Methodology
A comprehensive technical breakdown of how the MaleCNS fruit-fly connectome was converted into a recurrent reservoir computing system for time-series forecasting.
01 / The Connectome (MaleCNS v1.0)
The connectome used in FlyCast originates from the open-access MaleCNS v1.0 release, an imaging and reconstruction effort by HHMI Janelia FlyEM, the University of Cambridge, the MRC Laboratory of Molecular Biology, and Google Research.
The published dataset contains 166,691 identified neurons across the adult male fruit fly central nervous system (brain, optic lobes, and ventral nerve cord) and 11,691 annotated cell types. FlyCast treats bodies with non-null superclass annotations as computational nodes, resulting in a runtime graph of approximately 166,483 connected nodes and 25.58 million directed edges.
02 / Turning Wiring into a Sparse Reservoir
The 1.1 GB raw graph is transformed once into a Compressed Sparse Row (CSR) matrix. To reduce the extreme heavy-tailed distribution of raw synaptic contact counts while preserving connection hierarchies, synaptic strengths are scaled via:
The sparse CSR matrix is strictly constructed with W[post, pre] orientation so that matrix-vector multiplication next_input = W @ state faithfully reflects activity flowing from presynaptic to postsynaptic neurons.
To prevent runaway excitation or numerical divergence, incoming synaptic weights are normalized per postsynaptic neuron such that the L1 norm of incoming weights equals 1.0 (row normalization).
03 / Injecting User Data into Biological Neurons
Arbitrary financial, seasonal, or physical time-series signals do not correspond to literal fruit fly sensory stimuli. However, biological inspiration dictates that external information enters the nervous system via dedicated sensory populations.
FlyCast identifies neurons whose anatomical superclass marks them as sensory. A deterministic subset of512 sensory neurons (seed 42) is selected. Each input feature is projected across a partition of these sensory neurons with fixed projection weights sampled from:
04 / Leaky Recurrent Rate Dynamics
FlyCast executes a classic Echo State Network (ESN) leaky rate update for each discrete observation:
Default parameters: leak = 0.25, recurrent_gain = 0.95, and input_gain = 0.50. All states are stored as 32-bit floating point vectors.
05 / Direct Multi-Horizon Ridge Readout
The recurrent connectome weights remain completely frozen. Rather than storing all ~166K neuron states at each timestep, 4,096 virtual readout electrodes are deterministically sampled across the CNS.
For a forecast horizon H, FlyCast constructs direct multi-output targets:
A multi-output linear Ridge model is trained on the sampled reservoir states. Chronological splits (70% Train, 15% Validation, 15% Test) are strictly enforced without shuffling. Hyperparameter selection across candidate alphas (1e-4 to 100) is determined exclusively on validation RMSE. The final model is refitted on Train + Validation and evaluated on unseen Test rows.
06 / Baseline Benchmarks
Every experiment is evaluated alongside standard non-connectome baselines using identical chronological splits:
- Persistence: Projects the last observed value
y[t]forward across all future horizons. - Autoregressive Ridge: Uses lagged values of the target series (candidate lags: 10, 20, 40) with cross-validated Ridge regularization.
07 / Scientific Limitations
Electron microscopy connectomics provides static anatomical structure. It does not record the animal's dynamic firing patterns, electrophysiological conductance, neurotransmitter receptor densities, or neuromodulatory state.
Therefore, FlyCast represents an engineered computational machine operating over biological graph topology. It is not a digital resurrection or cognitive simulation of a living fruit fly.
08 / Reproducibility & Open Source
Every run records the dataset SHA-256 hash, connectome manifest hash, random seed (42), split boundaries, and exact hyperparameter selections. Any experiment can be audited and reproduced identically.