Advancements in bio-engineering have enabled the creation of in-vitro biological neural networks, offering an exciting avenue for a new kind of computational platform. A computing stack powered by living neurons could unlock self-organizing and dynamically rewiring systems with extreme connectivity and parallel processing power, all while running on sugar with unprecedented energy efficiency. Despite their potential, computing applications of these biological systems remain a nascent and limited technology that presents a challenging and radical departure from the precise, digital von Neumann architectures that dominate today’s computing landscape. Here, we outline a framework that leverages in-silico simulation to establish an engineering testbed with the ultimate goal of learning to harness neural in-vitro systems for computational purposes. We describe an optimization approach to uncover reproducible neural activity present in a system that can be leveraged to carry out basic information processing tasks. We demonstrate the feasibility of this approach by optimizing a simulated neural system to perform digit classification, offering a proof-of-concept for a potential pathway to leveraging neural computation in vitro.