Foundation model of neural activity predicts response to new stimulus types

Nature 640 (2025), doi 10.1038/s41586-025-08829-y; open access, PMC11981942. MICrONS Consortium. Read by the main session for research direction R3 ("models of one brain"), 4 October 2026. Provenance: papers/amadeus/wang2025_foundation_model_visual_cortex.provenance.json.

What was read

• Read in full: the Europe PMC full-text XML, converted to text: abstract, main text, every figure and Extended Data caption (figures are images, so only captions were read), Methods (data collection, the four modules, training, testing, tuning analyses, anatomical predictions), and the availability, contribution and competing-interest statements.
• Not read:
  • the reference list, absent from the XML;
  • the Reporting Summary and Peer Review File;
  • the companion MICrONS papers it cites for connectivity and dendrite predictions;
  • the correction of 8 April 2026 (doi 10.1038/s41586-026-10457-z), whose content is unknown here.

What they did

• Data. Two-photon calcium imaging of about 135,000 excitatory neurons across visual areas (V1, LM, RL, AL, AM, PM) and layers 2/3 to 5 in 14 awake, head-fixed mice on a treadmill, watching natural videos. Eye position, pupil size and running speed were recorded.
• Model. Four modules trained end to end to predict each neuron's activity over time (Poisson loss):
  • Perspective: ray tracing from eye tracking, to give what the retina sees.
  • Modulation: an LSTM over running and pupil, giving behavioural and arousal state.
  • Core: 3D convolutions plus a convolutional LSTM. It holds most of the capacity.
  • Readout: per neuron, a receptive-field position and a 512-dimensional feature-weight vector.
• Foundation paradigm. One shared core was trained on more than 900 minutes of recordings from 8 mice (the "foundation cohort", about 66,000 neurons), then frozen. For each new mouse only the perspective, modulation and readout were fitted.

Main results (verified)

• Single-session models beat the previous best dynamic model by 25–46% in normalised correlation on held-out natural videos. Higher visual areas were predicted about as well as V1. Most of the gain comes from more data; module lesions cost 0.9–9.6%.
• Data efficiency for a new individual:
  • Individually trained models needed more than 60 minutes of a new mouse's data to exceed a median normalised correlation of 0.65; foundation models needed under 30.
  • The gain held on stimulus types never used in training (static images, drifting Gabors, flashing dots, pink noise, random-dot motion). For Gabors, the foundation models passed 0.55 with 16 minutes of natural-video data; individual models needed more than an hour.
  • In the MICrONS mouse, with about 42 minutes per scan, foundation models scored 0.58–0.76 against 0.48–0.65 for individual models.
• In silico tuning matches in vivo. For strongly orientation-tuned neurons, the median error in preferred orientation was 4° (OSI above 0.5, 11% of neurons) and 7° (OSI above 0.3, 43%). Preferred location for strongly spatially tuned neurons: about 2° of visual angle.
• A stable "functional barcode" per neuron. The readout weights of the same neuron recorded in different sessions were more similar than those of nearby different neurons: for 919 of 1,013 re-recorded neurons, fewer than 5% of nearby neurons were closer.
• Function predicts anatomy.
  • The barcodes predict the visual area with 68% balanced accuracy (chance 25%).
  • They predict 11 morphologically defined excitatory cell types with 32% (chance 9%), beyond what imaging depth explains (likelihood-ratio p below 1e-9).
  • Companion papers, not read, report predictions of synaptic connectivity and dendritic morphology.

Limits

• One species (mouse), one system (visual cortex), passive viewing with head fixation. The model is driven by stimuli and behaviour: it predicts responses to inputs and does not model internal, self-generated activity.
• Calcium imaging is slow (about 6–7 Hz) and samples excitatory neurons only.
• The individual part of each model is small: a per-neuron linear readout plus small eye and behaviour networks. The core, shared across animals, does the computation.
• No reference list or correction text was read. The figures were not seen.

What it means for Amadeus (inference)

• The closest existing "digital twin of an individual", and its architecture is a shared base plus individual parameters. This mirrors the person slot. The key empirical lesson: when a strong shared core exists, the individual-specific part is learned from much less data (under 30 minutes instead of over 60), and it generalises to situations never trained on.
• Tension with "the base supplies nothing". In this model the shared core supplies the computation, and the individual is "how their neurons read out a shared representation". If the person slot must also carry the person's ways of thinking, not only a readout of shared features, the data needed per person rises sharply. A defensible split: the base holds species-general computation, as the core does here, and the slot holds everything that differs between people. Measuring how much differs is itself an experiment.
• Function recovers structure, not the other way round. Functional barcodes predicted cell type and area. The connectomics reading found that wiring predicts measured effects poorly (Randi 2023). Together they support P3: measure function; structural facts can follow from it.
• Individual function is stable and measurable. Same-neuron barcodes were stable across sessions. For a month of recording, an individual's functional signature should be identifiable. At cellular resolution in mice, though; human non-invasive recording is far coarser.
• Gap for the month-of-thinking idea. This model explains responses to stimuli. Spontaneous thought has no stimulus, so it needs a model of internal dynamics (see LFADS and BrainLM in this batch), anchored by labels.