“`html
Neural networks are computational frameworks crafted to replicate both the organization and operation of the human brain. Caltech scientists have been innovating a neural network constructed from strands of DNA rather than electronic components, executing computations through chemical reactions instead of digital signals.
A crucial attribute of any neural network is its capability to acquire knowledge by absorbing information and retaining it for future choices. Currently, researchers in the laboratory of Lulu Qian, a professor of bioengineering, have developed a DNA-infused neural network that possesses the ability to learn. This research signifies a preliminary step toward illustrating more intricate learning behaviors in chemical systems.
A paper detailing the study is published in the journal Nature on September 3. Kevin Cherry (PhD ’24) is the lead author of the research.
The ability to learn exists on various levels: Our minds reconfigure themselves to assimilate new insights, our immune systems encode information chemically about past encounters with pathogens for future reference, and even single-celled organisms acquire basic information about chemical gradients and utilize it to move towards nourishment. Learning is an essential aspect of intelligence, whether organic or artificial; for instance, “intelligent” devices can adapt to your preferences and provide tailored recommendations.
“Our aim was to create a molecular system from the ground up that could absorb examples, identify underlying patterns, and then respond to new information it had never encountered before,” Qian states. “Imagine a future artificial cell, learning from a biological cell as its instructor. It observes how the instructor reacts to various molecular signals, retains those experiences, and—through multiple lessons—discovers how to independently respond to similar but not identical signals.”
In 2018, Cherry and Qian developed a DNA-driven neural network capable of recognizing handwritten digits encoded in DNA as chemical patterns. Since it can be challenging, even for humans, to decipher messy handwriting, identifying handwritten numbers is a common trial for instilling intelligence into classical electronic artificial neural networks. These networks must account for variations in writing styles, then juxtapose an unfamiliar pattern with their so-called memories and ascertain what digit is represented by the image. In Cherry and Qian’s framework, instead of digital pixels forming a number, each molecular “image” comprised 20 distinct DNA strands designated to symbolize an individual pixel in a 10-by-10 array. While constructing the DNA system, a classical computer was employed to ascertain the quantity of each molecular component necessary to encapsulate the memories.
The recent study builds upon that research to create a system that can “cultivate” its own memories, encoded in chemical signals referred to as molecular wires. These wires can be chemically activated to store information. When the system encounters a molecular representation of a handwritten digit, it activates a set of wires that establish a connection between a number and its physical identifying features. Over time, the system accumulates a tangible record of what it has absorbed, stored in the concentrations of specific DNA molecules. The idea parallels how human brains learn. (There is a saying in neuroscience that “cells that fire together, wire together.” In this context, the wiring is molecular, and the memories reside within the chemistry itself.)
Each neural network can execute its computations to discern digits within a minuscule droplet containing billions of DNA strands of over a thousand varieties. Every type of strand was engineered from scratch to react exclusively with specific intended partners under certain conditions. When the series of chemical reactions concludes, the system yields an output. For instance, if a molecular image is identified as a handwritten “0,” a fluorescent signal is emitted that corresponds to the output, such as red for 0 and blue for 1.
“Our quest for a DNA neural network that learns spanned seven years—and the route was far from straightforward,” Cherry remarks. “In a complex molecular framework, rectifying one issue resembled patching a leak in a dam only to have another leak spring up elsewhere. Instead of addressing challenges incrementally, we had to step back and observe the entire landscape, then devise solutions that tackled all the challenges simultaneously. It was a daring decision, as it required starting from square one. With a fresh, comprehensive design, we ultimately realized what we had been pursuing: a molecular system that can learn. Reflecting on it, the science imparted a broader lesson: that the most formidable problems necessitate both a panoramic perspective and the bravery to begin anew when the stakes are at their peak.”
The research establishes a foundation for eventually developing “smart” medications that can adapt instantaneously to pathogenic threats or “smart” materials that can learn and adjust to external conditions (like a bandage that learns from your skin’s signals and responds to enhance faster wound healing).
The paper is titled “Supervised learning in DNA neural networks.” Cherry and Qian are the authors of the study. Funding was provided by Schmidt Sciences and the National Science Foundation.
“`