PeTriPOV

PeTriPOV is a BERT-style protein encoder that reads amino acid sequence and backbone geometry together, developed at IPSiM in Montpellier, a plant sciences research unit of CNRS and INRAE. It is the successor to the same group's PeTriBERT, a 40-million-parameter transformer released in 2022 that added three-dimensional coordinates to a masked language model and used the result for inverse folding: predicting which amino acid sequence would fold into a given backbone.

The change from PeTriBERT to PeTriPOV is a change in how structure enters the network. PeTriBERT encoded each residue's position and orientation absolutely, as extra components of that residue's input vector, which makes the representation depend on where the protein sits in space. PeTriPOV instead computes the rigid-body transformation between every pair of residue frames, featurizes it, and injects it as an additive bias on the attention scores — the same device AlphaFold 2 uses for its pair representation. Because a relative transform is unchanged by rotating or translating the whole molecule, the encoding is invariant by construction and needs no pose augmentation during training.

The 2025 preprint that introduces PeTriPOV uses it for a second task. Fine-tuned into a variant called PeTriPPI, the encoder scores whether a peptide and a protein interact, taking confidence maps produced by a stripped-down AlphaFold-Multimer pipeline as additional input channels alongside the sequence and structure.

#Key Features

• Relative frame encoding: An orthonormal frame is built for each residue from its Cα, C, and N atoms; the transform relating frames i and j is compressed to a centroid offset and three Euler angles, passed through a small MLP, and added to the attention logits.
• Rotation and translation invariance: Because only relative transforms enter the model, global rigid motion of the input structure leaves every prediction unchanged.
• Inverse folding from a backbone: Trained by masked language modeling over structures, the encoder proposes sequences compatible with a supplied PDB backbone.
• Two released variants: PeTriPOV shares one attention-bias projection across layers; PeTriMPOV gives each layer its own, which fits the training data better without overfitting.
• Fine-tunable to interaction scoring: PeTriPPI replaces the output head and folds per-residue pLDDT and pairwise PAE and distogram tensors in through adapter layers, the pairwise ones summed into the same attention-bias matrix.

#Technical Details

PeTriPOV is a five-layer transformer with 40 million parameters, trained by masked language modeling on protein structures drawn from the AlphaFold Database — the same corpus and hyperparameters used for PeTriBERT, so the two are directly comparable. On the held-out test set, PeTriPOV recovers masked residues with 0.55 accuracy against 0.52 for PeTriBERT, at a cross-entropy of 1.42 versus 1.54; the unshared-bias PeTriMPOV variant reaches 0.60 accuracy at 1.25 cross-entropy.

PeTriPPI was evaluated on a curated build of the Propedia v2.3 peptide-protein database, augmented by random non-interacting pairs to roughly 100,000 examples at 10% positives, with AlphaFold features precomputed on an HPC cluster. It reaches precision 0.83 and recall 0.74. The recall figure is the point of the model: gradient-boosted and random-forest baselines fed only summary statistics of the same AlphaFold outputs achieve comparable or better precision — random forest reaches 0.88 — but recall no higher than 0.47, because collapsing per-residue and pairwise confidence maps into moments discards the spatial pattern. Fine-tuning used a 1e-5 learning rate, weighted cross-entropy for class imbalance, and an initial frozen-encoder phase.

#Applications

The encoder serves protein and peptide design work that starts from a structure rather than a sequence. Inverse folding proposes sequences for a target backbone; a binder-design script generates candidate peptide ligands for a receptor; and the PPI head ranks candidate peptide-protein pairs so that only the highest-confidence designs advance to synthesis and assay. The authors frame this last use as an in-silico filtering stage placed ahead of in-vivo testing, where a high-precision, moderate-recall operating point is the right trade: a false positive costs an experiment, a missed candidate only one entry in a large generated pool.

#Impact

PeTriPOV is a small, lab-scale model, and its evaluations are the authors' own — there is no independent benchmarking, and the interaction results rest on a single dataset. Its interest lies in what it isolates. By holding the training corpus and hyperparameters fixed and changing only how geometry reaches the attention mechanism, the work gives a clean measurement of what relative frame encoding buys over absolute coordinates in a protein language model, a design choice that larger structure-aware encoders adopt without ablating. The accompanying HAPI library, which decouples AlphaFold 2's search pipeline from its network so that structures can be supplied as templates and MSAs skipped, is a reusable piece in its own right for anyone scoring interactions in bulk. Both repositories are Apache 2.0, but only for the code: the four pretrained checkpoints and the roughly 10 GB AlphaFold Database training corpus arrive as Google Drive archives linked from the PeTriBox README under no stated terms, and the preprint sits on bioRxiv under the server's no-reuse license, so the sole document carrying the performance numbers cannot be redistributed without permission.