Hybrid Neural Networks for Off-Target Prediction in Gene Editing: Integrating Sequence and Biophysical Features

Probably not a substantial claim: may not pass S3 (load-bearing).

DeepCRISPR uses a deep convolutional denoising neural network (DCD-NN) and CNN to predict both on- and off-target activities of sgRNAs with improved accuracy.

“DeepCRISPR uses a deep convolutional denoising neural network (DCD-NN) and CNN to predict both on-and off-target activities of sgRNAs with improved accuracy.”

• cites [7] 7. Chuai, G.; Ma, H.; Yan, J.; Chen, M.; Hong, N.; Xue, D.; Zhou, C.; Zhu, C.; Chen, K.; Duan, B.; Gu, F.; Qu, S.; Huang, D.; Wei, J.; Liu, Q. DeepCRISPR: Optimized CRISPR Guide RNA De- sign by Deep Learning. Genome Biology 2018, 19 (1) https://doi. org/10.1186/s13059-018-1459-4. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Higher GC content in the gRNA can stabilize the DNA-RNA duplex, potentially increasing on-target activity.

“Some studies suggest that higher GC content in the gRNA can stabilize the DNA-RNA duplex, potentially increasing on-target activity”

• cites [1] 1. Modrzejewski, D., Hartung, F., Lehnert, H., Sprink, T., Kohl, C., Keilwagen, J., & Wilhelm, R. (2020). Which Factors Affect the Occurrence of Off-Target Effects Caused by the Use of CRISPR/ Cas: A Systematic Review in Plants. Frontiers in plant science, 11, 574959. https://doi.org/10.3389/fpls.2020.574959 sourcePartly supported

  The Background and Results section ‘Hypothesis 3: GC-Content’ say that higher GC content may stabilize guide-RNA/genomic-DNA hybridization. They discuss this in relation to off-target effects, however, and do not say that higher GC content increases on-target activity. Page numbers are unavailable in our copy. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

Target sequences with a GC content higher than 70% may increase the likelihood of off-target effects.

“while others indicate that target sequences with a GC content higher than 70% may increase the likelihood of off-target effects.”

• cites [1] 1. Modrzejewski, D., Hartung, F., Lehnert, H., Sprink, T., Kohl, C., Keilwagen, J., & Wilhelm, R. (2020). Which Factors Affect the Occurrence of Off-Target Effects Caused by the Use of CRISPR/ Cas: A Systematic Review in Plants. Frontiers in plant science, 11, 574959. https://doi.org/10.3389/fpls.2020.574959 sourceSupported

May combine two claims.

Cas9 tolerates up to three mismatches between the gRNA and genomic DNA, which increases the chances of off-target effects in regions with higher GC content.

“For instance, recent studies have shown that Cas9 tolerates up to three mismatches between the gRNA and genomic DNA, which increases the chances of off-target effects in regions with higher GC content.”

• cites [2] 2. Guo, C.; Ma, X.; Gao, F.; Guo, Y. Off-Target Effects in CRISPR/ Cas9 Gene Editing. Frontiers in Bioengineering and Biotechnology 2023, 11 (1143157). https://doi.org/10.3389/fbioe.2023.1143157 sourcePartly supported

  The Introduction states that Cas9 can tolerate up to three mismatches between sgRNA and genomic DNA, allowing off-target cleavage. Section 2 says a prediction tool reports GC content and on/off-target scores, but it does not say that higher-GC regions have a greater chance of off-target effects. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

DeepCRISTL applies deep transfer learning to predict CRISPR/Cas9 editing efficiency in specific cellular contexts, fine-tuning predictions with smaller, biologically relevant datasets for superior context-specific accuracy.

“Similarly, DeepCRISTL, developed in 2024, applies deep transfer learning to predict CRISPR/Cas9 editing efficiency in specific cellular contexts, fine-tuning predictions with smaller, biologically relevant datasets for superior context-specific accuracy.”

• cites [8] 8. Elkayam, S.; Ido Tziony; Orenstein, Y. DeepCRISTL: Deep Transfer Learning to Predict CRISPR/Cas9 On-Target Edit- ing Efficiency in Specific Cellular Contexts. Bioinformatics 2024. https://doi.org/10.1093/bioinformatics/btae481. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Manipulating Cas9's biophysical attributes, such as minimizing nonspecific RNA-DNA interactions, significantly reduces off-target mutations while maintaining high on-target efficiency.

“For example, a study demonstrated that manipulating Cas9's biophysical attributes, such as minimizing nonspecific RNA-DNA interactions, significantly reduces off-target mutations while maintaining high on-target efficiency.”

• cites [5] 5. Fu, Y.; Foden, J. A.; Khayter, C.; Maeder, M. L.; Reyon, D.; Joung, J. K.; Sander, J. D. High-Frequency Off-Target Mutagenesis Induced by CRISPR-Cas Nucleases in Human Cells. Nature Biotechnology 2013, 31 (9), 822-826. https://doi.org/10.1038/nbt.2623. sourceCould not verify

  The discussion of RNA:DNA hybrid stability (section labeled Abstract; page unavailable) says its effect on off-target specificity needs further experiments; it does not report a successful intervention. The paper also reports that lowering guide-RNA and Cas9 expression reduced on-target mutation rates without appreciably changing relative off-target rates, but that is a different intervention from altering Cas9’s biophysical attributes. No passage establishes the claimed reduction in off-target mutations while maintaining high on-target efficiency. (LLM, conf=0.77)

  Experimental — accuracy not yet measured

May combine two claims.

The BoostMEC model improves off-target prediction by considering binding energy, cleavage efficiency and other biophysical factors, reducing false positives and enhancing off-target site identification.

“The BoostMEC model further improves prediction by considering binding energy, cleavage efficiency, and other biophysical factors, reducing false positives and enhancing off-target site identification.”

• cites [6] 6. Zarate, O. A.; Yang, Y.; Wang, X.; Wang, J.-P. BoostMEC: Predict- ing CRISPR-Cas9 Cleavage Efficiency through Boosting Models. BMC Bioinformatics 2022, 23 (1). https://doi.org/10.1186/s12859- 022-04998-z. sourcePartly supported

  The Methods (Feature construction) describe thermodynamic features used to predict cleavage efficiency, so that part of the claim is supported. The Discussion explicitly says that gRNA–DNA binding energy was not used in BoostMEC. The paper describes cleavage-efficiency prediction, not reduced false positives or improved off-target site identification. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

In the HEK293 cell line, model M2 achieved a recall of 68.54%, an F1 score of 71.37% and an AUC of 80.82%.

“The recall was 68.54%, and the F1 score was 71.37%, indicating a strong balance between true positive and false positive rates.”

• Also stated in: § Cell Line HEK293: · boosted

  “The model also showed a robust AUC of 80.82%, confirming its capability to distinguish between on-target and off-target effects.”

May combine two claims. The quote only partly states this claim.

In the U2OS cell line, model M10 achieved a recall of 74.87%, an F1 score of 75.87% and a precision of 77.39%.

“This model also achieved a recall of 74.87% and an F1 score of 75.87%, making it particularly effective in identifying true off-target effects while maintaining a balance between precision and recall.”

• Also stated in: § Cell Line U2OS: · boosted

  “Precision for this model was 77.39%, further demonstrating its reliability in minimizing false positives in this context.”

May combine two claims.

In the K562 cell line, the CNN variant (M2) of Architecture 1 delivered the best performance, with accuracy 83.78%, precision 89.74%, F1 score 82.39%, AUC 89.90%, recall 76.24% and balanced accuracy 83.78%.

“In the K562 cell line, the CNN variant (M2) in Architecture 1 consistently delivered the best performance, with the highest accuracy at 83.78%, precision at 89.74%, F1 score at 82.39%, and AUC at 89.90%.”

• Also stated in: § Cell Line K562: · hedged

  “The recall was also strong at 76.24%, indicating its effectiveness in identifying true off-targets.”

• Also stated in: § Cell Line K562: · plain

  “The balanced accuracy of 83.78% highlights the model's consistent performance across different datasets.”

May combine two claims. The quote only partly states this claim.

Adding attention mechanisms improved performance only slightly and was not very impactful when the models were evaluated across all cell lines together.

“Although the addition of attention mechanisms improved performance, it was not very impactful when experimented against all the cell lines together.”

May combine two claims. The quote only partly states this claim.

In Architecture 2, the CNN+GRU model achieved the highest recall (76.77%) and the highest F1 score (75.12%).

“while the CNN+GRU model led in recall at 76.77% and F1 score at 75.12%”

May combine two claims.

In Architecture 3 (Attention-Augmented Sequential Hybrid), the CNN+BiDirectional LSTM+Attention model achieved the highest accuracy (73.31%) and the highest balanced accuracy (73.27%).

“The CNN+BiDirectional LSTM+Attention model achieved the highest accuracy (73.31%) and balanced accuracy (73.27%)”

The quote only partly states this claim.

In Architecture 2, the CNN+LSTM variant achieved the highest precision (77.31%).

“The CNN+LSTM variant excelled in precision at 77.31%, crucial for accurate predictions”

May combine two claims.

In Architecture 3, the CNN+LSTM+Attention model achieved the highest recall (78.97%) and the highest F1 score (74.55%), with a precision of 71.20%.

“CNN+LSTM+Attention model led in recall at 78.97% and F1 score at 74.55%, balancing precision and recall while maintaining precision at 71.20%.”

May combine two claims.

In the HEK293 cell line, the CNN model (M2) of Architecture 1 outperformed other models, achieving the highest accuracy (72.62%) and the highest precision (74.65%).

“For the HEK293 cell line, the CNN model (M2) from Architecture 1 outperformed other models, achieving the highest accuracy at 72.62% and precision at 74.65%.”

May combine two claims. The quote only partly states this claim.

In Architecture 2 (Sequential Convolutional Hybrid), the CNN+GRU variant achieved the highest accuracy (74.59%) and the highest balanced accuracy (74.60%).

“ed the highest accuracy at 74.59% and balanced accuracy at 74.60%, proving effective in contexts requiring a balance between sensitivity and specificity.”

May combine two claims.

In the U2OS cell line, the CNN+LSTM+Attention model (M10) of Architecture 3 exhibited the highest accuracy (76.29%) and the highest balanced accuracy (76.23%).

“The CNN+LSTM+Attention model (M10) in Architecture 3 exhibited the highest accuracy at 76.29% and balanced accuracy at 76.23%.”

May combine two claims.

In Architecture 1, the GRU variant achieved the highest recall (80.21%) and the highest F1 score (75.41%).

“Moreover, the GRU variant led in recall at 80.21% and F1 score at 75.41%, balancing precision and recall effectively.”

Probably not a substantial claim: may not pass S3 (load-bearing).

CRISPR-Net employs a recurrent convolutional network to score gRNA-target pairs with mismatches and indels.

“For instance, CRISPR-Net employs a recurrent convolutional network to score gRNA-target pairs with mismatches and indels.”

• cites [3] 3. Lin, J., Zhang, Z., Zhang, S., Chen, J., & Wong, K. C. (2020). CRISPR-net: A recurrent convolutional network quantifies CRIS- PR off-target activities with mismatches and indels. Advanced Science, 7(13), 1903562. https://doi.org/10.1002/advs.201903562 sourceNot checked — no full text of the source

May combine two claims.

The hybrid model surpasses existing models in specific and general genomic environments, achieving the aims outlined in the hypothesis.

“This suggests that the hybrid model surpasses existing models in specific and general genomic environments, successfully achieving the aims outlined in the hypothesis.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The F1 score was used to determine the optimal threshold for predictions, although all listed metrics were evaluated across all architectures.

“Although all the metrics mentioned have been evaluated across all architectures, the F1 score has been used to determine the optimal threshold for predictions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The dataset was pre-processed to remove any rows containing negative cleavage frequencies, so that only biologically plausible data was retained for modeling.

“The dataset was pre-processed to remove any rows containing negative cleavage frequencies, ensuring that only biologically plausible data is retained for further modeling.”

Classified as probably not a substantial claim.

Cleavage frequency was converted into a binary variable using a threshold determined from the median cleavage frequency of the training dataset.

“The dependent variable, which is cleavage frequency, was turned into a binary variable by using a threshold determined from the median cleavage frequency of the training dataset so that anything more than or equal to the median was assigned as 1 while the rest as 0.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The SMOTE-Tomek resampling technique was used to correct class imbalance in the training data.

“Furthermore, the SMOTE-Tomek resampling technique was used to correct any imbalance in the distribution of classes among the training data samples.”

May combine two claims.

Among the learning rates tested, 0.001 allowed the model to efficiently learn complex genomic relationships, reducing the loss function over time and improving prediction sensitivity.

“This study tested various learning rates, finding that 0.001 allowed the model to efficiently learn complex genomic relationships, reducing the loss function over time and improving prediction sensitivity.”