Chris Mason

Implementing neural networks... in DNA 🧬 Neural networks are an example of a bio-inspired algorithm. We used neuroscience as an intuition pump for making learning systems. But what if we took our C.S. innovation in neural nets, and... implemented that using biology? How would we do it? A preprint led by Ryan Lee from the lab of @leochou0814 attempted to answer this question. They tried to unify three advances to make better modular DNA-based units for implementing "biomolecular neurons:" 1. enzymatic synthesis to encode a greater number of input-output connections on a 76 single DNA strand 2. solid-phase immobilization to spatially segregate DNA computing units into 77 network layers 3. universal addressing to enable the assembly of different circuits from the 78 same DNA components Using this framework, they showed a proof-of-concept for more efficiently encoding neural networks into DNA constructs. You might be asking yourself.... Why? Honestly, I'm not sure! The authors say this may enable "portable computing power for applications such as portable diagnostics, compact data storage, and autonomous decision making for lab-on-a-chips." At this stage, it's a really interesting basic demonstration of a concept. Who knows how we might figure out how to use this type of functionality if it becomes even easier and cheaper to do? Could synthetic cells some day operate on a combination of logic (explicit regulatory circuits) and fuzzier predictive models (like this DNA-encoded neural nets)? This would be similar to how computer programming has evolved. It's another example of how rich the intersection between biology and computation is. Link to full study: doi.org/10.1101/2023.07.20.5…