README.md

metadata

license: apache-2.0
tags:
  - biology
  - rna-seq
  - tabular-classification
  - tabular-regression
  - survival-analysis
  - in-context-learning
datasets:
  - hanzi-bulk-collab/bulk-datasets-private

PRTS

PRTS (Prior-trained RNA-seq Task Solver) predicts binary, multiclass, continuous and survival endpoints from bulk transcriptomes in context. A labelled cohort provides the context for predictions on new samples.

This repository provides the evaluated checkpoint for In-context learning predicts clinical endpoints from bulk transcriptomes across diseases and platforms, by Han Wang, Lennart Purucker, Huafeng Liu, Ziyin Luo, Junfeng Hao, Chao Zhu and Chunjie Tian.

• Code and applications
• PreBULK, RealBulk and benchmark data
• Architecture: 219,595,334 parameters in 471 tensors; four cell blocks and twenty row blocks.
• File: prts.pt, a PyTorch checkpoint containing config.model and the model state dictionary.
• Export: model configuration and all parameter tensors are unchanged from the author-supplied checkpoint. Training-run bookkeeping and machine-specific paths are omitted from this inference export.
• Checkpoint size and SHA-256

Load the model

pip install "git+https://github.com/Hanziwww/PRTS.git#subdirectory=prts" huggingface_hub

from huggingface_hub import hf_hub_download
from prts import load

checkpoint = hf_hub_download("HanZiwww/PRTS", "prts.pt")
model = load(checkpoint, device="cpu")  # use "cuda" for a CUDA GPU

See the inference documentation for expression scaling, task construction and prediction outputs. The default Thinking preset averages four passes; the Off preset is also available.

Scope and limitations

The model was evaluated on retrospective public transcriptomic cohorts. Accuracy varies with endpoint, cohort, platform and the reference samples supplied as context. Cross-cohort and unseen-site evaluation should be considered when assessing transfer to a new study. This repository supplies a research model; it is not a validated clinical diagnostic device.

Licence

The checkpoint and author-owned code are available under Apache License 2.0. Original datasets and third-party code retain their own terms.