date":"2026-02-05T09:05:39+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"50854727303697283662433376239296550246","date":"2026-02-05T06:10:27+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"56039077543229543725706260854394676660","date":"2026-02-05T02:22:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-05T01:41:41+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-04T16:21:59+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-03T14:32:48+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-03T02:43:08+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Genomics","date":"2026-02-03T02:38:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-genomics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gics","sideBox":"Learn more about [BMC Genomics](http://bmcgenomics.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/gics","title":"BMC Genomics","twitterHandle":"#BMCGenomics","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b78c5b21-ceb5-4c49-b5ef-d0932c7adb9f","owner":[],"postedDate":"February 9th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-30T01:38:31+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-09 17:31:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8736163","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8736163","identity":"rs-8736163","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}