Molecular and DNA Artificial Neural Networks via Fractional Coding

Area of Science:

• Biomolecular Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Prior molecular artificial neural networks (ANNs) using fractional coding were limited to binary inputs and scaled weighted sums.
• Existing molecular perceptrons could not handle arbitrary positive/negative weights or compute the true weighted sum, restricting their application scope.

Purpose of the Study:

• To develop molecular perceptrons capable of handling arbitrary weights and computing the sigmoid of weighted sums.
• To construct a molecular ANN with a hidden layer using fractional coding.
• To present molecular implementations of activation functions like Rectified Linear Unit (ReLU) and softmax.

Main Methods:

• Introduction of a novel molecular divider to compute sigmoid(ax).
• Design of a molecular ANN architecture with one hidden layer based on fractional coding.
• Mapping a trained ANN classifier for seizure prediction from electroencephalogram data to molecular reactions and DNA.

Main Results:

• Demonstration of molecular perceptrons that accommodate arbitrary weights and compute sigmoid of weighted sums, suitable for regression and multi-layer ANNs.
• Successful construction of a molecular ANN with a hidden layer.
• Presentation of performance metrics for a molecularly implemented seizure prediction classifier.
• Development of molecular activation functions for ReLU and softmax.

Conclusions:

• The developed molecular perceptrons and ANNs overcome previous limitations, enabling more complex computations and broader applications in machine learning.
• This work advances DNA-based computing for sophisticated AI tasks, including real-world applications like seizure prediction.
• Molecular implementations of activation functions expand the toolkit for building complex biomolecular computing systems.