Better data for better predictions: data curation improves deep learning for sgRNA/Cas9 prediction

Datasets

Portions of this text were previously published as part of a preprint (Browne, Edgell & Gloor, 2025). For model development, training, and testing we used three large and three small datasets that measured bacterial Cas9 activity by multiple different assays. Table 1 provides information on the experimental design used to generate these datasets and Table 2 provides information on sequencing depth and the number of available single guide RNAs (sgRNA) for training and validation. The three large datasets measure killing efficiency (depletion) targeting sgRNAs to the host genome which activates cleavage when combined with Cas9 expression; in these assays active sgRNAs are depleted from the library. In the small datasets, activity is measured through cellular growth (enrichment) when the sgRNA/Cas9 complex cleaves a plasmid harbouring the CcdB toxin. Our primary SpCas9 prediction model is constructed from a set of 30,138 sgRNAs and the read counts from nine experimental and one E. coli Epi300 control replicate generated by Ham et al. (2023). This depletion assay measured the activity of sgRNAs in the presence of the dual nuclease TevSpCas9 to cleave the genome of C. rodentium and concomitant loss of the active sgRNA from the pool. The three independent testing sets, also created by Ham et al. (2023), were developed in E. coli K-12, each with 10 control and 10 experimental replicates. Two of these datasets measured cleavage of a CcdB toxin-harbouring plasmid with the TevSpCas9 and SpCas9 nucleases, referred to as the pTox TevSpCas9 and pTox SpCas9 datasets. The two pTox datasets have overlapping sgRNA target sites, providing the opportunity to compare cleavage by the TevSpCas9 and SpCas9 enzymes directly. An additional testing dataset measured TevSpCas9 cleavage of the KatG gene cloned from Salmonella enterica inserted on a plasmid in E. coli, and is referred to as the KatG TevSpCas9 dataset.

We also obtained two additional depletion datasets from Guo et al. (2018) that tested a pool of 65,928 sgRNAs in E. coli. We reconstructed both datasets from publicly available sequencing data and used the original datasets for comparison. For these datasets, E. coli eSpCas9 (n = 45,071) and E. coli SpCas9 (n = 44,163), we use the scores as calculated by Guo et al. (2018) with the 43nt model input sequences from Wang & Zhang (2019), and an improved scoring scheme outlined below. The original datasets and scoring schemes are referred to as the original E. coli eSpCas9 dataset and the original E. coli SpCas9 dataset.

Expanded eSpCas9 and SpCas9 datasets

For the generation of a new E. coli eSpCas9 and a new E. coli SpCas9 model, we reconstruct the two datasets from Guo et al. (2018) using sequencing data generated in E. coli K-12 substr. MG1655. This reconstruction was performed in order to make use of all available data for model training. The original data was curated based upon a minimum of five sgRNAs mapping to an E. coli gene for inclusion—valid for their use case, but unnecessary for model training. The initial sgRNA pool comprised 65,928 sgRNAs targeting the eSpCas9 or SpCas9 nuclease to the E. coli genome. Each dataset consisted of two control and two experimental condition samples, with catalytically dead eSpCas9 or SpCas9 being used for the control condition. The paired-end reads were merged with usearch (Edgar, 2010) and we obtained read counts for sgRNAs from the sequencing output using an in-house script, with sequence reads being counted for both replicates in the control and experimental conditions.

Data processing

The raw read counts for each sgRNA target in control and experimental conditions were used as the input to ALDEx2, with the ALDEx2 diff.btw, a non-parametric log₂ fold-change metric, used as the activity score; hereafter referred to as log₂FC (Fernandes et al., 2014). This metric differs from the Z-score activity metric used by Guo et al. (2018) for the original E. coli eSpCas9 and SpCas9 datasets used for comparison testing in that it is the expected value of a Bayesian posterior derived from the data and ensures that all values are linearly related (Aitchison, 1986; Fernandes et al., 2013). Additionally, given the common use of a log₂FC metric in differential expression analysis, using such values for training may improve the interpretability of resulting predictions. The enrichment pTox TevSpCas9, pTox SpCas9, and KatG TevSpCas9 datasets used a control-condition read cutoff of 20. To test control-condition read count minimum cutoff impacts on activity, we constructed the C. rodentium TevSpCas9, E. coli eSpCas9, and E. coli SpCas9 datasets with incrementally increasing minimum average control-condition read count cutoffs—100 variants of each dataset ranging from a cutoff of 1 to a cutoff of 100 (Fig. 1A).

To determine the effect of adjacent nucleotides on prediction accuracy, we extracted the target site sequences from each datasets associated target DNA sequence—C. rodentium DBS100, E. coli K-12, and the respective plasmid sequences, examining and the impacts to model performance when included alongside the sgRNA and PAM (Fig. 1B). For each set of target sites, we extended the sequence by 500 nucleotides upstream and downstream of the sgRNA plus the NGG PAM. Prior to model input, these sequences are transformed into a model-friendly one-hot encoding format of 4×N shape: four nucleotide options by an N-length sequence input.

During the process of dataset generation, a fixed seed value was used. For the primary datasets used in model development, 20% of the data was separated and remained untouched until final model testing (Table 2). For replication, data generation scripts are made available at the associated repository. During dataset generation, and additionally during model testing, checks were performed to ensure identical input sequences were not present across training and testing sets. Target site sgRNA sequence similarity between each training and held-out test set was checked via hamming distance (Fig. S1).

Model architecture

Throughout we use the term “architecture” when referring to the programming setup, and “model” when referring to the output of the architecture after training. Given the focus of this work on improving predictive performance through a data-centric approach, we train all models on the existing crisprHAL model architecture from Ham et al. (2023) (Fig. 2). This architecture uses a deep learning approach with a dual branching structure comprising convolutional and recurrent neural network layers. The convolutional neural network (CNN) layers aim to learn local features of the target site which impact on-target activity through optimizing the weights of convolutional filters, whereas the recurrent neural network (RNN) aims to learn sequential features (Rumelhart, Hinton & Williams, 1985; LeCun, Bengio & Hinton, 2015). The dual branching structure passes an encoded sgRNA input through an initial convolutional block to either several successive convolutional blocks or a bidirectional gated recurrent unit (BGRU) RNN (Cho et al., 2014). Each convolutional block consists of a one-dimensional CNN followed by a LeakyReLU activation function and one-dimensional max pooling. Each branch then sends their output through two dense layers, each followed by a dropout layer, finally being concatenated into a single output node. This structure is unchanged from the original crisprHAL architecture, save for a small alteration of the BGRU to allow for better GPU usage. A major focus of the original architecture was transfer learning with all parameters frozen in the CNN-branch. Here we do not make use of transfer learning, instead focusing on performance improvement through optimal data curation.

Hyper-parameter adjustments were made in order to better utilize the larger training sets—the original model was designed for use with small dataset fine-tuning. As such, we reduced the learning rate to 0.0005 and increased the batch size to 1,024. Unchanged hyper-parameters include using the “Adam” optimizer with a mean squared error loss function (Kingma & Ba, 2017). For each final presented model, performance is reported following training on the entire training set using hyper-parameters determined by 5-fold cross validation (5CV) using the respective training set. These include the previously mentioned hyper-parameters in addition to the average control-condition read count minima(cutoff), upstream and downstream target site adjacent nucleotide inclusion to the model input, and total training epochs. Each final model architecture and parameter count is available in Table S1.

Prior models

To benchmark our TevSpCas9, eSpCas9, and wild-type SpCas9 models—crisprHAL_Tev, crisprHAL_eSp, and crisprHAL_WT—we use six existing models from three sources (Table 3). These models are: the eSpCas9 and wild-type SpCas9 models from Guo et al. (2018), Guo_eSp and Guo_WT; the eSpCas9 and wild-type SpCas9 models from Wang & Zhang (2019) named here as DeepSgRNA_eSp and DeepSgRNA_WT; and the crisprHAL transfer learning-based TevSpCas9 and wild-type SpCas9 models from Ham et al. (2023), crisprHAL_TL-Tev and crisprHAL_TL-WT. All models are publicly available: crisprHAL (https://github.com/tbrowne5/A-generalizable-Cas9-sgRNA-prediction-model-using-machine-transfer-learning), DeepSgRNA (https://github.com/biomedBit/DeepSgrnaBacteria), and Guo (https://github.com/zhangchonglab/sgRNA-cleavage-activity-prediction). Models have been installed and tested with appropriate configurations as laid out in their respective installation guides. For prior models which do not report specific performance metrics for a held-out test set, when relevant, we compare model performance against their reported 5-fold cross validation metrics.

Model performance metrics

To judge model performance we use the Spearman rank correlation coefficient (ρ) and Pearson correlation coefficient (r); referred to as Spearman correlation and Pearson correlation for simplicity. We report both these metrics to highlight the properties of the predictions since these metrics are concordant in Normally distributed data that is linearly dependent. Across most sgRNA/Cas9 activity prediction models Spearman correlation has been the primary, or only, metric reported, and therefore, this is the main metric we use to benchmark the models. We use the ‘spearmanr’ and ‘pearsonr’ functions from the package SciPy (Virtanen et al., 2020).

An improved bacterial SpCas9 activity prediction model

For the development of an updated and improved SpCas9 prediction model, we use the hybrid convolutional and recurrent neural network-based crisprHAL architecture from Ham et al. (2023) as a framework. For clarity, we denote the new model generated with the C. rodentium dual nuclease-derived TevSpCas9 dataset as crisprHAL_Tev. The C. rodentium TevSpCas9 dataset comprises 25,244 sgRNA target-site sequences and their corresponding on-target activity scores from a depletion dataset (Methods) that uses the non-parametric log₂FC scoring metric derived from the ALDEx2 tool. Here we used a minimum control read count of 56 (Table 2) derived as in the next section, and included 378 nt input sequence for training, derived as described in the third results section. This dataset was split into an 80% training set (n = 20,195) and 20% held-out testing set (n = 5,049), with additional checks performed to ensure no sequence overlap (Fig. S1). The training set was used in model performance optimization and final model generation (Table 3). Hyper-parameters for training epochs, input sequence length, and score cutoff were determined by 5-fold cross validation (5CV).

Following tuning and cross-validation, we observed the optimal mean 5CV performance of 0.754 Spearman correlation and 0.751 Pearson correlation (Fig. S2). We noted that this new model showed similar performance in the held-out 20% test set, with a 0.746 Spearman correlation and 0.744 Pearson correlation (Fig. 3A).

Curation of prior eSpCas9 and SpCas9 data for better models

We now show how the strategy used for the optimized crisprHAL_Tev model is generally useful when using datasets obtained from the literature. To provide maximal opportunity for our model to learn sequence features associated with eSpCas9 activity, we chose to regenerate the dataset without the sgRNA inclusion criteria present in the existing eSpCas9 dataset from Guo et al. (2018). From an initial pool of 64,021 sgRNAs, using a minimum average control-condition read count criteria of 31 (Table 2, and described in later sections), we obtain a final eSpCas9 dataset of 59,489 sgRNAs and corresponding scores—14,419 more than the original version of the dataset from Guo et al. (2018). The sgRNAs not present in the original eSpCas9 dataset are distributed across a similar range of scores to the pre-existing sgRNAs, however, do cluster more towards lower activity (Fig. S3). This expanded dataset was then split into 80% training (n = 47,591) and 20% testing (n = 11,898) sets, with the training set used for all subsequent eSpCas9 model five-fold cross validation testing (Table 3).

Our eSpCas9 model, crisprHAL_eSp, shows improved performance versus the existing eSpCas9 models with a Spearman correlation of 0.782 and a Pearson correlation of 0.780 (Fig. 3B) in the held-out data. This performance is obtained following 62 epochs as determined through five-fold cross validation. The best of the existing eSpCas9 models, DeepSgRNA_eSp, reports a mean five-fold cross validation Spearman correlation of 0.715, which is substantially lower than the value the new model shows with held-out data. The other eSpCas9 model, from Guo et al. (2018), obtains a Spearman correlation of 0.682 on its held-out 20% test set. To highlight the importance of the data, using the original E. coli eSpCas9 dataset, Guo et al. (2018) with the 43nt input used by DeepSgRNA (Wang & Zhang, 2019)—we obtain a Spearman correlation of 0.717 and a Pearson correlation of 0.716 on a held-out 20% test set (n = 9,015).

We also reconstructed a new dataset from the E. coli SpCas9 data from Guo et al. (2018) to support the proposed data curation methods discussed. Using an average control-condition read count cutoff of 73 (Table 2, and later sections), we obtain 33,495 sgRNAs and scores from an initial pool of 61,002—a reduction of 10,668 sgRNAs from original version of the dataset from Guo et al. (2018). Following this curation, the E. coli SpCas9 dataset was split into an 80% training (n = 26,795) and 20% testing (n = 6,700) sets (Table 3).

After training, the E. coli SpCas9 model, crisprHAL_WT, shows a Spearman correlation and Pearson correlation of 0.695 and 0.720 on the held-out test set (Fig. 3C). The Guo_WT and DeepSgRNA_WT models, constructed on the prior E. coli SpCas9 data, obtain 0.542 and 0.582 Spearman correlation—a difference of 0.153 and 0.113 respectively. To provide a direct comparison, akin to the eSpCas9 test, training the crisprHAL model using the original E. coli SpCas9 dataset results in a peak performance of 0.582 Spearman correlation and 0.611 Pearson correlation on their respective held-out 20% test set (n = 8,833). To emphasize the importance of curation, the original model obtains substantially lower performance when using the full-size dataset despite training on 33% more data than does the new model trained on the curated dataset.

Limiting dynamic range inhibits performance

In this section we show the importance of choosing an appropriate minimum read count cutoff for each dataset that maximizes the performance of the model trained on that curated dataset. Model training is reliant upon the accuracy of the data itself, both the input feature and label; i.e., the target site sequence and the activity score. The log₂FC scores are reliant upon sufficient read count abundance in the control condition to facilitate a robust dynamic range in scores for model use. When we examine the entire set of C. rodentium TevSpCas9 sgRNA activity scores versus their control-condition read counts we note a substantial score reduction associated with low control-condition read counts (Fig. 4A). Since these data are generated using a depletion assay, low control-condition read counts are expected to reduce the range of scores obtainable for highly active sgRNA sites; for example, a control read count of 1 allows only a read count of 0 to result in a positive score if an absolute value-based difference between calculation is used. This reduction can be seen across both the E. coli eSpCas9 and SpCas9 activity scores as well (Figs. 4B–4C). Given the potential for these low quality scores to impact model performance, we generated subsets of the C. rodentium TevSpCas9 dataset by incrementally increasing the control-condition read count minimum (Fig. 1A). Comparing a control read count of 1 to the previously used minimum of 20 (Guo et al., 2018), we see a substantial improvement; a 5CV mean Spearman correlation increase of 0.036 (Fig. 4E). As we increase this minimum count cutoff we see a steady increase in performance, peaking at a Spearman correlation of 0.754 with a read count cutoff of 56, shown in purple in (Fig. 4E). This cutoff removes 16.2% of the data; a loss of 4,894 sgRNAs (Fig. 4D). Examining how this cutoff affects the overall distribution of log₂FC scores, we note only a minor reduction in the number of low activity scores (Fig. S4). Given the implication that including these guides inhibits model training, we next examined the error between log₂FC scores and the predictions from crisprHAL_Tev for all sgRNAs eliminated by the 56 cutoff, separated by control read count (Fig. 5). As expected, we note a high mean absolute error for between-model predictions and sgRNAs with low control-condition read counts. When error is separated by positive and negative predictions, the loss of predictive utility is observed to occur in the positive predictions. This is unsurprising, as low control-condition read counts reduce the precision of the readout for potentially active on-target sites. Above control read counts of 10-to-20 the error rate is substantially reduced and slowly decreases, in line with the model performance versus control cutoff results.

When we examine the impact of optimizing the control-condition read count cutoff on the E. coli eSpCas9 dataset we again see an increase in model performance (green line in the figure), though not as substantial as seen in the C. rodentium TevSpCas9 example (purple). Model performance improves consistently as low cutoff values increase, peaking at an average control-condition read count cutoff of 31 reads across the two control replicates. This cutoff negligibly alters the distribution of log₂FC scores compared to a cutoff of 1 (Fig. S4). The use of a read count minimum is likely less impactful to E. coli eSpCas9 model performance due to the higher read counts across the dataset, with comparatively few sgRNAs with low control read counts. Only 4.37% of sgRNAs in the eSpCas9 set have a mean control read count of less than 20 compared to 7.71% and 8.30% in the SpCas9 and TevSpCas9 sets, respectively. It would follow that datasets containing more sgRNAs with low quality activity scores will see a larger performance benefit from their removal.

When this process is applied to the E. coli SpCas9 dataset, we again observe a similar trend, with model performance improving until a minimum average control read count of 73 is reached. At this cutoff the 5CV average Spearman correlation is 0.697, an improvement of 0.029 over a cutoff of 1, and an improvement of 0.014 over using a cutoff of 20 (Fig. 4E). Using this cutoff results in minimal alteration to the distribution of log₂FC scores, with a minor reduction in the already over-represented higher activity scores (Fig. S4). The performance boost using a cutoff of 73 is notable as 45.1% of the total data is removed in this example, reducing the dataset size from 61,002 to 33,495 sgRNAs. The loss of a large fraction of the dataset is unsurprising given that this dataset has the lowest read count coverage across all tested datasets. Thus, we conclude that removing lower quality scores outweighs the benefit of retaining more training data across multiple datasets.

Adjacent nucleotides contain on-target activity information

We next examined the effect on predictive accuracy as a function of the model input sequence length. Prior work has noted the on-target Cas9 activity information contained in the nucleotides adjacent to the sgRNA target site, especially in the nucleotides immediately downstream of the PAM (Wang & Zhang, 2019; Ham et al., 2023). To test the performance benefit of including additional nucleotides upstream and downstream of the sgRNA target site, we train our TevSpCas9, eSpCas9, and SpCas9 models with input sequences of increasing length. We started from just the sgRNA and increased the sequence length incrementally one nucleotide at a time either upstream or downstream, but not both, with performance measured by average five-fold cross validation Spearman correlation (Fig. 1B).

Across all three models, the most substantial boost to performance was obtained when the nucleotides immediately downstream of the PAM were included in the input sequence (Fig. 6A). For crisprHAL_Tev this region extends 11 nucleotides downstream of the PAM, before the performance benefit of adding more nucleotides to the input sequence begins to level off. This trend can be seen when examining average log₂FC activity scores for each di-nucleotide position within and surrounding the sgRNA target, where most information is located between the sgRNA, PAM, and 11 nucleotides downstream (Figs. 6B–6D). Analysis of more distant positional preferences for TevSpCas9 reveals little information, in line with model performance (Fig. S5). We note that the E. coli SpCas9 model obtains the greatest performance boost from this downstream nucleotide inclusion, with a 5CV average increase of +0.07 Spearman correlation units when including the PAM plus 10 nucleotides downstream compared to just using the sgRNA (Fig. 6A). Though the C. rodentium TevSpCas9 and E. coli eSpCas9 models do also see a performance benefit, more information appears to be contained within the sgRNA target site itself, especially in the PAM-distal region.

Examining upstream nucleotides, there is a marginal performance benefit for crisprHAL_Tev when including three nucleotides, with a potential case to be made for including more. However, upon examination of the positional di-nucleotide mean scores across the C. rodentium TevSpCas9 and E. coli SpCas9 datasets, little activity information appears to be contained in this region (Figs. 6B/6D). By contrast, we notice a trend within the E. coli eSpCas9 dataset which is present both upstream of the sgRNA target and beyond 10 nucleotides downstream of the PAM: A-T rich regions are disfavoured (Fig. 6C). Extending the di-nucleotide mean score plots 100 nucleotides upstream and downstream of the sgRNA-PAM site, this trend is maintained (Fig. S6). Though A-T rich adjacent regions are disfavoured, target sites with T rich PAM-proximal regions are very strongly favoured. This PAM-proximal T-preference is present across all three nucleases. There is also a G-preference for the N-position of the NGG PAM, in line with prior work suggesting local sliding of SpCas9 across adjacent PAM motifs (Corsi et al., 2022).

An interesting finding is the use of very long sequence length inputs for E. coli SpCas9 model performance. Appending 189 nucleotides upstream to the sgRNA target model input increases the mean 5CV Spearman correlation by 0.065, while appending the PAM plus 166 nucleotides downstream increases performance by 0.125 (Fig. 6A). Combined, a model built with a 378 nucleotide input sequence comprising 189 upstream, 20 sgRNA target, three PAM, and 166 downstream nucleotides obtains a mean 5CV Spearman correlation of 0.697, 0.143 over the performance seen when using an input of just the sgRNA-target. The sgRNA-target only model attains a mean 5CV Spearman correlation of 0.554. This finding is also repeated with the E. coli eSpCas9 model, with a moderate boost to performance seen when including 193 upstream and the PAM plus 190 downstream nucleotides—a total sequence length of 406. However, for eSpCas9, this can be more easily understood through the disfavouring of A-T rich regions, an attribute not seeming to be present for SpCas9 (Fig. S7). Examining the positional di-nucleotide mean scores for these adjacent regions in the E. coli SpCas9 dataset identifies limited discernible information compared to the sequence length optimization testing. However, if we instead examine sequence length-based performance improvements per added nucleotide, rather than cumulatively, results are similar to the di-nucleotide analyses in terms of signal. Examining per nucleotide position performance improvement—using the crisprHAL_WT model and final input sequence length versus just the sgRNA target site input—we obtain a change in mean Spearman correlation of +0.00034 per upstream and +0.00074 per downstream nucleotide added. While together the longer input sequences improve crisprHAL_WT model performance, as well as crisprHAL_eSp model performance, the per-position impact is minimal.

SpCas9 model predictions and methodology generalize

Since the training sets were all generated using depletion assays, we explored if the data-driven performance improvements extended to datasets generated using a different methodology. We benchmarked crisprHAL_Tev against existing machine learning (ML) models using three independent enrichment datasets: E. coli pTox plasmid TevSpCas9 (n = 260), pTox SpCas9 (n = 255), and KatG TevSpCas9 (n = 268) (Table 2). Across the three testing sets, using both Spearman and Pearson correlation metrics, crisprHAL_Tev achieves better performance than all six prior models tested—the original crisprHAL models, crisprHAL_Tev and crisprHAL_WT, Guo_WT, and DeepSgRNA_WT Cas9 models (Fig. 7). Specifically, the performance metrics for crisprHAL_Tev are: pTox TevSpCas9 0.690 Spearman correlation and 0.693 Pearson correlation, pTox SpCas9 0.650 Spearman correlation and 0.659 Pearson correlation, and KatG TevSpCas9 0.694 Spearman correlation and 0.615 Pearson correlation (Figs. 7A–7C). Since the original crisprHAL models are constructed with the pTox TevSpCas9 and pTox SpCas9 datasets, only their performance on the KatG dataset is used for comparison. Performance from crisprHAL_eSp on the test sets is reported in Fig. S8 due to differences in enzyme preferences. Since SpCas9 and TevSpCas9 do not appear to share the characteristic of eSpCas9 that disfavours AT-rich adjacent regions, the inclusion of such regions in crisprHAL_eSp reduces performance on the non-eSpCas9 datasets. When shorter input sequences from other models are used, that exclude the disfavoured AT-rich regions, crisprHAL_eSp performance on the TevSpCas9 and SpCas9 datasets improves (Fig. S8). We note that although the crisprHAL_Tev model is trained on a dataset which tests the TevSpCas9 nuclease, it still outperforms the models constructed on SpCas9 data on the pTox SpCas9 dataset, supporting the use of a TevSpCas9 data based model for SpCas9 prediction tasks.

A secondary result is the generalization of performance seen from crisprHAL_WT compared to both prior models constructed on the original E. coli SpCas9 data (Figs. 7D–7F). Performance on the three testing sets exceeds both prior models, with the exception of pTox SpCas9, where performance is similar to the Guo_WT model. Using Pearson correlation as test set performance metric, similar results are attained (Fig. S9). We note that this performance is obtained using the ultra long 378 nucleotide model inputs.

Finally, given the data-centric nature of this approach, we tested if the performance improvements identified were specific to the crisprHAL architecture. To perform this, we retrained the five-layer CNN architecture from the DeepSgRNA model using the three training datasets. Across both Spearman and Pearson correlation metrics, mean 5CV performance corresponded with the results obtained when using the crisprHAL architecture (Fig. S10). Specifically, the input sequence lengths of 37 for C. rodentium TevSpCas9, 406 for E. coli eSpCas9, and 378 for E. coli WT-SpCas9 obtained improved performance compared input sequence lengths of 43, 30, and 28 used by prior models, and the baseline 20nt sgRNA target site only input (Wang & Zhang, 2019; Guo et al., 2018; Ham et al., 2023). Using the final input sequence lengths with the five-layer CNN model resulted in mean 5CV Spearman correlations improvements of 0.036 (TevSpCas9), 0.102 (eSpCas9), and 0.160 (WT-SpCas9), compared to the baseline 20nt input. Performance also benefited from using the higher control-condition read count cutoff data, compared to the previously used cutoff value of 20 and baseline cutoff of 1 (Guo et al., 2018; Wang & Zhang, 2019). Compared to a cutoff of 1, 5CV Spearman correlations increased by 0.052(TevSpCas9), 0.008 (eSpCas9), and 0.023 (WT-SpCas9).