We are evaluating already-cached human CONTACT_MAPS predictions from AlphaGenome API v1, using SDK 0.10.0 and ALL_FOLDS (IMR-90 and GM12878). The returned signed values L are on the native 2048-bp grid. We are not requesting new predictions or access to raw training data.
Could you point us to a version-applicable specification or implementation that defines the units of L and the intended mapping f(L) to a positive, normalized contact signal on that same native grid? In particular, is the inverse exp(L), 2**L, or another formula, and does it require an additional offset, scale, or other operation? What positive quantity does that mapping represent after the documented interpolation to 2048 bp and ALL_FOLDS aggregation?
The metadata describes a log-fold change over a distance-dependent expectation, while the paper writes “log” without an explicit base. We have checked the v1 output declaration/decoder/proto, research contact head and prepared-target loader, and the official contact-map tutorial examples; these do not identify a version-bound inverse. A formula or a precise documentation/code link with its applicable API/model version would resolve the issue.
Our downstream use is positive-signal diamond insulation and boundary analysis, so selecting a base by comparing benchmark scores would be inappropriate. We only need a documented native-grid positive representation; we do not need unique recovery of raw counts, reconstruction of the original 1-kb map, or the original expected/pseudocount arrays. If no such positive representation is intended for API v1, please confirm that limitation instead.