Two AI models can predict where the same protein binds to DNA while relying on different clues. Tangermeme, a toolkit described in Nature Methods on 8 October, makes that difference visible: BPNet and Beluga learn strikingly different sequence rules for predicting MYC binding.[1]
BPNet’s predictions depend almost exclusively on the MYC motif—a short, recurring DNA pattern. Beluga draws on a broader set of motifs associated with DNA accessibility and the initiation of transcription. The comparison extends beyond one example: the study counts these patterns across the MYC binding sites being analysed.[1]
The toolkit finds influential stretches of DNA, groups them into variable-length segments called seqlets, and matches them to a motif database. Counting those annotations and their pairings turns a landscape of individual nucleotide scores into a description of the model’s learned regulatory rules.[1]
Change the sequence, then ask again
Tangermeme’s documented operations let researchers interrogate a trained model by changing its inputs. Each comparison asks a specific question about the prediction. Repeating a change across several background sequences helps distinguish an effect that travels across contexts from one tied to a particular stretch of DNA.[2]
Three ways to probe a sequence model
| Sequence change | Question for the model |
|---|---|
| Insert a motif into different backgrounds | How does this pattern change the prediction across contexts?Source 2 |
| Shuffle a motif or its surroundings | Which local sequence information does the prediction depend on?Source 2 |
| Vary the distance between motifs | Does their predicted interaction depend on their spacing?Source 2 |
Changing a prediction measures the model’s dependence on a sequence feature, giving researchers a concrete hypothesis to test in cells.[2]
A model’s clue can become a biological experiment
An earlier BPNet study shows what the next step can look like. In 2021, researchers tested a predicted interaction between Sox2 and Nanog motifs by editing them in mouse embryonic stem cells. Disrupting the Sox2 motif reduced nearby Nanog binding; disrupting the Nanog motif did not noticeably alter Sox2 binding. They examined three independently derived clones for each mutation.[3]
That experiment exposed a directional relationship: one binding site helped the other. The sequence edits and the measurements supplied the biological test. Tangermeme’s useful contribution is to make the preceding questions easier to formulate—what information a model uses, which neighbouring patterns matter, and which specific edit would distinguish competing explanations.[3][2]
Sources & further reading
- Tangermeme: a toolkit for understanding cis-regulatory logic using deep learning models Nature Methods · 8 Oct 2026 · Accessed 18 Oct 2026
- Tangermeme: biological sequence analysis and model interrogation Jacob Schreiber / GitHub · Accessed 18 Oct 2026
- Base-resolution models of transcription-factor binding reveal soft motif syntax Nature Genetics · 18 Feb 2021 · Accessed 18 Oct 2026