Research

Inspired by these compact molecular rules underlying biological connectivity, I next asked whether genetic encoding could serve as a general principle for network construction. In Scientific Reports, I defined a generative network model that uses simple heuristics to build scale-free and feed-forward topologies out of biclique motifs. Building on this idea, I proposed the Genetic neuroEvolution Model (GEM, Nature Communications), in which artificial neural network weights are not optimized directly, but instead emerge from learned compatibility rules between neurons, inspired by the developmental wiring models from my earlier Neuron and PNAS papers. Rather than improving task fitness by updating individual weights, GEM updates the neurons’ wiring rules, mirroring evolutionary selection on developmental programs. We find that GEM (1) provides sufficient representational power for strong performance on machine learning benchmarks while also compressing parameter counts, and (2) acts as an effective regularizer in meta-learning settings, favoring simpler and more stable circuit solutions. Together, these two papers crystallized a theme that has remained central to my research: development can be understood not just as a biological process, but as a source of structured priors that support efficient computation.

I next sought to test my computational ideas experimentally, by asking whether complex behavior can emerge from developmental wiring alone. In Nature Communications, I showed that the zebrafish optomotor response can mature even when neural activity is pharmacologically silenced throughout development. After washout of the anesthetic, animals still performed visually guided behavior with high accuracy, and brain-wide imaging revealed that the underlying neural populations came online already tuned, without requiring activity-dependent refinement. Thus, complex sensory-guided behavior can be assembled by activity-independent developmental mechanisms. I later formalized this view in Nature Reviews Neuroscience, proposing that circuit formation reflects three interacting systems: assembly, in which genetically guided programs establish core circuitry; updating, where salient experiences rapidly modify synapses; and tuning, in which ongoing plasticity stabilizes and fine-tunes networks. This summarizes the implications of my experimental findings: rather than treating learning and plasticity as the primary architects of neural function, many core competencies may first emerge through innate developmental programs, upon which later experience acts to update and refine.

As neuronal-level maps became available, I returned to the question of how brain-wide connectivity is organized. Collating eight connectomes across five species, I found that canonical network-science models do not capture brain architecture: neuronal degree and connection strength were consistently better approximated by lognormals than by random or scale-free distributions. To address this inconsistency, we reframed the brain as a physical network of spatially extended neurons, thereby providing a mechanistic model: stochastic multiplicative growth of arbor length provides a common developmental basis for the observed distributions of neuron length and degree. This framework links morphology, connectivity, and network organization in a unified quantitative theory, generating empirically falsifiable relationships among neuronal properties.

As a FutureHouse AI-for-Science Fellow, I am working on the design and evaluation of an AI Scientist (Kosmos) that integrates hypothesis-guided literature search and data analysis into iterative scientific workflows. This gave me unusual firsthand experience with a new methodological frontier: building AI systems that do not merely summarize knowledge, but produce mechanistic insight. In evaluating Kosmos on connectomics, I found that it could not only reproduce the central empirical findings of my preprint, but also recovered the underlying interpretation. I now see agentic systems as essential partners in extracting organizing principles from large-scale brain datasets, and am continuing to pioneer AI-assisted theory-building as a core methodology for connectomic discovery.