AIDO.Tissue

AIDO.Tissue is a spatially-informed pretraining framework for single-cell spatial transcriptomics, developed by GenBio AI with collaborators at MBZUAI and Carnegie Mellon University and posted as a preprint in July 2025. It is the tissue-scale member of the AIDO (AI-Driven Digital Organism) family, alongside AIDO.DNA, AIDO.RNA, AIDO.Protein, and AIDO.Cell.

Most single-cell foundation models — including Geneformer, scGPT, and scFoundation — are pretrained on dissociated cells, where dissociation has destroyed each cell's position in the tissue. Spatially aware models such as CellPLM and Nicheformer added spatial data to pretraining but still process each sample centered on a single cell, so cross-cell relationships must be inferred from pooled or pairwise representations rather than attended to at the gene level.

AIDO.Tissue changes the input unit. Each training example is a center cell concatenated with its k nearest spatial neighbors, a construction the authors compare to multiple sequence alignment inputs in protein language modeling, and self-attention then runs across genes and cells simultaneously. To keep that affordable, the encoder sees only expressed (non-zero) genes across all cells in the neighborhood, while the decoder reconstructs only the masked expression values of the center cell.

#Key Features

• Multi-cell neighborhood input: Each example stacks a center cell with its 8 to 64 nearest neighbors from spatial coordinates, so intercellular dependencies are modeled directly rather than inferred.
• Asymmetric encoder-decoder: Restricting the encoder to expressed genes and the decoder to the center cell cuts the sequence length attention must cover, making 64-neighbor contexts tractable over a 19,264-gene panel.
• Two-dimensional rotary position embeddings: Each token carries a gene index and a cell index, letting the model distinguish "the same gene in a different cell" from "a different gene in the same cell."
• Column-wise masking: Masking a gene across all cells at once stops the model copying the answer from a neighbor, forcing it to learn genuine expression structure.
• Neighborhood size beats parameter count: Raising k from 8 to 64 improves downstream performance consistently, while scaling 3M to 60M parameters yields marginal gains.

#Technical Details

Pretraining used roughly 22 million cells from about 76 tissue slides across three imaging-based platforms — Vizgen, NanoString, and 10x Genomics — with a fixed 19,264-gene panel and paired x/y cell centroids. Gene tokens are the element-wise sum of a learned gene-name embedding, an MLP projection of the expression value, and the 2D rotary positional embedding; the objective is mean squared error on masked positions of the center cell, with separate mask ratios for zero and non-zero entries following the abundance-bias treatment introduced in xTrimoGene. The 3M and 60M parameter models were trained for 150,000 iterations at global batch size 128 with Adam and a learning rate warmed from 2 × 10⁻⁵ to 2 × 10⁻⁴ under cosine decay.

Evaluation follows the three benchmark tasks defined by Nicheformer. On the CosMx human liver dataset, AIDO.Tissue reaches an F1 of 0.77 on 22-way spatial cell-type classification versus 0.73 for Nicheformer, and 0.75 versus 0.50 on six-way niche-type prediction; on cell-density regression over the Xenium human lung dataset it achieves a mean absolute error of 4.583 against Nicheformer's 7.084. The released checkpoints, distributed as GB.Tissue-3M and GB.Tissue-60M, use an eight-neighbor context; their model card reports niche F1 of 0.67 and density MAE of 4.44 with R² = 0.55, against 0.53 / 6.17 for scFoundation and 0.50 / 7.08 for Nicheformer. Weights and fine-tuning configurations ship through the ModelGenerator framework under the GenBio AI Community License, which permits non-commercial use only.

#Applications

The model targets researchers working with imaging-based spatial transcriptomics — CosMx, Xenium, and MERFISH experiments that yield hundreds of thousands of cells with coordinates attached. Fine-tuned niche-type prediction assigns each cell to a microenvironment class such as an immune-rich region, stromal domain, or tumor core, turning a cell-by-gene matrix into an interpretable tissue map; on the liver benchmark this recovers zonation structure with spatially coherent boundaries. Cell-density regression estimates local cellular composition within a radius, a quantity that differs systematically between healthy and tumor tissue. The model also emits per-gene embeddings that pool into spatially aware cell representations for clustering.

#Impact

AIDO.Tissue's most transferable result concerns scaling: for spatial transcriptomics, widening the spatial receptive field pays off more than adding parameters. That reframes where effort should go in a subfield that has largely imported scaling assumptions from single-cell and language modeling, suggesting future gains will come from richer spatial context and more diverse slide collections rather than larger backbones. The asymmetric encoder-decoder is a practical contribution in its own right, showing that gene-level cross-cell attention is affordable at 64-cell neighborhoods. The work remains a short preprint awaiting peer review, and its evaluation is narrow — three tasks on two tissues, without cross-platform generalization tests or comparison against graph neural network baselines. The non-commercial license further limits industrial adoption.