Loki2 documentation
======================

Loki2 is a single-cell pathology foundation model that uses an encoder–decoder architecture to reliably translate nucleus morphology into molecular-level information. Loki2 is trained on UniSeg, a pan-tissue resource of 15 million cells with paired segmentation and transcriptomic measurements from multiple spatial technologies. This training strategy enables three capabilities from hematoxylin and eosin (H&E) images alone: universal cell type inference, in silico spatial transcriptomics by retrieving transcriptionally matched cells from reference atlases, and continuous morphological pseudotime inference of cell state transitions. By further aggregating single-cell features, Loki2 supports in silico protein staining and whole-slide clinical inference. These results show that routine histology contains far richer molecular and cellular information than previously recognized, and that Loki2 provides a general framework for accessing this information across technologies, tissues, disease contexts, and scales.

Please find our github on `GitHub <https://github.com/GuangyuWangLab2021/pathweb>`_.

.. image:: images/Loki2.png
  :width: 100%


.. toctree::
   :maxdepth: 2
   :caption: API

   autoapi/index


.. toctree::
   :maxdepth: 2
   :caption: Tutorials

   notebooks/Installation
   notebooks/Loki2_cell_type_inference
   notebooks/Loki2_morph_retrieve
   notebooks/Loki2_morph_psdtime_CRC
   notebooks/Loki2_morph_psdtime_PRAD
   notebooks/Loki2_in_silico_immunostaining
   notebooks/Loki2_multiple_instance_learning