# Aggregating rounds

Once the rounds are registered, aggregation matches each cell in the reference plate to its counterpart in every other round and merges their features into a single row.

``` bash
python bin/aggregate_plates.py 4126
```

Pass the **reference** plate ID -- the first one given to `align_plates.py`.

| Option | Default | Purpose |
|----|----|----|
| `--method {0,1,2,3}` | 3 | Matching strategy -- see below |
| `--threshold N` | 25 | Distance threshold for accepting a mapping |
| `--std N` | 6 | Standard deviations from the mean distance before exclusion |
| `--gallery N` | 4 | Size of the mapping gallery grid |
| `--sample-alignments` | off | Use per-sample rather than per-well alignments |
| `--seed N` | random | Random seed for sampling |
| `--env NAME` | -- | Environment file to load |


# Matching methods

| `--method` | Strategy                   |
|------------|----------------------------|
| `0`        | Minimum distance           |
| `1`        | KD-tree minimum distance   |
| `2`        | Greedy nearest neighbour   |
| `3`        | **Mask overlap** (default) |

Mask overlap is the default because it is the most conservative. The three distance-based methods will always find *a* nearest neighbour, even when the correct partner is absent -- a cell lost between rounds gets matched to whichever neighbour happens to be closest. Requiring the two nuclei to physically overlap refuses that match instead of inventing it.

Use a distance method only when overlap fails systematically, for instance when the rounds are separated by enough time that cells have visibly moved.


# What you get

The aggregated plate behaves like any other from here:

``` bash
uv run cellview import plate 4126
```

Each cell now carries the features from every round, so a single row holds all the markers measured across the experiment.

> **Warning: Warning**
>
> **Orphan nuclei are dropped.** A nucleus with no matching cell mask is removed during the merge rather than carried through with null features. Because cells are larger than nuclei, border filtering removes them more often, so the loss concentrates at canvas edges and around unimaged regions. Raw counts near those edges are biased downwards by an unrecorded amount.

Before trusting the result, run the [quality-control report](quality-control.md).