Dual-feature selectivity enables bidirectional coding in visual cortical neurons

This dataset contains neural recordings and computational analyses
supporting the identification of dual-feature selectivity in visual
cortex. We recorded spiking activity from macaque visual areas V1 (458
neurons) and V4 (394 neurons) while animals viewed naturalistic images, as
well as from mouse visual cortex areas V1 (598 neurons), LM (350 neurons),
and LI (126 neurons). Using functional digital twin models—deep
learning-based predictive models trained on these recordings—we
systematically characterized neuronal selectivity across the full dynamic
range of responses. The dataset includes: (1) 200,000 synthetically
rendered scenes (236×236 pixels, PNG format) used to probe neuronal
responses; (2) optimized most and least exciting inputs (MEIs/LEIs)
generated through gradient-based synthesis for each neuron; (3) indices
identifying the most and least activating natural images (MAIs/LAIs) from
large-scale screening of ImageNet and of the Rendered Data; (4) predicted
neuronal activation profiles across all stimuli; and (5) metadata
including baseline firing rates, and response reliability metrics. These
data reveal that many visual neurons exhibit bidirectional
selectivity—responding strongly to preferred features while being
systematically suppressed by distinct non-preferred features around
elevated baseline firing rates. This coding strategy appears conserved
across species (macaque and mouse) and visual areas (from primary to
higher-order cortex), suggesting a general principle of sensory coding
that balances representational capacity with interpretable single-neuron
responses.