Modelling the dynamic range of second-generation toehold switches

Three open-source tools were developed for machine-learning based design of RNA devices, namely toehold switches. They have been developed for second-generation toehold switch designs, but could be adapted for first-generation toehold switches. Our software is made freely available for the scientific community under GNU GPLv3.

(1) GrammarParser.py:

Used to return the domain structure of a toehold switch, i.e, parsing the sequence into its structural segments based on the grammar of dot-bracket representations. The dot-bracket representation could be obtained by ViennaRNNA's RNAfold.

Input: Toehold switch sequence (toehold_seq) & predicted dot-bracket representation (dot-bracket_rep)
Output: Parsed domain sequences
Usage: python GrammarParser toehold_seq dot-bracket_rep

(2) predict_linear.py:

Used to predict the efficacy (i.e, dynamic range) of a new toehold switch. Takes the non-redundant engineered feature values in order as arguments, and returns the dynamic range of the construct:

Input (six Features in order): Overall Switch MFE Bottom Region MFE RBS Linker MFE Net MFE Frequency of MFE structure Specific Heat at 37 deg C
Output: Dynamic range of the given toehold switch construct
Usage: python predict.py InputFeatures

(3) nn_model.ipynb:

This Jupyter notebook contains the code to validate our neural network model. To use it for predicting efficacies of new toehold switches, the path to a csv file with the instances/features may be specified in the final code block in the indicated place; the header line and first two fields are ignored. See here for an example. Then use the following code, where the path to the saved model (see section on NN_models) may also be provided:

import pandas as pd 
from keras.models import load_model
new_feats= pd.read_csv('/path/to/csv_file')
new_feats = new_feats.iloc[:,2:]
#normalize the features as in the original script; use the StandardScaler fit on the training data. Then:
model = load_model('/path/to/saved_model/',custom_objects={'r_square':r_square})
print(float(model.predict(new_feats)))

(4) toehold_efficacy_predict.sh:

An integrated script written in bash that provides an end-to-end pipeline for the prediction of toehold efficacy for second-generation toehold switches.  

The script does the following:

(i) calls ViennaRNA RNAfold to parse the input sequence into its dot-bracket representation

(ii) calls GrammarParser to then extract the segments of the toehold switch based on this dot-bracket representation

(iii) calls more ViennaRNA RNAfold utilities to obtain engineered feature values for the given toehold switch sequence

(iv) passes these feature values as input to 'predict_linear.py` and returns the dynamic range of the toehold switch sequence. Alternatively the user may use nn_model.py described above for making the predictions.

Input: Toehold switch sequence(s) in .fasta or .txt format.
Output: Predicted dynamic range(s) of given toehold switch sequence(s), and engineered feature values.
Usage: sh ./toehold_efficacy_predict.sh Input_Seq_File

Requirements:

(1) ViennaRNA must be installed and available in the path.

(2) Python 3.x with pandas, numpy, matplotlib, sklearn, tensorflow and keras libraries.

(3) bash shell

Data

The data folder consists of the engineered features for the 228 toehold instances from the literature [1,2,3]:

(1) all_features.csv : consisting of all the engineered sequence and structural features

(2) selected_features.csv : consisting features with the best predictive capabilities, post feature-selection (see our manuscript for details).

(3) miR_toeholds.csv: consisting features of the toehold switches designed for the two miRNA biomarkers of cervical cancer identified in our study, viz. mir-21-5p and miR-20a-5p.

NN_Models

This folder contains the saved_models of the best-performing neural network models (adj. R2 > 0.5) for making batch-predictions on new toehold switches. The models are suffixed with a number, indicating their rank. To find the ON/OFF ratio of designed toehold switches using any of these models, load the model and predict.

import pandas as pd 
from keras.models import load_model
new_feats= pd.read_csv('/path/to/csv_file')
new_feats = new_feats.iloc[:,2:]
#normalize the features as in the original script; use the StandardScaler fit on the training data. Then:
model = load_model('/path/to/saved_model/',custom_objects={'r_square':r_square})
print(float(model.predict(new_feats)))

Video demo

Please find a short video tutorial here explaining how to use our scripts, validated using Takahashi et al ¹.

Citation

Please cite our work:

Baabu PRS, Srinivasan S, Nagarajan S, Muthamilselvan S, Selvi T, Suresh RR, Palaniappan A. End-to-end computational approach to the design of RNA biosensors for detecting miRNA biomarkers of cervical cancer. Synth Syst Biotechnol. 2022 Apr 4;7(2):802-814. doi: 10.1016/j.synbio.2022.03.008. PMI D: 35475253; PMCID: PMC9014444.

biorxiv preprint (2021).

Contact

For comments, queries and more information, please contact Ashok Palaniappan

Primary sources for building our dataset:

Green AA, Silver PA, Collins JJ, Yin P. Toehold switches: De-novo-designed regulators of gene expression. Cell 2014. https://doi.org/10.1016/j.cell.2014.10.002.
Pardee K, Green AA, Takahashi MK, Braff D, Lambert G, Lee JW, et al. Rapid, LowCost Detection of Zika Virus Using Programmable Biomolecular Components. Cell 2016;165:1255–66. https://doi.org/10.1016/j.cell.2016.04.059.
Green AA, Kim J, Ma D, Silver PA, Collins JJ, Yin P. Complex cellular logic computation using ribocomputing devices. Nature 2017. https://doi.org/10.1038/nature23271.

Takahashi, M.K., Tan, X., Dy, A.J. et al. A low-cost paper-based synthetic biology platform for analyzing gut microbiota and host biomarkers. Nat Commun 9, 3347 (2018). https://doi.org/10.1038/s41467-018-05864-4 ↩