LiverDCP

A hepatocyte in a healthy liver and a hepatocyte in a cirrhotic one draw on the same genome, but the protein interactions actually available in each are different — partners drop out as expression shifts, and a protein's functional role moves with them. A context-free protein embedding hands both hepatocytes the same vector, so a model asked which target matters in cirrhosis has already lost the distinction it needs.

LiverDCP answers this with a Disease-Cell-Protein (DCP) paradigm: disease states, cellular composition, and protein interaction networks are represented jointly inside a single graph, so a protein's embedding is conditioned on the disease and the cell type it is being asked about. It was developed by the Zhao lab at the University of Texas Medical Branch in Galveston, with a collaborator at Baylor College of Medicine, and posted as a preprint in August 2026, instantiating the paradigm for the human liver across 286 disease-by-cell-type contexts.

Its closest conceptual sibling is PINNACLE, which also produces cell-type-resolved protein representations. The two differ in axis and in training regime: LiverDCP adds a disease dimension the multi-organ PINNACLE does not model, and where PINNACLE trains models independently within individual cellular contexts, LiverDCP couples all 286 contexts through a shared subgraph and one set of parameters — learning the interaction principles common to every context while preserving what is specific to each.

#Key Features

• Disease and cell type as joint context: Each protein receives a distinct 256-dimensional embedding in every one of 286 combinations of 11 liver diagnoses and 26 cell types, giving 777,609 context-specific representations over 12,085 proteins.
• A single jointly trained model: One shared graph attention backbone is trained once across all contexts rather than refit per disease or cohort, so the released artifact is a fixed checkpoint, not a pipeline to rerun.
• Geometry-aware two-phase training: Pretrained ESM-2 sequence features enter through a zero-initialized side adapter, and the graph encoder freezes once a validation quality threshold is met, adding sequence signal without distorting the embedding structure already learned.
• Context-resolved GWAS interpretation: A probe on frozen embeddings classifies GWAS risk genes at AUROC 0.946 in MASH and 0.970 in ALD, against 0.790 and 0.776 for sequence-only ESM baselines.
• Training-free target prioritization: Mahalanobis distance over the embeddings ranks OpenTargets Phase-II-and-beyond targets at an enrichment factor of 3.93 in the top decile for MASH, versus 2.38 for ESM-2, and recovers THRB — the only FDA-approved MASH target — at rank 25.

#Technical Details

The architecture is a hierarchical graph neural network: a graph attention backbone shared across every context-specific interactome, combined by semantic-attention fusion, with a metagraph connecting diagnoses to cell types. The hidden dimension is 256 across four attention heads. The graph is assembled from LiverHomo, the group's single-cell liver atlas of 1.16 million cells from 320 patients across 16 public datasets, and from roughly 1.9 million PrePPI-AF structural interaction predictions; per-diagnosis, per-cell-type differential expression defines each context's protein universe, and CellPhoneDB cell-cell interactions supply the metagraph edges.

Training optimizes knowledge-guided link prediction over PPI and metagraph edges. In the second phase, 480-dimensional ESM-2 embeddings are injected through a low-rank side adapter while the encoder freezes once validation quality passes 0.7, keeping the sequence signal from overwriting the graph-learned geometry. Base node features are random initializations, so embeddings are tied to the training graph's protein and context index space.

#Applications

Researchers working on liver disease can query the embedding table directly: for a given diagnosis and cell type it returns a representation that already encodes which interactions are available in that state. The demonstrated uses are placing GWAS risk genes into the cell types where their products are functionally active, and nominating therapeutic targets — including repurposing candidates — for MASH, a disease with a nearly empty approved-drug space. Because the precomputed embeddings are consumed frozen, downstream work needs no GPU training, and for most users the embedding table, not the encoder, is the practical artifact. Reconstructing the full module requires graph metadata rebuilt from the Zenodo deposit, as the checkpoint is not self-describing.

#Impact

LiverDCP extends context-aware protein representation from cell type alone to the disease-by-cell-type product, and argues that coupling contexts during training rather than fitting each separately is what lets shared interaction principles transfer. The authors state its limits plainly. The interaction scaffold is predicted rather than measured — PrePPI-AF is AlphaFold-informed Bayesian inference, and cell-type filtering uses scRNA expression as a proxy for protein presence. The model is human- and liver-only by construction, and because its embeddings are bound to the training graph's index space, extending the paradigm to another tissue means training a new DCP rather than transferring this one. Code is GPL-3.0 and the weights and embeddings CC BY 4.0, and the work is a preprint awaiting peer review.