PULSAR-all-tissue

PULSAR (Patient Understanding Leveraging Single-cell universAl Representation) is a multi-scale, multicellular foundation model that turns a set of single-cell embeddings (UCE, 1,280-d) into a 512-d donor-level embedding. This repo hosts the pan-tissue pre-trained model (stage 1 of PULSAR pre-training, CELLxGENE corpus of ~36.2M cells from 6,807 samples) including the full self-supervised training head (decoder, mask projector and mask value), so it can be used both for inference and as a starting point for self-supervised fine-tuning with the training scripts in the GitHub repository. The PBMC model KuanP/PULSAR-pbmc was continually pre-trained from this checkpoint.

Usage

Inference (encoder only). The five training-head tensors (mask_element, in_proj_mask_embedding.proj_1.*, in_proj_mask_embedding.proj_2.*) are reported as unused when loading this class, which is expected.

from pulsar.model import PULSAR

model = PULSAR.from_pretrained("KuanP/PULSAR-all-tissue")

Self-supervised fine-tuning (full model with training head):

from pulsar.model import PULSARForSelfSupervisedFinetuning

model = PULSARForSelfSupervisedFinetuning.from_pretrained("KuanP/PULSAR-all-tissue")

Model Sources

• Repository: https://github.com/snap-stanford/PULSAR

• Paper: https://www.biorxiv.org/content/10.1101/2025.11.24.685470v1

• Related: https://huggingface.co/KuanP/PULSAR-pbmc , https://huggingface.co/KuanP/PULSAR-aligned

• Developed by: Kuan Pang (Stanford University)

• License: MIT