Introducing perceptein, a protein-based artificial neural network in living cells

Westlake University in China and the California Institute of Technology have designed a protein-based system inside living cells that can process multiple signals and make decisions based on them.

The researchers have also introduced a unique term, “perceptein,” as a combination of protein and perceptron. Perceptron is a foundational artificial neural network concept, effectively solving binary classification problems by mapping input features to an output decision.

By merging concepts from neural network theory with protein engineering, “perceptein” represents a biological system capable of performing classification computations at the protein level, similar to a basic artificial neural network. This “perceptein” circuit can classify different signals and respond accordingly, such as deciding to stay alive or undergo programmed cell death.

Cells naturally process multiple classification cues, such as stress and developmental signals, to initiate cell functions with distinct outcomes. Immune cells respond to threats based on the signals they detect. The p53 signaling pathway determines whether to repair damage or self-destruct to prevent cancer.

Scientists have struggled to create artificial systems that can replicate this decision-making process inside cells. Most existing attempts rely on DNA or RNA, which can be slow and less direct. Instead of DNA-based systems, the researchers built their decision-making circuit with proteins, de novo protein heterodimers and engineered proteases.

By creating protein pairs that bind together in specific ways, the proteins arrange into the perceptein network, where some proteins activate themselves and inhibit others. This ensures that when multiple signals are present, only the strongest one triggers a reaction, ignoring weaker signals.

In the study, “A synthetic protein-level neural network in mammalian cells,” published in Science, researchers showed that perceptein circuits could distinguish signal inputs with tunable decision boundaries, offering the possibility of controlling complex cellular responses without transcriptional regulation.

The team assembled six perceptein protein components and two input proteins necessary for a complete two-input, two-output circuit. They selected two well-known proteases, split tobacco etch virus protease and tobacco vein mottling virus protease, and fused them in a way that controls for protease cleavage and degradation.

To test the activation of the perceptein circuit, researchers engineered a stable human embryonic kidney reporter cell line. This cell line contained a construct that simultaneously expressed two fluorescent proteins: Citrine and mCherry.

Each fluorescent protein was tagged with a cleavage-activated N-degron (degradation signal) specific to one of the two input proteases in the perceptein circuit. When a corresponding protease was active, it would cleave the degron, reducing fluorescence. This setup allowed the researchers to visually and quantitatively assess activity based on fluorescence levels. The team confirmed that each protease variant specifically reduced fluorescence only from its target reporter.

Further validation steps demonstrated that input proteins correctly reconstituted their target proteases. By altering perceptein component levels, they could effectively fine-tune the decision outcomes, and performance remained strong even when input timing varied or noise was introduced.

To showcase practical application, the researchers connected the perceptein circuit’s output to a caspase-3 apoptosis pathway. This linkage allowed the circuit to trigger cell death based on specific input conditions, transforming fluorescence-based outputs into life-or-death decisions for the cells.

The study demonstrates the feasibility of constructing artificial neural network-inspired circuits in mammalian cells using synthetic proteins to perform complex signal classifications. These circuits have potential applications in programmable therapies, where cells could respond to disease-specific signals with tailored outputs, such as selective apoptosis or other cellular responses.

There are also obvious implications for constructing complex computational systems out of interacting proteins as a form of biology-based artificial intelligence, though such considerations are outside the scope of the current research effort.

© 2024 Science X Network