{"id":5652,"date":"2026-11-11T13:19:20","date_gmt":"2026-11-11T06:19:20","guid":{"rendered":"https:\/\/bioturing.com\/blog\/?p=5652"},"modified":"2026-11-11T14:07:00","modified_gmt":"2026-11-11T07:07:00","slug":"sesame-peanuts-a-transcript-to-cell-assignment-method-for-overlapping-cells","status":"publish","type":"post","link":"https:\/\/bioturing.com\/blog\/sesame-peanuts-a-transcript-to-cell-assignment-method-for-overlapping-cells\/","title":{"rendered":"Sesame-peanuts: A transcript-to-cell assignment method for overlapping cells"},"content":{"rendered":"\n

1. Introduction<\/strong><\/h2>\n\n\n\n

1.1. Motivations<\/strong><\/h3>\n\n\n\n

In single-cell spatial transcriptomics, most downstream analyses run on the CELL\u00d7GENE expression matrix. This matrix is built by a process called transcript-to-cell assignment, which is commonly done by giving each transcript to the cell whose polygon it falls in. But a tissue section is a 3D object imaged in 2D. Where cells overlap in depth, their transcripts also overlap, and no 2D polygon can separate them.<\/p>\n\n\n

\n
\"\"
Figure 1.<\/strong> A crop showing macrophages sitting next to, or on top of, a sheet of tumour cells. The second image shows position of marker transcripts of the crop. Dataset: Xenium Human Lung Cancer [1]; marker genes from [2].<\/figcaption><\/figure>\n<\/div>\n\n\n

Figure 1 shows the physical limit of cell segmentation. There is no cut that separates the macrophages\u2019 transcripts from the tumour\u2019s: whatever boundary we draw, it will include tumour transcripts. If we let them mix as-is, three kinds of damage follow, ordered from least to worst:<\/p>\n\n\n\n

    \n
  1. Wrong cell-type prediction.<\/strong> An immune cell that infiltrates a tumour or an inflamed tissue can have its transcripts outnumbered by those of the cells it overlaps. Its true identity is drowned, and it is mis-annotated.<\/li>\n\n\n\n
  2. Wrong gene expression.<\/strong> Even when the cell type comes out right, the cell\u2019s expression profile is influenced by who its neighbours are.<\/li>\n\n\n\n
  3. False discoveries.<\/strong> As a result of (2), a tumour gene such as CAPN8 can look up-regulated in macrophages simply because they sit next to a tumour that makes it. Differential expression tests handle dropout and over-dispersion, but they have no concept of this kind of noise, so it comes out as false positives.<\/li>\n<\/ol>\n\n\n\n

    1.2. Problem statement<\/strong><\/h3>\n\n\n\n

    To undo the mixing, we need to move the leaked transcripts in both directions: send the foreign transcripts inside a polygon back to the cell that made them, and give each cell back its own transcripts that landed outside its polygon. Some transcripts belong to no cell at all and should be left as background.<\/p>\n\n\n\n

    In other words, the goal is to link every transcript to the cell that made it (Figure 2), then produce the CELL\u00d7GENE matrix.<\/p>\n\n\n

    \n
    \"\"
    Figure 2.<\/strong> The goal: each transcript linked to the cell that made it, or left as background (\u00d7).<\/figcaption><\/figure>\n<\/div>\n\n\n

    1.3. Related works<\/strong><\/h3>\n\n\n\n

    Two published methods also use expression to decide where transcripts belong. Neither removes the mixing.<\/p>\n\n\n\n

    Baysor<\/strong> [3] treats the slide as a mixture model of gene expression and transcript positions, and assigns transcripts to cells with expectation maximization. Each assignment is scored by the three factors of (1), where factor (b) is the problem: it rewards neighbouring transcripts for sharing a cell, so where two cells overlap and their transcripts are mixed, Baysor blends them into one region instead of sorting them apart.<\/p>\n\n\n

    \n
    \"\"<\/figure>\n<\/div>\n\n\n

    Proseg<\/strong> [2] comes closer. It also models each cell\u2019s expression, and lets cells reclaim transcripts from their neighbours by growing and shrinking their shapes one bin at a time. But its output is still a cell shape, and the matrix counts what falls inside it (Figure 3). Two cells cannot claim the same spot, so where cells overlap, the mixing remains<\/p>\n\n\n

    \n
    \"\"
    Figure 3.<\/strong> Proseg refines cell shapes, but still counts the transcripts inside each shape. Figures from [2].<\/figcaption><\/figure>\n<\/div>\n\n\n

    1.4. Sesame-peanuts<\/strong><\/h3>\n\n\n\n

    We propose sesame-peanuts, a completely unsupervised transcript-to-cell assignment method that sorts mixed transcripts without drawing new boundaries. Its inputs are:<\/p>\n\n\n\n