Overview

Our project “PERspectives” aims to implement the perceptron algorithm in vitro-via the senders and the receivers population- in order to perform pattern recognition tasks. Since we are referring to a basic machine learning algorithm, our project would not be fully implemented without a software application in order to perform the training and in silico design of the genetic circuit. Thus, we created the PERspectives app, which has the following purposes:

1. Implement the software training of the perceptron, offering some tuning parameters and some accuracy metrics as a result
2. Correspond each classes’ weight to the correct RBS, via our RBS library
3. Allow advanced users to import their own RBS sequences as well as delete already existing ones (via the RBS name)

Figure 1. Input page when first running the application.py script.

Interface

In terms of the interface, it is split in two sections; a left one, where the user can provide the inputs, the training dataset as well as modify the RBS library and a right one, where the results of training and modification can be viewed.

Furthermore, users can download sample datasets not only for training but also for importing novel RBS sequences via download buttons in the right section.

Figure 2. Download buttons for sample datasets (training and RBSs).

Building the app

In terms of training the model, we tested it in 2 different datasets: one used in the pattern recognition section of our model and the breast cancer dataset offered by the sklearn python package. Since we have already shown the usage of our software on the first dataset for extracting the correct RBS sequences per class, in this section we will use it for the latter one.

After importing our dataset via the “Import Your Dataset” button on the right section, and using a learning rate of 0.03 for epochs we received a training accuracy of 89.3%.

Figure 3. Training Results for the sklearn breast cancer dataset.

Our Use Case

However, in terms of our biological implementation of the perceptron, the application offers an RBS for each class from the RBS library which correctly corresponds to the weight value(%) received by the training of the perceptron. For our example above, the result received was the following:

Weight Class	Weight Value	Sign	Translation Rate [a.u]	RBS_Name	RBS_Sequence
mean radius	12.235467715472511	Negative	19580.10876	synthetic_RBS_20	GGGCCCAAGUUCACUUAAAAAGGAGAUCAACAAUG AAAGCAAUUUUCGUACUGAAACAU CUUAAUCAUGCUGCGGAGGGUUUCUA
mean texture	13.559208074786707	Positive	30195.18	BCD6	UUGGCAACGGCUUAUGGGAGGGAUGACA
mean perimeter	11.116287961601783	Negative	13106.64212	synthetic_RBS_26	GGGCCCAAGUUCACUUAAAAAGGAGAUCAACAAUGAA AGCAAUUUUCGUACUGAAA CAUCUUAAUCAUGCUAUGGAGGUUUUCUA
mean area	12.573070309567562	Negative	39353.4991	synthetic_RBS_4	UUGGCAACGGGUUAUCGGAGGGAUGACA
mean smoothness	15.98002809868399	Positive	73526.53956	synthetic_RBS_7	UUGGCAACGGCUUAUGGGAGGGAUGACA
mean compactness	1.9926475349802477	Positive	31126.53073	synthetic_RBS_14	UUGGCAACGGCUUAUGGGAGGGAUGACA
mean concavity	2.3803258156223253	Positive	78686.23401	synthetic_RBS_6	GGGCCCAAGUUCACUUAAAAAGGAGAUCA ACAAUGAAAGCAAUUUUCGUACUGAAACAU CUUAAUCAUGCGCCGGAGGUUUUCUA
mean concave points	0.5270090142745103	Positive	106815.75	BCD12	GGGCCCAAGUUCACUUAAAAAGGAGAUC AACAAUGAAAGCAAUUUUCGUACUGAAAC AUCUUAAUCAUGCAUCGGACCGUUUCUA
mean symmetry	17.35597929400305	Positive	78686.23401	synthetic_RBS_6	GGGCCCAAGUUCACUUAAAAAGGAGAUCAA CAAUGAAAGCAAUUUUCGUACUGAAACAUCUU AAUCAUGCGCCGGAGGUUUUCUA
mean fractal dimension	7.870227407285138	Positive	134925.9866	synthetic_RBS_13	UUGGCAACGGCUUAUGGGAGGUAUGUCA
radius error	3.583395175979027	Negative	30195.18	BCD6	UUGGCAACGGCUUAUCGGAGGUAUGACA
texture error	5.996772332174852	Positive	31126.53073	synthetic_RBS_14	UUGGCAACGGGUUAUGGGACGUAUGACA
perimeter error	5.655963608371979	Negative	26985.04	BCD8	UUGGCAACGGCUUAUGGGAGGGAUGCCA
area error	22.02322108639225	Negative	30195.18	BCD6	UUGGCAACGGCUUAUCGGAGGGATGACA
smoothness error	22.103359834529517	Positive	39353.4991	synthetic_RBS_4	UUGGCAACGGCUUAUGGGAGGUAUGUCA
compactness error	19.125114259736407	Negative	4723.721793	synthetic_RBS_43	GGGCCCAAGUUCACUUAAAAAGGAGAUCAAC AAUGAAAGCAAUUUUCGUACUGAAACAUCUUAAU CAUGCUGCGGAGGGUUUCUA
concavity error	27.345908734968848	Negative	5809.43	BCD20	UUGGCAACGGGUUAUCGGAGGGAUGACA
concave points error	44.075566057707576	Negative	1279.278805	synthetic_RBS_62	UUGGCAACGGCUUAUGGGAGGGAUGACA
symmetry error	12.08302765909511	Positive	39353.4991	synthetic_RBS_4	UUGGCAACGGGUUAUCGGAGGGAUGUCA

Table 1. Corresponding RBS to each weight class, after the training process of the PERceptron application for the breast cancer dataset.

In terms of the RBS Library used, it should be mentioned that it has been modified from the initial Library offered by the dry lab during the design process of our project, due to the fact that we received some laboratory data for 11 of the Sender’s RBS, as they are presented in the table below:

RBS	Value Wet	Value Dry	a_luxi_fitted	Proportional Strength
BCD2	43704	43680	128	287840
S11	34943	34125	100	224875
BCD14	23202	23205	68	152915
S13	19903	20475	60	134925
S10	17106	17062	50	112437
BCD12	16981	16209	47.5	106815
S4	5926	5972	17.5	39353
BCD8	4050	4095e	12	26985
S12	2887	2900	8.5	19114
S5	597	597	1.75	3935

Table 2. Fitted RBS sequences (as presented in the Model section) and their proportional strengths, based on the strongest proportional value of 297840 offered by RBS BCD2

So, since our RBS Library should be modified due to new experimental data, our team thought that a great idea would be the capability to modify the imported RBS, since future iGEM Teams might come with novel characterization data or even new RBS parts that they would like to use. As a result, the PERspective App offers advanced users with the capability to modify our own RBS Library via “delete” and “insert” options on the left-hand side of the application.

Figure 4. RBS Library displayed before and after insertion of novel RBS sequences. We can view on the top of the list the imported “rbs_sample_01” from the downloaded sample dataset, with TIR value of 300,000.

Figure 5. Delete attempts when an RBS sequence is found or not found (based on it’s input name).

Figure 6. A significant drop in training accuracy is observed for large datasets, such as the sklearn breast cancer one.

In order to run the application, visit our repository on gitlab.igem.org, here. You can find instructions on the requirements and how to run the application in the README.md file, included in the repository.

Conclusions

Regarding its value, we believe that the PERspectives Application has the potential to become a solid bioinformatics tool for future iGEM Teams due to its versatility as a simulation and design tool. As future goals, our team aims to create a more dynamic UI, probably moving from the Streamlit python implementation to a full-stack one, which will include a relational database in SQL, an API such as php and a website interface using HTML,CSS and JavaScript languages, for a more customized setup.

In terms of the RBS Library, it would be of high importance the update of the initial RBS Library with translation rate sequences that have been validated from wet lab experiments on the sender’s constructs as well as minor modifications on the Perceptron algorithm in order to avoid bad parameter tuning, especially in the case of large datasets with more than 10 classes. That was especially displayed on the breast cancer dataset where a small change in the learning rate from 0.03 to 0.1 resulted in a major drop in training accuracy.