PRIDICT: Predicting Efficiencies of Prime Editing Guide RNAs

By Guest Blogger

Guest Blogger June 25, 2024

Prime editing is a versatile genome editing technology that allows precise modifications of DNA (replacements, small insertions, and deletions) without introducing DNA double-strand breaks (Anzalone et al., 2019; Chen & Liu, 2023). This method uses a prime editor (typically a Cas9 H840A nickase fused to a reverse transcriptase) together with a prime editing guide RNA (pegRNA). The pegRNA consists of a target-complementary spacer, a scaffold sequence, and an extension sequence containing a primer binding site and a reverse transcription template (RTT). The design of the pegRNA has been shown to play a critical role in the editing efficiency of prime editing, whereby a multitude of configurations can be used to introduce specific target modifications. Since this means we could design hundreds to thousands of different pegRNAs for each desired edit, it simply becomes impossible (or at least unreasonable) to test them all. This is where PRIDICT comes in!

Evaluate prime editing efficiencies in high-throughput libraries

My labmates and I wanted to find a way to predict the efficiencies of pegRNA designs based on patterns observed from high-throughput experiments, where we can evaluate tens of thousands of designs in parallel. To do this, we performed high-throughput ‘self-targeting’ screens with oligo libraries where pegRNAs are coupled to their target allowing us to edit and evaluate all of the pegRNAs in one experiment (Figure 1). Through this method, we found key determinants of prime editing efficiency such as the length of the RTT overhang, the GC content of the primer binding site, and many more. Once we had our datasets with all the different features and editing efficiencies, we could train our prediction model PRIDICT (Mathis & Allam et al., 2023). We trained our models in close collaboration with the Krauthammer lab (University of Zurich), especially with my deep learning-wizard colleague Ahmed Allam. Continued training led to the recent development of PRIDICT2.0 (Mathis et al., 2024). The additional datasets used in the training of PRIDICT2.0 make this model more suited for prediction of very diverse edit types and cell contexts.

Graphic showing various pegRNA designs leading to a self-targeting library, then PE is added in to HEK293T cells (MMR-) and K562 (MMR+) cells labeled Prime editing, leading to editing analysis, leading to machine learning, and ending in PRIDICT/PRIDICT2.0, labelled as sequenced-based prediction.

Figure 1: High-throughput screening of pegRNA efficiency and building prediction tools with machine learning (PRIDICT, PRIDICT2.0).

How to apply PRIDICT in prime editing experiments

Prediction models only become useful once they find their way into the hands of users. For this reason, we made our tools available via the PRIDICT website and also for offline use via GitHub (which includes prediction in batch mode). To predict the best pegRNA design for a specific edit, users can first define their edit in brackets and add the upstream and downstream flanking sequences to their input string (Figure 2). This string can then be used as input for the prediction on the website or offline via GitHub. The model provides users with different designs of pegRNAs, ranked by their PRIDICT/PRIDICT2.0 score. For the best prime editing performance, we suggest that users select some of the best designs (e.g., top 5) and test them in their labs.

Figure 2: Guidelines to predict pegRNA efficiencies with pridict.it.

This approach leverages the strengths of PRIDICT/PRIDICT2.0 while accounting for some of the inherent inaccuracies of computational prediction models. In addition, we provide options for nicking gRNAs (used for PE3 and PE5 prime editing strategies) based on a DeepSpCas9 score ranking (Kim et al., 2019).

PRIDICT saves time and resources by selecting the most promising pegRNA designs and especially facilitates the use of prime editing in pooled settings with hundreds to thousands of pegRNAs.

Chromatin-based prime editing efficiency prediction

While sequence-based prediction models from self-targeting screens such as PRIDICT/PRIDICT2.0 provide valuable insights into pegRNA editing efficiencies for specific targets (Mathis & Allam et al., 2023, 2024; Yu et al., 2023; Koeppel et al., 2023), their predictive power is sometimes limited across different targets due to variations in chromatin context. Our recent studies (Mathis et al., 2024) and concurrent research by the Shendure lab (Li et al., 2024) have highlighted how chromatin substantially affects prime editing outcomes across many genomic locations.

In our study, we used the TRIP (Thousands of Reporters Integrated in Parallel; Akhtar et al., 2013; Schep et al., 2021) method to link prime editing efficiency to chromatin data from over a thousand genomic locations. This allowed us to train ePRIDICT (Figure 3), a model that predicts how well a genomic site can be edited based on its chromatin environment. After selecting the best pegRNA design with PRIDICT/PRIDICT2.0, ePRIDICT can help you understand how chromatin might influence editing results at your target.

A graphic showing the steps of high-throughput screening. TRIP library integration is first; next is prime editing; then editing analysis; then assign chromatin features; then machine learning; and finally chromatin-based predictions.

Figure 3: High-throughput screening of the effect of chromatin on prime editing efficiency and building a chromatin-based prediction tool (ePRIDICT).

In summary, machine learning-based prime editing prediction models (such as PRIDICT/PRIDICT2.0) can greatly facilitate prime editing experiments, thus saving you time, sweat, and money on extensive optimization efforts. This holds true for any prime editor experimentalist, and especially shines once large-scale prime editing experiments come into play.

Nicolas Mathis received his PhD from Gerald Schwank's lab at the University of Zurich (Switzerland), where he developed a strong passion for computational biology. His expertise includes applying machine learning techniques to make sense of biology, with a special focus on gene editing techniques.

References and resources

References

Akhtar, W., De Jong, J., Pindyurin, A. V., Pagie, L., Meuleman, W., De Ridder, J., Berns, A., Wessels, L. F. A., Van Lohuizen, M., & Van Steensel, B. (2013). Chromatin Position Effects Assayed by Thousands of Reporters Integrated in Parallel. Cell, 154(4), 914–927. https://doi.org/10.1016/J.CELL.2013.07.018

Anzalone, A. V., Randolph, P. B., Davis, J. R., Sousa, A. A., Koblan, L. W., Levy, J. M., Chen, P. J., Wilson, C., Newby, G. A., Raguram, A., & Liu, D. R. (2019). Search-and-replace genome editing without double-strand breaks or donor DNA. Nature, 576(7785), 149–157. https://doi.org/10.1038/s41586-019-1711-4

Chen, P. J., & Liu, D. R. (2023). Prime editing for precise and highly versatile genome manipulation. Nature Reviews Genetics, 24(3), Article 3. https://doi.org/10.1038/s41576-022-00541-1

Kim, H. K., Kim, Y., Lee, S., Min, S., Bae, J. Y., Choi, J. W., Park, J., Jung, D., Yoon, S., & Kim, H. H. (2019). SpCas9 activity prediction by DeepSpCas9, a deep learning–based model with high generalization performance. Science Advances, 5(11), eaax9249. https://doi.org/10.1126/sciadv.aax9249

Koeppel, J., Weller, J., Peets, E. M., Pallaseni, A., Kuzmin, I., Raudvere, U., Peterson, H., Liberante, F. G., & Parts, L. (2023). Prediction of prime editing insertion efficiencies using sequence features and DNA repair determinants. Nature Biotechnology 2023, 1–11. https://doi.org/10.1038/s41587-023-01678-y

Li, X., Chen, W., Martin, B. K., Calderon, D., Lee, C., Choi, J., Chardon, F. M., McDiarmid, T. A., Daza, R. M., Kim, H., Lalanne, J.-B., Nathans, J. F., Lee, D. S., & Shendure, J. (2024). Chromatin context-dependent regulation and epigenetic manipulation of prime editing. Cell. https://doi.org/10.1016/j.cell.2024.03.020

Mathis, N., Allam, A., Kissling, L., Marquart, K. F., Schmidheini, L., Solari, C., Balázs, Z., Krauthammer, M., & Schwank, G. (2023). Predicting prime editing efficiency and product purity by deep learning. Nature Biotechnology, 41, 1151–1159. https://doi.org/10.1038/s41587-022-01613-7

Mathis, N., Allam, A., Tálas, A., Kissling, L., Benvenuto, E., Schmidheini, L., Schep, R., Damodharan, T., Balázs, Z., Janjuha, S., Ioannidi, E. I., Böck, D., Steensel, B. van, Krauthammer, M., & Schwank, G. (2024). Machine learning prediction of prime editing efficiency across diverse chromatin contexts. Nature Biotechnology. https://doi.org/10.1038/s41587-024-02268-2

Schep, R., Brinkman, E. K., Leemans, C., Vergara, X., van der Weide, R. H., Morris, B., van Schaik, T., Manzo, S. G., Peric-Hupkes, D., van den Berg, J., Beijersbergen, R. L., Medema, R. H., & van Steensel, B. (2021). Impact of chromatin context on Cas9-induced DNA double-strand break repair pathway balance. Molecular Cell, 81(10), 2216-2230.e10. https://doi.org/10.1016/J.MOLCEL.2021.03.032

Yu, G., Kim, H. K., Park, J., Kim, J., Kim, J., Kim, H. H., Kwak, H., Cheong, Y., & Kim, D. (2023). Prediction of efficiencies for diverse prime editing systems in multiple cell types. Cell, 186, 1–17. https://doi.org/10.1016/j.cell.2023.03.034