LFP2Gene: Recovering Local Molecular Composition from Local Field Potentials

Local field potentials (LFPs) are generated by the synaptic and cellular machinery of the tissue immediately around an electrode. We hypothesize that self-supervised LFP representations therefore retain signatures of their local molecular environment. Bridging International Brain Laboratory Neuropixels channels to reference MERFISH atlases via the Allen Common Coordinate Framework, we ask how much local gene expression a frozen encoder recovers on 34 held-out subjects. A good deal, and more than any alternative we tested under the same frozen-readout protocol. Because both LFP and gene expression vary with anatomical location, we read this as recovery of the local molecular environment rather than as evidence of molecular information independent of anatomy. Recovery also falls steadily as the target neighbourhood shrinks from 800 to $100\\,\mu$m, so the signal is regional in character rather than cellular. To ask how much structure is genuinely *shared*, rather than only how well it can be decoded, we compare two estimators that differ only in whitening: PLS and CCA agree on five shared dimensions and diverge beyond them, while a permutation null would accept far more. What LFP representations carry about their molecular surroundings is therefore real, low-dimensional, and organised at the scale of brain regions.