This project develops machine learning approaches that describe statistical systems in biology. By combining analytic results calculated from the exact probabilistic description of the system with machine learning inference, our new methods present exciting opportunities to model previously inaccessible complex dynamics. The resulting Boltzmann machine-like learning algorithms present a new class of modeling techniques based on the powerful in- ference of arti cial neural networks. Further development of this approach will bring the groundbreaking advances from the surge of recent interest in machine learning into the biological modeling eld. The mathematical methods we develop will be used to derive e cient algorithms for multiscale simulation, directly applicable to large scale biological modeling. In particular, the algorithms will be used to study the dynamics of stochastic biochemistry at synapses, with direct relevance to learning and memory formation in the brain. Current studies of these processes are limited by the long timescales involved and the highly spatially organized structures featured. In addition to leveraging the machine learning expertise we are developing, we also employ new electron microscopy datasets to produce 3D reconstructions of neural tissue with unprecedented accuracy. Consequentially, we will be able to study the fundamental mechanisms underlying synaptic plasticity, as well as the biochemical basis of oscillatory behavior in networks of neurons that occurs during sleep. Furthermore, the interactions of these highly stochastic ion channels with electrical in neurons will be explored through groundbreaking hybrid simulation environments. The software that we will develop combines existing popular simulation tools into multiscale approaches, and will be distributed as a powerful tool to the broader biological modeling community. Its usage in further computational experiments can present a key advancement in the development of pharmaceuticals, allowing the direct study of the interactions of biochemistry and whole neuron electrophysiology without making limiting assumptions to sim- plify the simulations. This has promising implications for intervening in age-related learning de cits, as well as in neurological disorders such as Alzheimers. Finally, this proposal will bring together our existing multiscale modeling community, the National Center for Multi-scale Modeling of Biological Systems (MMBioS), with the MSM consortium. The interactions of these organizations and their communities of expert researchers will foster new collaborative work on exciting multiscale problems in biology, including applications of the machine learning frameworks and software we are developing. 1
A wide variety of biological systems can be described statistically, from molecular biochemistry up to the network level activity of neurons. This work develops machine learning approaches to approximate these systems, enabling new simulation methods that bridge di erent levels of description. The resulting computational studies aim to shed light on the basis of learning and computation in the brain, and will enable the development of pharmaceutical targets for learning de cits associated with aging and neurological disorders such as Alzheimers. 1
