HIPPIE — High-dimensional Interpretation for Physiological Patterns in Intercellular Electrophysiology

HIPPIE is a multimodal generative model that learns a shared latent representation of neurons from extracellular electrophysiology recordings. It is trained jointly on waveform shape, inter-spike interval (ISI) distribution, and autocorrelogram (ACG) features across 11 publicly available datasets, with explicit conditioning on recording technology.

The checkpoint released here was pretrained on all 11 datasets using technology conditioning (4-class: neuropixels, silicon probe, juxtacellular, tetrode), making it suitable for transfer to new datasets without retraining the full model.

---

Model Details

Property	Value
Architecture	Conditional VAE (ResNet18-based 1D conv encoder/decoder)
Latent dimension	30 (10 per modality × 3 modalities)
KL weight (β)	1.0
Input modalities	Waveform (50 samples), ISI (100 bins), ACG (100 bins)
Technology conditioning	4-class embedding (dim=5)
Pretraining datasets	11 (see below)
Framework	PyTorch + PyTorch Lightning

Architecture

Each modality is independently encoded by a 1D ResNet18 with channel widths [64, 128, 256, 512]. A fusion encoder concatenates all three modality representations and projects to the shared latent space. The decoder mirrors the encoder structure and is conditioned on both the latent code and the cell-type label (decoder-only conditioning — the encoder is label-agnostic at inference time).

For datasets lacking ACG data (bimodal recordings), zero vectors are injected for the ACG modality.

---

Training Data

Pretrained on 11 labeled electrophysiology datasets spanning mouse, rat, and macaque:

Dataset	Species / Region	Technology	# Cell Types
Hull et al.	Mouse cerebellar cortex	Neuropixels	3
Hausser lab	Mouse cerebellar cortex	Neuropixels	3
Lissberger lab	Macaque cerebellum	Silicon probe	3
CellExplorer (Petersen et al. 2021)	Mouse cortex / hippocampus	Neuropixels	Multi
IBL brain-wide	Mouse whole-brain	Neuropixels	Multi-region
Allen Neuropixels	Mouse cortex	Neuropixels	Multi
DANDI 000041	Rat frontal cortex	64-ch silicon probe	Multi
DANDI 000473	Mouse PFC	Neuropixels	Multi
DANDI 000955	Rat somatosensory cortex	32-ch NeuroNexus	2
A1 (Lakunina et al. 2020)	Mouse auditory cortex	Silicon probe	Multi
Juxtacellular S1	Mouse somatosensory cortex	Juxtacellular micropipette	Multi

Technology Conditioning Vocabulary

ID	Label	Datasets
0	neuropixels	IBL, Allen, Hausser, Hull, CellExplorer, DANDI 000473
1	silicon_probe	DANDI 000041, DANDI 000955, A1, Lissberger
2	juxtacellular	Juxtacellular mouse S1
3	tetrode	Reserved

---

Performance

KNN balanced accuracy (cosine distance, k=1..20 sweep) on held-out test splits:

Dataset	HIPPIE	NEMO	PhysMAP	VAE
Hull (cerebellum)	0.955	0.862	0.813	0.941
Lissberger (cerebellum)	0.796	0.781	0.743	0.704
A1 (auditory cortex)	0.658	—	0.624	0.588
Juxtacellular S1	0.677	—	0.657	0.537

---

How To Use

Fine-tune on a new dataset

import torch
from huggingface_hub import hf_hub_download

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="Jesusgf23/hippie",
    filename="hippie_techcond_v1.ckpt"
)

# Load with the HIPPIE training script
# (requires the hippie_benchmarking repository)
# python scripts/train_multimodal_transductive.py \
#   --dataset YOUR_DATASET \
#   --pretrain-checkpoint PATH_TO_CKPT \
#   --use-tech-conditioning \
#   --config class_decoder_source_bn_aug_reg \
#   --z_dim 30 --beta 1.0

Extract latent embeddings

import torch
import numpy as np

# Load model state dict
ckpt = torch.load(ckpt_path, map_location="cpu")
state_dict = ckpt["state_dict"]

# Inputs (all float32, batch-first):
#   waveform:  (N, 1, 50)   — min-max normalized to [-1, 1]
#   isi:       (N, 1, 100)  — log-transformed ISI distribution
#   acg:       (N, 1, 100)  — autocorrelogram (zeros if unavailable)
#   tech_id:   (N,)  int    — 0=neuropixels, 1=silicon_probe, 2=juxtacellular

# See hippie/multimodal_model.py for the full forward pass.

Waveform preprocessing

import torch
import torch.nn.functional as F

def preprocess_waveform(raw: np.ndarray, wave_len: int = 50) -> torch.Tensor:
    wave = torch.as_tensor(raw, dtype=torch.float32).unsqueeze(0).unsqueeze(0)
    wave = F.interpolate(wave, size=(wave_len,), mode="linear", align_corners=False)
    wave = wave.squeeze(0)
    mn, mx = wave.min(), wave.max()
    if mx > mn:
        wave = (wave - mn) / (mx - mn)
    return wave * 2.0 - 1.0

---

Citation

If you use this model, please cite:

@article{gonzalez2025hippie,
  title   = {HIPPIE: High-dimensional Interpretation for Physiological Patterns in Intercellular Electrophysiology},
  author  = {Gonzalez-Ferrer, Jesus and Lehrer, Julian and Alvarez-Esteban, Bruno and
             Schweiger, Hunter E. and Geng, Jinghui and Mostajo-Radji, Mohammed A. and
             Haussler, David and Teodorescu, Mircea},
  journal = {Nature Communications},
  year    = {2025},
  note    = {In revision}
}

---

License

Apache 2.0. See LICENSE for details.