fix: use bin centers, not left edges, for pLDDT/PAE/PDE classificatio…

[Pana-TsK]

Summary

all_atom_plddt_loss, pae_loss, and pde_loss each discretize a continuous error value into classification bins via binned_one_hot(x, v_bins), which assigns x to whichever entry of v_bins is nearest. All three built
v_bins as bin_min + arange(no_bins) * bin_size — the left edge of each bin — instead of the bin center.

get_bin_centers (core/metrics/confidence.py), which decodes predicted bin probabilities back into pLDDT/PAE/PDE scores at inference time, and all_atom_distogram_loss's target-binning both correctly use bin_min + bin_size/2 + arange(no_bins) * bin_size.

Impact

The mismatch meant these three losses were trained against a grid shifted half a bin width from the grid their own output is decoded against, on essentially every training step (confidence losses are enabled by default). This is a systematic, silent miscalibration of every reported pLDDT/PAE/PDE score — it doesn't affect predicted coordinates (distogram/diffusion/FAPE
losses are unaffected), only the confidence heads' training signal and the scores derived from them.

Changes

• openfold3/core/loss/confidence.py: add the bin_size / 2 offset in all_atom_plddt_loss, pae_loss, and pde_loss, mirroring the pattern distogram.py already uses correctly.

Testing

Added one regression test per loss in openfold3/tests/test_confidence_loss.py that patches binned_one_hot to capture the v_bins it's actually called with and asserts it matches get_bin_centers's output for the same bin_min/bin_max/no_bins. Confirmed each new test fails against the pre-fix code and passes after. Full existing test suite for these losses (test_confidence_loss.py, test_distogram_loss.py, test_diffusion_loss.py, test_loss_weights.py) passes with no regressions. ruff format --check and ruff check clean.

[jandom]

Looping in @gnikolenyi – @Pana-TsK you're absolutely on a roll here, well done!