Announcer
The following program features simulated voices generated for educational and philosophical exploration.
Adam Ramirez
Good evening. I'm Adam Ramirez.
Jennifer Brooks
And I'm Jennifer Brooks. Welcome to Simulectics Radio.
Adam Ramirez
Tonight we're examining connectomics—the systematic mapping of neural connections at synaptic resolution. The Human Connectome Project mapped large-scale fiber tracts using diffusion imaging. C. elegans has a complete electron microscopy connectome of all three hundred neurons. Drosophila now has a complete adult brain connectome with hundreds of thousands of neurons. The question is what we actually gain from these maps. Does knowing every synaptic connection tell us how the circuit computes, or do we need additional information about synapse strengths, neuron types, and dynamic activity patterns?
Jennifer Brooks
There's an analogy to genomics. Sequencing the human genome was a monumental achievement, but the sequence alone didn't explain gene function. You need expression data, regulatory networks, protein interactions—layers of functional information beyond the static blueprint. Similarly, a connectome provides anatomical structure but doesn't specify how signals propagate, how plasticity modifies connections, or how neuromodulators alter effective connectivity. The map is not the territory, and the wiring diagram is not the computation.
Adam Ramirez
To explore these questions, we're joined by Dr. Sebastian Seung, a neuroscientist at Princeton University whose lab has been instrumental in developing connectomics methods and analyzing C. elegans and retinal connectomes. His work bridges the technical challenges of reconstructing circuits from electron microscopy with the conceptual challenges of interpreting what those circuits tell us. Dr. Seung, welcome.
Dr. Sebastian Seung
Thank you for having me.
Adam Ramirez
Let's start with the technical side. Reconstructing a connectome from electron microscopy requires imaging tissue at nanometer resolution, then tracing neurites through thousands of serial sections to identify synaptic connections. This generates petabytes of data. The original C. elegans connectome took over a decade of manual reconstruction. How have automated methods changed the feasibility of mapping larger circuits?
Dr. Sebastian Seung
Automation has been transformative. We now use convolutional neural networks to segment neurons in electron microscopy images—identifying cell boundaries, tracing processes, detecting synapses. This reduces the manual effort by orders of magnitude. Humans still validate the reconstruction and correct errors, but instead of tracing every neurite by hand, they're proofreading automated reconstructions. The Drosophila brain connectome—about a hundred and fifty thousand neurons—would have been infeasible without these machine learning methods. We're now working on mouse cortical circuits, which are another order of magnitude larger.
Jennifer Brooks
Even with automation, you're still generating a static snapshot of connectivity. Synapses are dynamic—they strengthen with potentiation, weaken with depression, form and retract over hours to days. How much does a single connectome miss by capturing structure at one point in time?
Dr. Sebastian Seung
That's a real limitation. The connectome tells you which neurons are connected, but not the current strength of those connections. For some questions, that matters enormously. If you're trying to understand learning and memory, synaptic weights are critical. But for other questions, the anatomical connectivity is sufficient. If you want to know which sensory neurons connect to which motor neurons to produce a reflex, the wiring diagram gives you that. The connectome provides boundary conditions—it tells you what connections are physically possible. Functional measurements tell you which connections are currently active and strong.
Adam Ramirez
Let's talk about C. elegans. It's had a complete connectome since the 1980s—all chemical synapses and gap junctions mapped. That's forty years of having the wiring diagram. Has that connectome enabled us to predict the worm's behavior from first principles?
Dr. Sebastian Seung
Not entirely. We can explain some circuits—the tap withdrawal reflex, thermotaxis, certain locomotion patterns. But predicting complex behavior remains difficult. Part of the problem is that we don't know all the parameters. Neurons have different ion channel densities, different intrinsic excitabilities, different neuromodulator sensitivities. Synapses have different strengths and dynamics. The connectome gives you the graph structure, but not the edge weights or node properties. It's like having the circuit diagram of a computer without knowing the values of the resistors and capacitors.
Jennifer Brooks
That suggests we need to combine connectomics with other methods—physiology to measure synaptic strengths, transcriptomics to identify neuron types, imaging to track activity. How do we integrate these different data modalities into a unified model of a circuit?
Dr. Sebastian Seung
That's the current frontier. You can use calcium imaging to measure activity patterns, then match those patterns to specific neurons in the connectome. You can use RNA sequencing to classify neurons into types based on gene expression, then map those types onto the anatomical structure. You can use optogenetics to perturb specific neurons and see how activity propagates through the circuit. The challenge is doing all of this in the same preparation, so you can register the datasets to each other. Some labs are now attempting this—imaging activity, then fixing the tissue for electron microscopy to get the connectome, then sequencing the RNA to get cell types. It's technically demanding, but it provides the multilayer view you need.
Adam Ramirez
Even with all that data, you still need a model that explains how structure produces function. What kinds of models do people build from connectomes?
Dr. Sebastian Seung
The simplest models are simulations—take the connectome as a graph, assign synaptic weights based on functional measurements or educated guesses, simulate the dynamics using differential equations for membrane potentials and synaptic currents, then see if the model reproduces observed behavior. More sophisticated approaches use machine learning to infer parameters that best fit the data. You constrain the network topology to match the connectome, but learn the weights to reproduce activity patterns or behavioral outputs. This is computationally intensive, but it allows you to test whether the anatomical structure alone is sufficient to explain function.
Jennifer Brooks
There's a conceptual tension here. On one hand, we're saying the connectome is insufficient without functional data. On the other hand, we're using the connectome to constrain models. How much does the anatomical structure actually constrain the space of possible functions?
Dr. Sebastian Seung
That's a deep question. In principle, a highly recurrent network with plastic synapses could implement many different functions depending on the weights. The anatomical structure might be permissive rather than prescriptive. But in practice, evolution likely shaped connectivity to facilitate certain computations. Feed-forward pathways for sensory processing, recurrent loops for memory, lateral inhibition for contrast enhancement—these motifs appear repeatedly across species and circuits. The anatomy may not uniquely determine function, but it biases the network towards certain solutions.
Adam Ramirez
Let's talk about scaling. C. elegans has three hundred neurons. Drosophila has a hundred and fifty thousand. A mouse brain has a hundred million. The human brain has eighty-six billion. At what scale does connectomics become impractical?
Dr. Sebastian Seung
It depends on what resolution you need. Full synaptic resolution connectomics—every synapse between every neuron—is currently practical for insects, feasible for small mammalian circuits like a cubic millimeter of mouse cortex, and far out of reach for whole mammalian brains. But you might not need complete coverage. You could focus on specific circuits—the hippocampus, a cortical column, the cerebellum. Or you could use statistical sampling—reconstruct enough of the network to estimate connectivity statistics, then extrapolate to the full structure. For whole-brain human connectomics at synaptic resolution, we're talking about exabyte-scale datasets and decades of effort even with automation.
Jennifer Brooks
That raises the question of whether we should pursue complete connectomes or focus on functional circuits. If the goal is understanding computation, maybe we should prioritize circuits where we have strong behavioral assays and functional data, even if that means leaving the rest of the brain unmapped.
Dr. Sebastian Seung
I think both approaches are valuable. Complete connectomes are reference datasets—they enable discovery of unexpected connections and circuit motifs that you might miss if you only focus on known pathways. But targeted connectomics of well-studied circuits is more immediately useful for understanding function. The retina is a good example. We have decades of physiological recordings characterizing how different cell types respond to visual stimuli. The connectome tells us how those cells are wired together, which allows us to build mechanistic models of visual processing. That synergy between structure and function is where we make real progress.
Adam Ramirez
How much does circuit structure vary between individuals? If we map the connectome of one Drosophila, does that tell us about other flies, or is each connectome unique?
Dr. Sebastian Seung
There's variability at multiple levels. The coarse connectivity—which neuron types connect to which—is highly stereotyped across individuals. But the fine details—exactly which neurons form synapses, the number of synapses per connection—show considerable variability. For some circuits, this variability is noise—developmental stochasticity that doesn't matter for function because the circuit is robust. For other circuits, particularly those involved in learning and memory, the variability is meaningful—it encodes individual experience. Distinguishing meaningful variability from noise requires comparing multiple connectomes and correlating differences with behavioral differences.
Jennifer Brooks
Let's return to the comparison with genomics. The Human Genome Project cost billions of dollars and took over a decade. Now sequencing costs a few hundred dollars. Will connectomics follow a similar trajectory, or are there fundamental barriers that prevent costs from dropping dramatically?
Dr. Sebastian Seung
Some aspects will improve with technology. Imaging is getting faster with better electron microscopes and automated systems. Reconstruction is getting better with improved machine learning. But there are hard limits. You need nanometer resolution to see synapses, which means physically sectioning the tissue and imaging every section. That's fundamentally slower than extracting DNA and sequencing it. Storage and computation for petabyte datasets are expensive. I expect costs will drop by orders of magnitude over the next decade, but probably not to the point where connectomics becomes as cheap and routine as genomics.
Adam Ramirez
What about alternative approaches? Can we infer connectivity from functional data without direct anatomical reconstruction?
Dr. Sebastian Seung
You can infer effective connectivity from correlation patterns in neural activity. If two neurons consistently fire together with a specific temporal lag, that suggests a connection. But correlation doesn't prove causation—the correlation could be driven by a common input rather than a direct connection. You can use causal inference methods to improve the estimates, but they rely on assumptions about network dynamics. Anatomical connectomics provides ground truth that functional methods can't achieve. Ideally, you'd use both—functional methods to identify candidate connections quickly and cheaply, then validate with anatomy where it matters.
Jennifer Brooks
Looking forward, what's the most important scientific question that connectomics could answer in the next decade?
Dr. Sebastian Seung
How circuit structure changes with learning. We have snapshots of circuits before and after training, but we don't have continuous monitoring of how specific synapses strengthen or weaken during learning. New methods that combine in vivo imaging with post hoc electron microscopy could track individual synapses over time. That would tell us whether learning involves forming new connections, strengthening existing ones, or both. It would also tell us whether plasticity is as local as we think, or whether there are coordinated changes across distributed circuits. That's fundamental to understanding how memories are stored and retrieved.
Adam Ramirez
Dr. Seung, thank you for the detailed picture of where connectomics stands and where it's headed.
Dr. Sebastian Seung
It was a pleasure.
Adam Ramirez
That's our program. Until tomorrow, stay critical.
Jennifer Brooks
And keep questioning. Good night.