Datasets:

KuanP
/

PULSAR_DONORxEMBED_zero_shot

dataset_info:
  features:
    - name: donor_id
      dtype: string
    - name: embedding
      sequence: float32
    - name: disease_label
      dtype: string
    - name: dataset_id
      dtype: string
  splits:
    - name: train
      num_bytes: 6025971
      num_examples: 2804
  download_size: 7645385
  dataset_size: 6025971
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
license: mit

Dataset Card for DONORxEMBED-zero-shot

This dataset contains pre-computed PULSAR donor embeddings for 2,804 peripheral blood mononuclear cell (PBMC) samples from 41 independent studies, totaling over 10 million cells with paired diagnostic annotations. The embeddings are generated using PULSAR (Patient Understanding Leveraging Single-cell universAl Representation), a multi-scale, multi-cellular foundation model designed to transform PBMC scRNA-seq samples into unified donor representations.

This dataset hosts the zero-shot model (PULSAR-pbmc) embeddings. A disease-aligned variant is also available (see https://huggingface.co/datasets/KuanP/PULSAR_DONORxEMBED_aligned).

Dataset Details

Dataset Description

Curated by: Kuan Pang (Stanford University, kuanpang@stanford.edu)
License: MIT

Uses

Direct Use

This dataset is intended for:

Disease classification through reference-based mapping using k-nearest neighbor search in the embedding space
Building searchable vector databases for flexible, scalable disease classification
Benchmarking machine learning models for immune-related disease prediction
Cross-cohort generalization studies in immunology
Population-level analysis of immune variation across diseases

Dataset Structure

The dataset contains 2,804 samples with the following fields:

donor_id (string): Unique identifier for each donor/sample
embedding (sequence of float32): 512-dimensional PULSAR donor embedding vector
disease_label (string): Disease annotation following Mondo Disease Ontology standards
dataset_id (string): Source dataset identifier

Dataset Creation

Curation Rationale

This dataset was created to enable reference-based disease classification and to demonstrate PULSAR's capacity to capture disease-relevant donor-level signatures. The comprehensive reference cohort allows for scalable disease classification through vector database search, enabling continuous expansion without model re-training.

Source Data

The dataset was assembled from 41 sources:

32 datasets from CZ CELLxGENE Census database
1 dataset from Single Cell Portal (SCP)
8 additional curated sources

Citation

BibTeX:

[More Information Needed]