Compare through a stated relationship
Sheaf methods connect local data through explicit maps. BioTwin applies this idea to selected molecular representations and computes consistency residuals under those maps. The residual tells a researcher where the assumed relationship is poorly satisfied; interpreting it requires checking the quantities, units and mapping.
The relationship between views becomes something you can examine, rather than a hidden assumption in a combined score. That gives a disagreement a useful next step: inspect the input, improve the mapping or seek the observation that separates the explanations.
A small molecule changed the structural story
In a reported T4 lysozyme comparison, a Cα description could not represent retained benzene that an all-atom clearance model could see. The largest disagreements involved that structure, which an orientation check also flagged. The diagnostic value came from exposing a specific representational omission.
Use disagreement to choose the next observation
Paired with molecular fit/access and ensemble analysis, this approach could help determine whether a pocket conclusion depends on omitted atoms, a conformation or an assumed mapping. The applied comparison was bespoke; a broad automatic consistency atlas remains research. Earlier sheaf variants did not always improve on simpler models. Agreement under a chosen map also cannot make shared input data into independent corroboration.
RMSD, correlation and feature comparisons remain useful when their quantities match. Sheaf-based analysis makes the cross-representation relationship itself part of the test.
| Question | Familiar approach | BioTwin | Why it matters |
|---|---|---|---|
| Compare models | Coordinate alignment or descriptor comparison measures differences in a shared representation. | Uses explicit maps between representations and calculates residuals. | The comparison can expose a relationship that fails, even when the outputs have different forms. |
| Explain a mismatch | A structural discrepancy prompts inspection of input and preprocessing choices. | A reported Cα/all-atom control localized an omission involving retained benzene. | A missing physical feature can be investigated before interpreting disagreement as biology. |
| Combine evidence | A joint score can summarize several features for a task. | The proposed composition retains incompatible readouts and their mappings for review. | A team can choose a discriminating measurement instead of hiding the conflict in an average. |
The mathematics is established. The BioTwin evidence is a narrow diagnostic application; residuals can reveal a bad mapping as well as a meaningful structural difference.