Design over nature

For each selected cell pair: the designed sequences and the natural promoters they were bred from, side by side. A promoter is the stretch of DNA in front of a gene that decides how strongly it is switched on. Its selectivity is the predicted activity in the target cell type divided by activity in the cell type it should leave quiet. The model reports natural-log expression scores, so we exponentiate the ON-minus-OFF score difference to obtain this fold ratio. Design over nature then divides the design’s target/off-target activity ratio by that of the best natural promoter — both measured in the same genomic window, in the same run, under the same two context strings. These are ratios on the normalised TPM + 1 scale, never raw TPM. The checkpoint preserves ranking but compresses magnitudes, so they are not calibrated absolute expression.

Why the same genomic window matters. Regulatory DNA elements interact with one another, and the model was trained on long endogenous sequences. The same 600 bp promoter can therefore receive a different prediction in different surroundings. Every comparison below uses a design and a natural promoter measured in the same window; results from separate windows are not substituted.

The two databases are stacked here, never merged. A natural promoter’s rank is partly recall of the model’s own training data — it was trained on essentially every gene’s promoter — and a generated design’s rank cannot be, because the sequence exists in no genome. Pooling them into one ranking would put those two kinds of rank in one column, so the tables below stay separate.

What makes the comparison legitimate. The natural promoters below are not quoted from some other experiment: they were members of the generation run’s own starting population, scored in the same window, in the same run, under the same two verbatim context strings, on the same checkpoint. That is the only condition under which a design’s and a natural promoter’s fold activity ratios are directly comparable — a result from any other run or window does not meet it, and cannot be substituted here.