In order to generate a model that can predict gene-regulatory activity from ribozyme sequence, we first generated a data set of diverse ribozyme sequences and their gene-regulatory activities to train the model. We wanted the designed training set library to contain ribozymes with diverse sequences and structures so that the model would be able to learn which nucleotide interactions between the loop regions can support cleavage in physiological environments. In addition, because the model would be used to design ribozymes that exhibit low gene-regulatory activities in the absence of ligand, it was important that the training set contain thousands of ribozymes with low gene-regulatory activities, which are normally rare in ribozyme libraries with randomized sequences.

We designed two ribozyme libraries to capture the desired diversity and cleavage activities. The first library was designed by randomizing the loop regions (loops I and II) with nucleotide sequences of lengths 5–10 on the base of sTRSV hammerhead ribozyme scaffold (Figure 2A). Libraries were cloned into the 3′ untranslated region (UTR) of a GFP expression cassette encoded on a low-copy plasmid, such that ribozymes with high cleavage activities result in cells expressing low GFP levels. A constitutive mCherry expression cassette was included on the ribozyme library screening plasmid and used to normalize for intrinsic gene expression noise due to factors like cell cycle state and plasmid copy number (Liang et al., 2012).

Figure 2 with 1 supplement see all

Download asset Open asset

The gene-regulatory activities of individual ribozyme sequences in the libraries were measured via FACS-Seq(15) (Figure 2B). FACS-Seq uses fluorescence-activated cell sorting (FACS) to partition cells harboring libraries of regulatory elements into different activity bins based on the cell’s ratio of GFP to mCherry levels and next-generation sequencing (NGS) to determine the number of times a ribozyme sequence occurs in a particular bin. For a given sequence, the distribution of counts across the bins is used to generate a mean measured value of gene-regulatory activity associated with that ribozyme. The first designed ribozyme library was transformed into Saccharomyces cerevisiae cells and FACS-Seq was used to obtain gene-regulatory activity values for approximately 75,000 ribozyme sequences with at least 20 reads each; the resulting data set is designated as FACS-Seq 1 (Figure 2—figure supplement 1A). 20,000 of these ribozymes have at least one loop with a length of ≥8, which is the size of the smallest aptamer loops. 48% of the sequences exhibit activity that is greater than half of that exhibited by a nonfunctional ribozyme, and only 7% (1432 sequences) displayed activity near that of the native sTRSV hammerhead ribozyme. The activity data indicates that most of the possible nucleotide sequence combinations for loops I and II do not result in functional nucleotide interactions that support high cleavage activity at physiological conditions.

The second library was designed and characterized to ensure a larger fraction of low gene-regulatory activity ribozyme sequences. The library was designed with one of the loops (loop I or II) harboring a randomized sequence of larger length (N9–N13) and the other loop (loop II or I, respectively) harboring a shorter randomized sequence (N5–N7). The design of the larger loop was meant to capture contexts in which an aptamer is placed onto either loop I or II, such as in tertiary ribozyme switch designs, and to allow the model to extend to the design of those more complex architectures (Figure 2C).

The second library was characterized through two modified FACS-Seq experiments designed to increase the representation of low gene-regulatory activity ribozymes, producing the data sets FACS-Seq 2 and FACS-Seq 3. In FACS-Seq 2, cells harboring the ribozyme library were pre-sorted by separating out the cells with the highest 80% expression levels from those with the lowest 20%. The two populations were cultured separately for several days and then combined in a ratio such that the low-expressing cells comprised 80% of the final population (Figure 2D), which was subsequently analyzed by FACS-Seq. This recombined population, FACS-Seq 2, still contained ribozymes that spanned a broad range of expression (Figure 2—figure supplement 1B). Of the 31,000 sequences that had a loop length of ≥8, even with the prescreening, 24% still exhibited high gene-regulatory activities (greater than half of the activity exhibited by a nonfunctional ribozyme). However, although the loops of this library were of much larger size, which generally leads to high gene-regulatory activity (Townshend et al., 2015), another 1,424 low -activity sequences were isolated, thus doubling the number of ribozymes in our training set with gene-regulatory activity close to that of the native ribozyme. To obtain additional training data on low gene-regulatory activity ribozymes, we subjected the second ribozyme library to a more stringent pre-sort prior to the FACS-Seq analysis. Specifically, we separated out the cells with the lowest 5% expression levels and ran this cell population through the FACS-Seq protocol, creating another training set designated FACS-Seq 3 (Figure 2—figure supplement 1C). Due to the stringent pre-sorting process, the final library was enriched in low gene-regulatory activity ribozymes, adding another 1,158 (of the 5,000 sequences in FACS-Seq 3 that had a loop length of ≥8) unique low gene-regulatory activity ribozymes with large loops to the training set.

Through the design and testing of multiple ribozyme libraries, we generated individual measurements of gene-regulatory activities for roughly 150,000 unique ribozyme sequences (10,000 of which exhibit gene-regulatory activities close to that of the native ribozyme), achieving a data set with a broad range of activities. The libraries contain diverse sets of loop structures, with many combinations of loop sizes containing hundreds to thousands of unique sequences (Figure 2E). Loops ranged in size from 3 nucleotides (e.g., an N5 loop where the first and last nucleotides could base pair) to 13 nucleotides. Of the larger loops (length ≥10 nucleotides), between ⅓ and ½ of the loops in the training set contain a stem branching off of the loop, providing significant structural diversity to our training set. The library also contains a diversity of stem lengths, with the first stem (normally 6 base pairs long) extending up to 11 base pairs and the second stem (normally 4 base pairs) extending up to 8 base pairs. Approximately 60,000 of the ribozymes in the characterized library have a loop with length >8, and 4,000 of these large-loop sequences displayed near-native gene-regulatory activity. The thorough data sets generated from this work can be used to train a robust model for predicting ribozyme activity from sequences spanning diverse loop sizes and structures.

Before assessing different machine learning models for ribozyme design, we developed a method by which the features of a given ribozyme (i.e., sequence, structure) are encoded for input into the model. Given the mechanisms underlying ribozyme cleavage, it is important for the encoding to retain as much information as possible about the physical system and the relationship between the features. For example, when performing image analysis, the data is kept in a two-dimensional array (with extra dimensions added to represent color) that retains the physical relationship between different areas of the image.

Our approach to encoding the ribozyme features focuses on retaining the features’ physical relation to each other. The method starts by using the RNA secondary structure prediction program RNAstructure (Reuter and Mathews, 2010) to determine the structure (and therefore the sequence) of loops I and II of the ribozyme (Figure 3A). While the sequence of the ribozyme is known, which nucleotides comprise a loop is not known until the structure is determined due to the fact that different loop sequences can produce different loop structures. One such example is if the first two nucleotides in the loop after the stem can form a Watson–Crick pair, which would extend the stem further and therefore make the following two nucleotides the actual first ones in the loop. Another example is ribozymes with longer loop sequences that can form stems extending off of the loop, which functionally removes some of the nucleotides from the loop. Once the loop sequence is determined, turning stems I and II and loops I and II of the ribozyme are turned into an image with two spatial dimensions, each nucleotide occupying one ‘pixel’ and another dimension that represents the base and whether or not it is paired (Figure 3B). To simulate the stem-loop I and stem-loop II interactions, the two-dimensional images that represent stem-loops I and II are paired with each other, creating a three-dimensional (3D) image that represents the stem-loops. The described encoding method preserves the spatial orientation of the nucleotides in stem-loops I and II and includes all nucleotides in the stem-loops for analysis. One potential drawback of the encoding method is that it makes some assumptions regarding the spatial arrangement of the nucleotides, specifically that the stem-loop nucleotides occupy a relatively thin plane, which extends in one direction as the loop grows larger. Evidence from crystal structures for the native hammerhead ribozyme supports this assumption of the nucleotides being arranged in two thin opposing planes (Chi et al., 2008).

Figure 3 with 2 supplements see all

Download asset Open asset

The described ribozyme encoding method is optimally paired with a CNN. The algorithm begins by analyzing small groups of features at a time before combining them to make a final prediction. In our model, we applied a 3D convolution that was two nucleotides long along each edge, creating a window that captures each group of eight adjacent nucleotides (Figure 3C). The window scans along the length of the array, moving one nucleotide at a time, creating a set of one-dimensional arrays. The one-dimensional arrays are converted to a single prediction through a set of fully connected layers and dropout layers. The resulting model is sufficiently complex to learn which sequences affect cleavage, while being simple enough that it does not overfit to the training data with a low number of training epochs (Figure 3—figure supplement 1). The main disadvantage of a neural network model is that it is not transparent in how it makes its predictions, such that it is difficult to obtain mechanistic insight into ribozyme cleavage from the model. However, model interpretability is a secondary concern to accuracy and robustness to structural differences as the primary function of the model is to design new ribozyme switches given a known aptamer sequence.

As it is generally preferred to use as simple a model as possible for a given objective, we compared the performance of the described model with a version of a simple linear regression model known as a lasso model (Figure 3—figure supplement 2). We devised a series of tests to determine which model would be most useful for the predictive design of ribozyme switches. The model’s objective is to make accurate predictions of basal gene-regulatory activities for ribozymes that have an aptamer replacing either loop I or II.

We first compared the two models on accuracy of predicting gene-regulatory activity from ribozyme sequence. The data set of roughly 150,000 ribozyme sequences with measured gene-regulatory activities was randomly separated into a training subset (75% of the sequences) and a test subset (25% of the sequences). The neural network model and lasso model were trained on the training subset using backpropagation to minimize mean-squared error on the predicted log10(GFP/mCherry) value when compared to the measured log10(GFP/mCherry). The trained models predicted the gene-regulatory activities of ribozymes in the test subset. A model’s accuracy was evaluated by generating an R² value when analyzing the measured and predicted basal gene-regulatory activity values. We first tested the model’s ability to predict the simplest ribozymes we had measured data on in order to establish a baseline of predictive power. These were ribozymes with four random nucleotides on each loop (Figure 4A). Both models were able to predict the activities in the simulated test subset (25% of the constructed training set) with high accuracy; however, the neural network (R² = 0.91) was more accurate than the lasso algorithm (R² = 0.82). (Figure 4B). In addition to these predictive methods, we also tested a support vector machine (R² = 0.45) and a random forest (R² = 0.92) algorithm on the same N4N4 data set (Figure 4—figure supplement 1A). The random forest also scored well in its prediction of the gene-regulatory activity of simple ribozymes. Although they had a better R² value overall, ribozymes that were predicted to have low gene-regulatory activity by the neural network model were equally likely to have low measured gene-regulatory activity than those predicted by the lasso model (mean log10(GFP/mCherry) value of –1.41 to –1.40 for the 10 ribozymes with the lowest predict gene-regulatory activity).

Figure 4 with 3 supplements see all

Download asset Open asset

We next evaluated the model’s ability to predict gene-regulatory activities of ribozymes with more complex structures. An earlier study performed a high-throughput screen on ribozymes with an aptamer integrated into one of the two loops of the sTRSV hammerhead ribozyme (Townshend et al., 2015). We used that published data set as a test set for evaluating the model’s predictive ability on more complex ribozyme structures. In order to simulate the design of a novel ribozyme switch, we used ribozymes with a theophylline aptamer as one loop for a test set for predictions of gene-regulatory activity in the absence of ligand (Figure 4C). We used the trained neural network model and the lasso model to predict the basal gene-regulatory activity of each sequence in the theophylline ribozyme switch data set and compared the predicted activities with measured values as described above (Figure 4D). Both models were less accurate in predicting the activities of more complex ribozyme sequences than when predicting the activities of N4N4 ribozymes; however, the neural network model generated an R² value that was much higher than the R² value produced by the lasso model (0.27 vs 0.08). As before, we also tested a support vector machine (SVM) (R² = 0.002) and a random forest (R² = 0.06) algorithm, which performed worse than either the neural network or lasso algorithms (Figure 4—figure supplement 1B). The neural network model was also more accurate in predicting which ribozymes would display low gene-regulatory activity. The 10 ribozymes with the lowest neural network-predicted activities had a mean log10(GFP/mCherry) value of –1.04, while the best lasso-predicted ribozymes had a mean log10(GFP/mCherry) value of –0.69. This example offers an approximation to how the models would perform when designing novel ribozyme switches, and the data indicate that the neural network model has more accurate predictive power.

In addition to the level of accuracy, it is also desirable for the model to be robust to differences in loop structures. The ribozymes comprising the training data set exhibit a variety of loop lengths, but aptamers can comprise structures that are more complex than a simple stem-loop. We therefore compared the two models in the accuracy of their activity predictions for ribozymes that have loop structures that differ from the ribozymes from which they have been trained on. To do this, we grouped the ribozymes from our training data set into different segments based on the size of the loops and length of the stems of each ribozyme such that each ribozyme in a segment shares the same structure (Figure 4E). We then took a group, held it out as a test set, and combined the remaining segments into a data set that was used to train the two models, which were then used to predict the gene-regulatory activities of the test set. This was repeated for each segment of ribozymes with a shared structure in our data set, generating a measure of accuracy (R²) for each model. For each test segment of ribozymes with a shared structure, the neural network’s predictions had an R² value an average of 0.18 greater than the predictions from the lasso model (Figure 4F). Furthermore, the neural network’s R² values were higher than the lasso model for every single structural segment, and the segments where the stem length was different than the standard ribozyme stem had an even greater difference in R² (0.53).

Taken together, the results from the testing of the neural network and lasso model indicate that the former is the better choice from a forward design standpoint due to making more accurate predictions and being more robust to novel loop structures than the lasso model. Additional optimization was performed to determine the optimal size of the training set (Figure 4—figure supplement 2) and the effect of the structural difference between ribozymes in the training set and the test set on accuracy (Figure 4—figure supplement 3). It is particularly important to note that the accuracy of the neural network model declines rapidly as the structure of ribozymes in the training set differs from the structure of the ribozymes whose gene-regulatory activity the model is predicting (Figure 4—figure supplement 3C). These properties will be important in developing a model to design ribozymes with aptamers as one of the two loops, as those sequences and specific structures are not necessarily present in our training data set. However, the diverse set of structures in the training set is to our advantage as it increases the chance that there are shared structures between the training set of ribozymes and the ribozymes with aptamers.

We next examined the model’s ability to perform the forward design of tertiary-interaction ribozyme switch libraries. Specifically, by using the model to screen in silico for sequences that have a greater likelihood of exhibiting low gene-regulatory activities in the absence of ligand, these sequences can comprise a smaller ‘designed library’ that can subsequently be efficiently screened in the presence and absence of the cognate ligand for high-performing switches (Townshend et al., 2015). A model that can determine which switches are more likely to exhibit low basal gene-regulatory activity can be used to design libraries enriched for sequences that are more likely to be switches, enabling a more effective screen. The model was tested with five small molecule aptamers, specifically sequences that bind folinic acid (Chang, 2014), theophylline (Zimmermann et al., 1998), tetracycline (Berens et al., 2001), chloramphenicol (Burke et al., 1997), and neomycin (Weigand et al., 2008). The aptamer-ligand pairs span a diverse set of chemical structures (Figure 5A), RNA sequences, and secondary structures (Figure 5B), enabling us to examine the ability of the model to design ribozyme switch libraries for diverse aptamer sequences.

Figure 5 with 2 supplements see all

Download asset Open asset

The libraries for each of the five aptamers were designed similarly. The ligand-binding loop in the aptamer and any stems extending off that loop were attached directly to either stem I or II to create loop I or II. The loop opposite the aptamer loop was set as all possible sequences comprising N5–N8 to create a library of 174,080 ribozyme sequences per aptamer (Figure 5C). The model evaluated each sequence generated in this library in silico using the neural network model (trained on the generated training data set) to assign a predicted gene-regulatory activity (in the absence of ligand) value to each sequence (Figure 5—figure supplement 1). As the ribozymes used in the training set did not contain an aptamer structure on either loop, the model had to extrapolate its learnings to a new ribozyme sequence and structure. We ranked the examined sequences for each aptamer by their predicted gene-regulatory activity values and selected the 200 sequences with the lowest predicted values as our ‘designed library’.

The activities of every member of the designed libraries for each aptamer were characterized in a single FACS-Seq experiment, FACS-Seq 4. The 200 sequences comprising each aptamer’s designed library were combined into a 1,000 sequence library that was synthesized using a custom oligonucleotide pool. In addition to the designed library, we also constructed a random library using oligos encoding N5–N8 for the variable loops. The random library and the designed library were combined and analyzed through a FACS-Seq experiment. Through the FACS-Seq analysis, we obtained basal gene-regulatory activity values for 90% of the sequences from the designed libraries across all aptamers and nearly 20,000 sequences from all the aptamers in the random libraries (Figure 2—figure supplement 1D).

We compared the distribution of gene-regulatory activity values in the absence of ligand for sequences in the designed library with sequences in the random library to assess the model’s ability to design ribozymes that exhibit low levels of basal gene-regulatory activity for each aptamer (Figure 5D). We evaluated the libraries using three principal metrics: (i) the mean gene-regulatory activity for sequences in each library (log10(GFP/mCherry)); (ii) the fraction of sequences with basal gene-regulatory activities in the lower third of the expected activity range (bounded by the gene-regulatory activity a fully functional and fully non-functional ribozyme, i.e., below log10(GFP/mCherry) = 0.0); and (iii) the statistical confidence measurement via a Wilcoxon Rank-Sum test, indicating if the designed and random libraries were drawn from different underlying distributions (Table 1). For all aptamers, sequences exhibiting low gene-regulatory activity in a random library are rare; no more than 3% of sequences in any random library exhibit basal gene-regulatory activity levels in the lower third of the activity range.

For all aptamers, the designed library was enriched in low gene-regulatory activity sequences compared to that aptamer’s random library. However, the designed and random libraries exhibited different distributions of gene-regulatory activities, indicating that certain aptamers are more amenable to producing ribozymes with low basal gene-regulatory activity values as determined by the three principal metrics. For example, the analysis of the folinic acid aptamer libraries indicates that the neural network model was successful in designing ribozymes with low levels of gene-regulatory activity for this aptamer. The designed library had a mean gene-regulatory activity value of log10(GFP/mCherry) = –0.36 vs. 0.62 for the random library. The fraction of ribozymes in the designed library with low gene-regulatory activity was 77% vs. 3% for the random library. Finally, the results of the rank-sum test determined that the two libraries had a 0.0% probability (p=1.2 × 10^–91) of coming from the same distribution.

Three of the remaining four aptamers (theophylline, tetracycline, chloramphenicol) exhibited similarly successful results in the application of the neural network model for designing ribozymes with low gene-regulatory activities. Specifically, the designed libraries had mean gene-regulatory activity values of 0.31 (theophylline), 0.30 (tetracycline), and 0.38 (chloramphenicol) vs. 0.55, 0.44, and 0.49, respectively, for the corresponding random libraries. The fraction of ribozymes in the designed libraries with low gene-regulatory activity was 15% (theophylline), 8% (tetracycline), and 0.6% (chloramphenicol) vs. 2%, 1%, and 0.1%, respectively, for the corresponding random libraries. While for this set of aptamers the design libraries exhibited more overlap with the corresponding random libraries than the folinic acid libraries in terms of activity, the results of the rank-sum test similarly showed a very low probability of the two libraries for each aptamer coming from the same distribution (theophylline: p=1.0 × 10^–34; tetracycline: p=8.0 × 10^–36; chloramphenicol: p=1.6 × 10^–23). Taken together, the model was successful in each case of designing ribozymes that exhibit low gene-regulatory activities and enriched for such sequences relative to a random library.

Of the five aptamers examined, designing low gene-regulatory activity sequences for the neomycin aptamer proved particularly difficult. The designed neomycin aptamer library had a mean gene-regulatory activity value of log10(GFP/mCherry) = 0.64 vs. 0.73 for the random library. However, the fraction of ribozymes in the designed library with low gene-regulatory activity was 0% vs. 0.01% for the random library. Finally, the results of the rank-sum test determined that the two libraries had a very low probability of coming from the same distribution (p=4.8 × 10^–42). The results indicate that while the machine learning model was able to enrich for sequences exhibiting lower gene-regulatory activity relative to a random library, it was not successful in designing sequences in the lower third of the activity range for this aptamer.

Even with the variability inherent in aptamer sequences, the data indicate that the neural network model robustly identifies variable loops that more frequently exhibit lower basal gene-regulatory activity when the opposite loop is constrained to an aptamer sequence than would be expected by random variable loops. Thus, the model is useful in designing ribozyme switch libraries for subsequent screening. Although high-throughput screening methods such as FACS-Seq can measure many thousands of ribozyme sequences in a single experiment, bottlenecks including transformation efficiency of the cell host can limit the number of sequences that can be practically evaluated (Olsen et al., 2000; Xiang et al., 2019). Thus, the model can be used as a tool to design libraries of lower sequence diversity biased to those sequences that are likely to exhibit low gene-regulatory activities (Mishra et al., 2008).

We next examined the model’s ability to perform the forward design of tertiary-interaction ribozyme switches starting with a given aptamer sequence. In particular, rather than designing tailored libraries that can subsequently be screened for functional switches, the prediction accuracy of the model was sufficient to directly test a small number of the returned highest ranked sequences. To examine the forward design capability of the model, we chose the 15 ribozyme sequences for each aptamer that the model predicted to have the lowest gene-regulatory activities in the absence of ligand (i.e., the ‘predicted best’). As the model’s objective is solely to design ribozyme sequences with low gene-regulatory activity, inherent in using it to design tertiary-interaction ribozyme switches is the assumption that once such a ribozyme sequence is achieved with an aptamer in loop I or II, binding of the ligand to the aptamer will disrupt the loop interactions and thus the ability of that ribozyme sequence to adopt a catalytically active state at physiological conditions.

Each identified ribozyme sequence was individually synthesized and cloned into the previously described switch characterization plasmid in the 3´ UTR of the GFP reporter. Flow cytometry was used to measure the GFP to mCherry ratio of a clonal cell population harboring a switch characterization plasmid under standard growth conditions and in the absence and presence of high concentrations (1–5 mM) of the cognate ligand for the aptamer (Figure 6—source data 1 ). Gene-regulatory activities were normalized to an inactive control ribozyme to set the theoretical upper limit on activity levels as 100% (Figure 6—source data 1 ).

The model performs consistently well in predicting variable loops that will pair with the folinic acid aptamer to result in very low gene-regulatory activity. Similar to the aptamer’s designed library, the top 15 ribozymes designed with the folinic acid aptamer exhibited very low basal gene-regulatory activities, with a median of 3.2% of the nonfunctional ribozyme and a very low interquartile range (the difference between the 75th and 25th percentiles) of 1.8% (Figure 6A). In the presence of high concentrations of folinic acid, the median activity increased to 4.7%, resulting in a median activation ratio of 1.5 (Figure 6B). The top three performing ribozymes exhibited activation ratios spanning 2.2–2.9 (Figure 6C). While two of the top three switches by activation ratio had low basal gene-regulatory activities (2.5% and 4.0%), the third exhibited the highest basal gene-regulatory activity (23.4%) of the top 15 predicted ribozymes. Within the top 15 ribozymes, the sequences that displayed an increase in activity in the presence of folinic acid had the aptamer located in loop I of the ribozyme. Although the model correctly predicted ribozymes exhibiting low gene-regulatory activities with the aptamer located in loop II, that particular architecture either did not allow for ligand binding to the aptamer or for ligand-dependent change in ribozyme cleavage. Overall, the folinic acid-responsive ribozymes did not exhibit large increases in gene-regulatory activities in response to ligand, even though the activities in the absence of ligand were very low. Given the relatively high affinity of the aptamer to its ligand (Chang, 2014), these results imply that folinic acid does not cross the cell membrane efficiently, thus resulting in very low intracellular concentrations (Xiang et al., 2019; Niethammer and Kleihauer, 1974), or does not bind to the aptamer in a way that effectively disrupts ribozyme cleavage.

Figure 6

Download asset Open asset

Figure 6—source data 1

Activity of designed ribozymes in the presence and absence of their cognate ligand.

https://cdn.elifesciences.org/articles/59697/elife-59697-fig6-data1-v2.docx

Download elife-59697-fig6-data1-v2.docx

The model also performs consistently well in predicting variable loops that will pair with the tetracycline aptamer to result in low gene-regulatory activity. The top 15 ribozymes designed with the tetracycline aptamer exhibited low basal gene-regulatory activities, with a median of 26.3% of the nonfunctional ribozyme. There was relatively low variation (interquartile range of 15.8%) in activity across the 15 sequences tested. In the presence of high concentrations of tetracycline, the median activity increased to 75.6%, resulting in a median activation ratio of 2.8. The top three performing ribozymes exhibited activation ratios spanning 4.5–5.8, and all exhibited low basal gene-regulatory activities (spanning 10.7–14.3%). The model predictions for the design of ribozymes with the tetracycline aptamer and folinic acid aptamer provide a notable contrast in that in both instances the model reliably predicted sequences with low gene-regulatory activities. The greater success of the tetracycline-responsive ribozymes as genetic switches relative to the folinic acid-responsive ribozymes likely results from particular features of the ligand-aptamer pair, including the ability of tetracycline to more efficiently cross the cell membrane (Argast and Beck, 1984).

The model designs with the chloramphenicol aptamer further illustrate that features specific to the ligand-aptamer pair can limit the resulting switch performance, independent of the model’s ability to predict variable loop sequences that lead to low gene-regulatory activities. The top 15 ribozymes designed with the chloramphenicol aptamer exhibited low basal gene-regulatory activity with a median of 41.9% of the nonfunctional ribozyme, similar to that of the ribozymes designed with the tetracycline aptamer. In addition, there was relatively low variation (interquartile range of 20.7%) in activity across the 15 sequences tested. Unlike the tetracycline-responsive ribozymes, the median gene-regulatory activity only increases slightly to 46.1% in the presence of high concentrations of chloramphenicol, resulting in a median activation ratio of 1.1. Among the top three performing ribozyme sequences was one that exhibited a statistically significant activation ratio of 1.3, although its basal gene-regulatory activity was higher than the median chloramphenicol ribozyme basal activity (51.3%). The data demonstrate that the model is able to predict a ribozyme sequence with the chloramphenicol aptamer that exhibits relatively low gene-regulatory activity; however, these sequences exhibited little response to the cognate ligand and thus limited function as genetic switches. Similar to the folinic acid ligand-aptamer pair, these results imply that chloramphenicol does not cross the yeast membrane efficiently (Abdel-Sayed, 1987), its toxicity at high concentrations limits the intracellular concentrations attainable, or it does not bind to the aptamer in a way that effectively disrupts ribozyme cleavage.

The model designs with the theophylline aptamer highlight that certain aptamers result in greater variability in the model’s ability to predict variable loops that pair with that aptamer to result in low gene-regulatory activity. Although the model produced several ribozymes with low gene-regulatory activities, on average the ribozymes among the predicted top 15 sequences exhibited high gene-regulatory activities in the absence of theophylline, with a median of 77.0% of the nonfunctional ribozyme. In addition, there was substantial variability in the distribution of gene-regulatory activities (interquartile range of 28.1%) across the 15 tested sequences, which was the highest amongst the aptamers tested. In the presence of high concentrations of theophylline, the median activity increased to 80.6%, resulting in a median activation ratio of 1.2. However, due to the variability in activities across the top 15 predicted sequences, the top three performing ribozyme switches exhibited basal gene-regulatory activities as low as 12.7% and activation ratios up to 4.0. Overall, limitations in ligand permeability across the cell membrane or ability of ligand binding to disrupt ribozyme cleavage did not appear to restrict the design of theophylline-responsive ribozymes to the degree observed with some of the other ligands.

Finally, the model’s inability to predict variable loops that will pair with the neomycin aptamer to result in low gene-regulatory activity limited the successful design of high-performing switches with this aptamer. Similar to the aptamer’s designed library, the top 15 ribozymes designed with the neomycin aptamer did not exhibit low gene-regulatory activities, with a median of 72.5% of the nonfunctional ribozyme. There was relatively large variation in activity across the 15 sequences tested, with an interquartile range of 29.2%. Despite the high gene-regulatory activities in the absence of ligand, a substantial increase in activities was observed in the presence of neomycin, with a median of 98.7%, resulting in a median activation ratio of 1.4. The top three performing sequences exhibited basal gene-regulatory activities as low as 44.3% and activation ratios up to 2.0, thus reaching near the theoretical upper limit of gene-regulatory activity in the presence of ligand (i.e., 100%). The data demonstrate that similarly to tetracycline the neomycin ligand-aptamer pair exhibits properties desirable in the design of genetic switches, particularly the availability of the ligand at intracellular concentrations sufficient to disrupt ribozyme cleavage upon binding its aptamer. While overall the median activities of the sequences predicted by the model for the neomycin aptamer are similar to those predicted for the theophylline aptamer, the model did predict two low gene-regulatory activity sequences for theophylline in the top 15, which was not similarly achieved for neomycin. These results indicate that features of the neomycin aptamer make the model less able to predict loop sequences that will pair with the aptamer to result in ribozyme sequences exhibiting low gene-regulatory activities.

Taken together, the results indicate that the neural network model can be used as a tool for the forward design of individual tertiary-interaction ribozyme switches, eliminating the need for more time-consuming methods like ad hoc design or high-throughput screening. Overall, the design tool produced ribozyme sequences with basal gene-regulatory activities below 33% (or at a level to enable an activation ratio of >3 for aptamer-ligand pairs with optimal features for genetic switches) for four of the five aptamers examined. In addition, the model’s predictions resulted in genetic switches exhibiting statistically significant activation ratios for all aptamers examined. While certain aptamers resulted in consistent predictions from the model, others resulted in greater variability. The results from applying the model to these five aptamers support that, in general, testing 15 of the top sequence predictions is sufficient to produce ribozyme sequences with low gene-regulatory activities for those aptamers whose sequences are intrinsically amenable to creating cleaving ribozymes. The results also show that while designing ribozymes with low basal gene-regulatory activities is an important aspect for generating a switch, there are features specific to each aptamer-ligand pair that can affect switch performance and are not captured by the model.

Finally, we examined the model predictions to determine any trends that provide insight into the sequence-function relationship of the hammerhead ribozyme’s loops. We analyzed the sequences of the variable loops in the designed libraries for each aptamer to determine if the model was designing sequences that were specifically optimized for each aptamer or settled on a conserved motif that generally works well and applied that across diverse aptamer sequences. We determined which variable loop sequences appeared across different aptamers (Figure 7A). The results of this analysis indicate that 1.5% of the variable loop sequences in the designed libraries are generally favored by the model independent of the aptamer sequence. Specifically, 4 sequences were in every aptamer’s designed library (of 200 sequences) and a further 11 sequences were in four of the five libraries. These sequences share a 5-nucleotide motif on loop II (UGGAG) that the model deemed generally favorable for cleavage (Figure 7B). We searched in our training set for ribozymes with UGGAG on loop II and a random loop I, and the data indicate that harboring this sequence greatly increases a ribozyme’s probability that it will exhibit low gene-regulatory activity levels (Figure 7C). Sequences with this motif displayed a mean log10(GFP/mCherry) value of –0.27 vs. a mean log10(GFP/mCherry) value of 0.02 for ribozymes with random loop sequences. Thus, this analysis indicates that the model learned from the training data that this motif is generally likely to result in low gene-regulatory activity levels. In addition, the designed libraries indicate that aptamers may be preferentially located on specific loops to achieve a ribozyme exhibiting low gene-regulatory activities. Specifically, 75% of the folinic acid and 51% of the theophylline designs have the aptamer on loop II, while 83% of the tetracycline, 100% of the neomycin, and 100% of the chloramphenicol designs have the aptamer on loop I. The designed libraries also demonstrate that the preferred variable loop size varies by aptamer. The neomycin and chloramphenicol aptamers show a strong preference for loop lengths of five or six (91 and 100% of designs, respectively), whereas the tetracycline aptamer shows a strong preference for longer loop lengths of seven or eight nucleotides (85% of designs). In contrast, the folinic acid and theophylline aptamers are distributed equally across all loop lengths.

Figure 7

Download asset Open asset

Though certain motifs were predicted independent of the aptamer sequence, roughly 75% of the sequences in each aptamer’s designed library were unique to that library. The lack of a shared motif for the variable loop regions across different aptamers indicates that the model is responding to specific features of the aptamer. The analysis also indicates that the predicted 15 best sequences for each aptamer had different motifs represented (Figure 7D). The UGGAR motif that was shared between the designed libraries for multiple aptamers was present in the top designs for the folinic acid, theophylline, and chloramphenicol aptamers and produced switches with basal gene-regulatory activities below 30% for each of these aptamers. The top performing switches for folinic acid and theophylline included sequences with the aptamer on loop II, with different motifs (UANNNG and ANAUGGG, respectively). The top designs for tetracycline and neomycin had different motifs that were enriched in guanine and adenine nucleotides (AGGGA and AGGGG, respectively). One commonality among the various motifs present in the top predicted sequences for the aptamers tested is an enrichment of purine nucleotides, adenine (34.5%) and guanine (49.3%), when compared with the pyrimidines uracil (13.6%) and cytosine (2.6%).

Overall, the area of machine learning concerned with interrogating and explaining neural networks, especially neural networks as applied to genomics, is still new, but as it develops, one could begin applying lessons from that field to explain what contributes to predicted ribozyme activity. Because our model appears to be an accurate predictor of said activity, it could also provide further insights into the biophysical mechanisms governing ribozyme activity.

For the ribozyme libraries, oligonucleotides were ordered that contained N5−N14 in place of the two loops (Supplementary file 1). The oligonucleotides were designed such that the forward oligonucleotides contained the first half of the ribozyme, from the 5´ spacer sequence through the catalytic core, including the random library for loop I. The reverse oligonucleotides contained the catalytic core through the 3´ spacer sequence, including the random library for loop II. The overlap in the catalytic core area allowed for annealing during the PCR assembly. The spacer sequences were designed to minimize interactions between the surrounding sequences and the switch sequence. The switch libraries were constructed through a one-pot PCR reaction using the KAPA HiFi PCR Kit (Kapa Biosystems). For the PCR, 10 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 53°C for 30 s, and an extension step of 72°C for 15 s. This was followed by 20 cycles with a melting step of 94°C for 30 s, an annealing step of 65°C for 30 s, and an extension step of 72°C for 15 s. For library 1, both the forward and reverse oligos containing N5–N10 were used in one reaction. For library 2, two reactions were prepared – one with forward primers containing N5–N7 and reverse primers containing N9–N14 and one with forward primers containing N9–N14 and reverse primers containing N5–N7 (Supplementary file 1). These mixtures were amplified with the primers Gaprepair_F and Gaprepair_R (Supplementary file 1).

The assembled ribozyme libraries were cloned via homologous recombination into the plasmid backbone pCS1748 via electroporation in yeast. Briefly, 50 mL yeast culture (OD600 = 1.3–1.5) was incubated with Tris-DTT buffer (2.5 M DTT, 1 M Tris, pH 8.0) for 20 min at 30°C, pelleted, washed, and resuspended in buffer E (10 mM Tris, pH 7.5, 2 mM MgCl₂) to 250 µL. To 60 µL of the yeast cell suspension, 2 µg of linearized plasmid and 1 µg of library insert DNA were added, and the DNA-cell suspension was electroporated (2 mm gap cuvette, 540 V, 25 µF, 1000 Ω). Transformed cells were diluted to 1 mL volume in yeast peptone dextrose medium and incubated for 1 hr at 30°C. The incubated cultures were spun down and resuspended in 100 µL of water, which was then transferred directly into 50 mL YNB-URA liquid media for growth in 30°C. 100 μL of a 1/10,000 dilution of each transformation was also plated on YNB-URA agar plates to obtain the transformation efficiencies. Six transformations were in parallel done for each library to ensure sufficient coverage.

Every 12 hr, the cells harboring the ribozyme libraries were back-diluted into fresh 100 mL YNB-URA media and grown at 30°C to ensure that cells remained in exponential phase. This was repeated for six re-passagings. Replicate transformations were combined or discarded once the transformation efficiencies were determined so that the number of transformed cells reached an appropriate amount, generally 10× the targeted number of sequences. The libraries were passaged for 3–4 days prior to sorting.

For cell sorting, the passaged library was back-diluted to an OD of 0.1 in two separate 100 mL YNB-URA media flasks. The two cultures served as technical replicates, ensuring that the expression, sorting, and sequencing did not introduce significant variability in the measurement of ribozyme gene-regulatory activity levels. The two cultures were grown for 7 hr, then spun down at 600 rpg and resuspended in 2 mL phosphate buffered saline (PBS) (Life Technologies) + 1 % Bovine serum albumin (BSA). The resuspended cultures were kept on ice until sorting.

Sorting was performed on a FACSAria II FACS machine (Beckman Coulter) at the Stanford FACS Core. 8 bins were generated using a Matlab script (https://github.com/btownshend/TwoColor, Townshend, 2018), which ensured that each bin except the first and last spanned the same log10(GFP/mCherry) distance. GFP was excited at 488 nm and measured with a splitter of 505 nm and bandpass filter of 525/50 nm. mCherry was excited at 532 nm and measured with a splitter of 600 nm and bandpass filter of 610/20 nm. Initial gates based on the forward-scatter area and side-scatter area were used to gate out cell debris and nonviable cells. The bins were set so that the upper bin and the lower bin each contained roughly the top and bottom percentile of expression levels, respectively, with the remaining six bins capturing the remaining levels of expression. 20–25 million cells from each of the replicate cultures were sorted into the various bins based on the ratio between GFP and mCherry expression. Cells falling into bins 1–4 were collected in the first sort, and all other cells were discarded. In the second sort, cells falling into bins 5–8 were collected. 5 mL YNB-URA was added to each bin from the sort, though more YNB was added to some bins so that the ratio was at least 1:2 PBS:YNB. The cultures for the recovered bins were grown up overnight and 1 mL of each bin was miniprepped the next day using the ZR Fungal/Bacterial DNA MiniPrep kit (Zymo Research) according to the manufacturer’s instructions. The rest of the culture from each recovered bin was spun down and frozen as backup.

After several days of post-transformation passaging, certain experiments performed a pre-sort on the library to achieve a desired frequency of ‘low-expressing’ cells. In these pre-sorts, cells harboring the libraries were sorted into low-expression and high-expression groups as appropriate. The gates for sorting were set so that a certain bottom percentile of expressing cells were designated as ‘low-expressing’ while the top percentile of expressing cells were ‘high-expressing’. For library 2, the low-expressing cells were defined as those in the bottom 20% of the entire library 2 population. For library 3, the low-expressing cells were defined as those in the bottom 5% of the entire library 3 population. Post-sorting, the sorted cell groups were passaged into fresh 100 mL YNB-URA media and maintained for 2 days before the final eight-bin sort. If the low- and high-expressing populations were both included in the final sort, as in library 2, they were mixed at the point of the back-dilution before the final 7 hr of growth so that the appropriate ratio of low- and high-expressing cells was obtained.

Miniprepped DNA from the binned sorts were first PCR-amplified using primers GFP_F and ADH1t_R (Supplementary file 2) and the KAPA HiFi PCR Kit (Kapa Biosystems) to isolate from the miniprepped DNA the portion of the plasmid that contained the ribozyme library. For the PCR, 15 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 57°C for 30 s, and an extension step of 72°C for 15 s. The resulting PCR products were barcoded using the primers listed in Supplementary file 2, with each bin PCR-amplified using a different combination of forward and reverse primers that added different 5′ and 3′ barcodes. For the PCR, 15 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 57°C for 30 s, and an extension step of 72°C for 15 s. The forward primer contained a barcode indicating whether the bin contained the induced or uninduced library (BT1334 and BT1335), while the reverse primer contained a barcode indicating which bin was being amplified (BT609–BT1462). For the PCR, 15 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 57°C for 30 s, and an extension step of 72°C for 15 s. The relative concentrations of each amplified bin were determined using qPCR with primers BT235 and BT236 and Fast Plus EvaGreen qPCR master mix (Biotium). Each amplified bin was tested at 1/100, 1/400, 1/1000, and 1/4000 dilutions. The concentration measured for these four dilutions was averaged together to determine the concentration for that bin. The amplified bins were mixed together so that the final mixture would contain a fraction of molecules from each bin in proportion to the number of cells that were originally sorted into that bin. For example, if 50 million total cells were sorted into bins during the FACS portion and 5 million of those were sorted into bin 3, then the number of molecules from the amplified bin 3 made up 10% of the final mixture. The final mixture was PCR-amplified using primers BT235 and BT236 (Supplementary file 2) to increase the concentration. Samples were quantified on both a Bioanalyzer 2100 (Agilent Technologies) and a Qubit 3.0 Fluorometer (Thermo Fisher Scientific). The resulting product was sequenced using a MiSeq machine (Illumina) at the Stanford PAN Facility with a 2 × 150 kit. Samples were run using Illumina standard procedures, with PhiX (Illumina) added (to 15% by molarity) to further increase diversity at nucleotide positions that would otherwise have a significant fraction of the sample sharing the same base call and result in lower read quality.

NGS reads were paired using the paired-end read merge program PEAR (Zhang et al., 2014). The paired reads were analyzed by a Python script that recorded for each sequence the number of times that sequence appeared in the reads from each bin, normalized by the total number of reads that were obtained from that bin. A normal distribution was fitted to the normalized number of reads for each bin to obtain a mean GFP/mCherry value. The fits were performed using custom Python code available at https://github.com/smolkelab/Ribozyme_switch_generator. The fits resulted in an estimate for each sequence of the GFP/mCherry ratio.

We used the RNA secondary structure prediction software RNAstructure (Reuter and Mathews, 2010) for the computational folding of ribozymes as it has good accuracy, with a correlation between predicted and known RNA structures among the highest of all RNA secondary structure prediction methods (Puton et al., 2013), on short RNA sequences and can fold several sequences of ribozyme length (roughly 100 nucleotides) per second on a standard desktop computer. To confirm this, we used the stochastic folding tool from RNAstructure, which gives a distribution of predicted structures, on each of the designed ribozyme switches for the five aptamers that were predicted to have the lowest basal gene-regulatory activity. For each of these ribozyme switches, we gathered a distribution of 100 predicted structures and did not observe an alternate structure among this test set. We have included this data in the supplementary information (Supplementary file 5). In order to facilitate the high-throughput folding of hundreds of thousands of ribozymes, we use a Python script to control RNAstructure’s command line interface and collect the minimal free energy structure for each sequence of a list of ribozyme sequences. The Python script then automatically analyzes each ribozyme’s structure, pulling out important features such as loop size, loop sequence, and stem length. The library of folded ribozymes is stored as a Python list where each entry is a single ribozyme sequence with its corresponding structure, structural components, and, for ribozymes whose activities have been measured through FACS-Seq, basal and activated log10(GFP/mCherry) values. This list is stored long term as a Python list using the Pickle package. The code for performing the high-throughput computational determination of ribozyme structure can be found at https://github.com/smolkelab/Ribozyme_switch_generator/blob/master/Generate_candidate_list.py and https://github.com/smolkelab/Ribozyme_switch_generator/blob/master/Fold_candidate_list.py.

The data set of ribozymes with known gene-regulatory activities is retrieved from storage by using Pickle to load the list of ribozyme sequences and structures into memory. The ribozymes are then grouped by ribozyme structure into Python dictionaries where all the ribozymes in each dictionary have the same structure. When training and test sets need to be made, the relevant dictionaries are combined. These dictionaries are then put through a function that generates an array from the dictionary that represents the encoded ribozymes. The array generation functions work by iterating through each ribozyme in the input dictionary, analyzing its structure and loop sequence as described. The code for encoding of a ribozyme’s sequence and structure for machine learning can be found at https://github.com/smolkelab/Ribozyme_switch_generator/blob/master/Train_additional_models.py.

To implement a lasso algorithm, we use the built-in function linear_model.LassoCV from the scikit-learn machine learning library (Pedregosa et al., 2011). This model is fit to the array generated for the training data using gradient descent and generates an activity prediction based on the test array. We use Google’s tensorflow (Martín, 2015) to run the CNN using the Keras package for the easy implementation of the described design (keras.io). The model is trained on the training data by using the Adam optimizer (Kingma, 2014) to minimize the mean squared error on the training set, iterating over 100 epochs. After training, the model can be used to generate a predicted basal activity for a test array. The code for training a model and using it for prediction can be found at https://github.com/smolkelab/Ribozyme_switch_generator/blob/master/Train_additional_models.py and https://github.com/smolkelab/Ribozyme_switch_generator/blob/master/Predict_activities.py.

For a given aptamer, the bulge where the ligand binds was identified through literature. If there was no literature suggesting a ligand-binding area, the bulge closest to the terminal stem was chosen. The stem leading up to this bulge was removed and the remaining nucleotides of the aptamer were attached to either the first or second stem of the ribozyme, therefore completing that loop (Figure 5—figure supplement 2). The opposite ribozyme loop was randomized to an N5–N8 sequence. The full set of designed ribozyme switch sequences was put through the machine learning protocol for the computational determination of secondary structure, encoding, and gene-regulatory activity predicting. The 200 sequences with the lowest predicted basal gene-regulatory activity levels were chosen as the best candidates. This protocol of turning an aptamer into a library of possible ribozyme switches, predicting their basal gene-regulatory activity, and selecting the 200 sequences that are predicted to have the lowest basal gene-regulatory activity was repeated with the five aptamers to theophylline, folinic acid, neomycin, tetracycline, and chloramphenicol, resulting in 1000 candidate sequences (Supplementary file 6).

The oligos used for constructing the designed library of variable loops were ordered as an oligo pool (Twist Bioscience) and assembled into a library of switches through PCR. The forward primers contained the 5′ spacer through the first half of stem II, including either the aptamer or the variable loop I. The reverse primers contained the second half of stem I through the 3′ spacer, including either the aptamer or variable loop II (Supplementary file 4). Two PCR mixes were made – one mix with forward primers containing aptamers and reverse primers containing variable loop II and one mix with forward primers containing variable loop I and reverse primers containing aptamers. These mixtures were PCR-amplified with the primers switch_GFP_AvrII_Fwd and switch_ADH1t_XhoI_Rev (Supplementary file 4). For the PCR, 10 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 53°C for 30 s, and an extension step of 72°C for 15 s. This was followed by 20 cycles with a melting step of 94°C for 30 s, an annealing step of 65°C for 30 s, and an extension step of 72°C for 15 s. In addition to a designed library, a random library was generated by using primers with random sequences of N5–N8 in the place of the variable loop. These primers were used to construct a random library in the same PCR-based protocol through which the designed library was constructed.

The assembled libraries were transformed alongside the digested plasmid backbone pCS1748 into electrocompetent yeast as described previously. Three transformations were performed for each library to ensure sufficient coverage.

All riboswitch and backbone plasmid sequences are listed in Supplementary file 4. Designed ribozymes and ribozyme switches were assembled from primers and amplified through PCR with the primers switch_GFP_AvrII_Fwd and switch_ADH1t_XhoI_Rev (Supplementary file 4) using PFU Ultra II HS (Agilent Technologies). For the PCR, 10 cycles were performed with a melting step of 94°C for 30 s, an annealing step of 53°C for 30 s, and an extension step of 72°C for 15 s. This was followed by 20 cycles with a melting step of 94°C for 30 s, an annealing step of 65°C for 30 s, and an extension step of 72°C for 15 s. The plasmid backbone pCS1748, which contains both yEGFP and ymCherry under the control of two separate TEF1 promoters, was digested overnight with restriction enzymes AvrII and XhoI. The assembled ribozyme cassette was cloned into the plasmid backbone through a Gibson Reaction (Gibson et al., 2009) and transformed into Top10 E. coli for recovery and sequence verification.

Individual plasmids with designed ribozymes cloned between yEGFP and the ADH1 terminator were transformed into chemically competent CSY3 using the Frozen-EZ Yeast Transformation II chemical competence kit (Zymo Research) according to the manufacturer’s instructions. Three colonies from each transformation were picked into a 96-well plate with 500 μL YNB-URA in each well. Cultures were grown overnight in a 30°C shaker. 10.5 μL of saturated culture was then transferred into 200 μL YNB-URA either with or without the inducing molecule. The concentrations used were 5 mM for folinic acid, 1 mM for tetracycline, 25 μg/mL for chloramphenicol, 5 mM for theophylline, and 5 mM for neomycin. Cultures were grown for 6 hr in a 30°C shaker. The plates were spun at 3500 rpm for 5 min and the supernatant was discarded. The remaining cells in each well were resuspended in 200 μL PBS + 1% BSA and filtered.

Single-cell fluorescence levels were recorded using the VYB flow cytometer (Miltenyi), measuring both GFP and mCherry fluorescence levels for up to 10,000 cells per sample. GFP (the FL7.A channel) was excited at 488 nm and measured with a splitter of 505 nm and bandpass filter of 525/50 nm. mCherry (the FL4.A channel) was excited at 532 nm and measured with a splitter of 600 nm and bandpass filter of 610/20 nm. Initial gates based on the forward-scatter area and side-scatter area were used to gate out cell debris and nonviable cells. Cells that displayed an FL4.A reading of less than 10^-3.5 were removed from analysis as these cells had lost the mCherry gene through recombination. The normalized expression level (GFP/mCherry) for an individual cell was calculated by dividing the cell’s FL7.A reading by its FL4.A reading. For each sample, the GFP/mCherry values of all the cells in the sample were averaged to calculate an overall GFP/mCherry for the sample. For each switch, the samples are averaged together to obtain a mean GFP/mCherry value for that switch.

The GFP/mCherry values of all switches were normalized by dividing the GFP/mCherry value for that switch by the GFP/mCherry value for yeast transformed with a control ribozyme that does not cleave. For each switch, its uninduced GFP/mCherry value was divided by the GFP/mCherry value of an uninduced control ribozyme sample, while the switch’s induced GFP/mCherry value was divided by the GFP/mCherry value of an induced control ribozyme sample. The activation ratio for a switch was calculated by dividing the normalized induced GFP/mCherry by the uninduced GFP/mCherry.

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

We thank Dr Ben Kotopka, Dr James Payne, and Deze Kong for comments during the preparation of the manuscript; Dr Brent Townshend for guidance in developing and executing the FACS-Seq protocols; and Dr Ben Kotopka for expertise in developing and troubleshooting machine learning models. Cell sorting/flow cytometry analysis for this project was done on instruments in the Stanford Shared FACS Facility. This work was supported by funds from the National Institutes of Health (grant to CDS) and the National Institute of Standards and Technology (training grant to CMS).

© 2021, Schmidt and Smolke

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.