From Worm to Human

Abstract

Machine learning models are rapidly approaching or surpassing human performance on many metrics. In comparison, neuroscience is progressing at a slow pace. To reach the research velocity possible in software paradigms, we highlight a potential path to high-quality emulations of the brains of key model organisms such as the roundworm Caenorhabditis elegans, the zebrafish Danio rerio, and the mouse Mus musculus.

Notably, multiple required technologies are rapidly improving. Firstly, high-resolution imaging of neuronal structure and connectivity in light microscopy or electron microscopy. Connectomics has progressed from mapping the brain of a 302-neuron nematode C. elegans to a complete adult fruit fly brain reconstruction containing approximately 140,000 neurons. Required reconstruction and proofreading cost per reconstructed neuron has fallen from roughly $16,500 to approximately $100 for zebrafish larvae.

Secondly, functional imaging has improved drastically, approaching high-resolution imaging of entire brains in the young zebrafish and cortical sections in mice containing up to 1 million neurons.

Thirdly, simulation capabilities have progressed far. Detailed emulated models of neurons and synapses are available, ranging from simple proxies to biologically accurate neurons. Under pessimistic assumptions we estimate that real-time human brain emulations require roughly 6×10²⁰ FLOP/s of compute, 700 GB memory storage per GPU, and 24 GB/s interconnect bandwidth. Mid-2020s AI clusters reach 4×10²⁰ FLOP/s, 180 GB memory per GPU, and 1.8 TB/s interconnect, with large investments into larger clusters ongoing.

Early emulation attempts are already running. Recent progress includes benchmarking-focused emulations of zebrafish brains, an entire mouse cortex, and incomplete simulations as large as 80 billion neurons, beginning to reach human-scale requirements.

Rigorous, detailed, and biologically accurate emulations of multiple model organisms are surprisingly tractable. C. elegans, larval zebrafish, and Drosophila present compelling near-term targets where structural, functional, and molecular datasets are attainable. Progress in these organisms can drive the development of technologies and structure-to-function mapping methods required to scale up brain emulation work to larger organisms, including eventually human-scale emulations.