PIPENN-EMB

Sequence-based interface predictors are routinely benchmarked on test sets filtered to 25% sequence identity, and still much of what they appear to know comes from having seen a relative of the query. The features give it away: a position-specific scoring matrix is built by running PSI-BLAST until homologs turn up, so a protein with nothing recognisable in the databases arrives at the network with an uninformative feature vector. That is exactly the protein a biologist most wants predictions for — an uncharacterised gene product from a newly sequenced pathogen genome.

PIPENN-EMB, from the bioinformatics group at Vrije Universiteit Amsterdam, answers this by throwing out the hand-crafted feature stack. Every residue of the input sequence is instead represented by a vector taken from ProtT5-XL, the frozen 3-billion-parameter encoder from ProtTrans, and a small supervised network on top decides which residues sit in a protein-protein interface. Because the embedding is produced in a single forward pass over the sequence, there is no alignment step at inference and no dependence on finding relatives.

It is the embedding-based successor to PIPENN (Stringer et al., 2022), the group's earlier ensemble of sequence-feature predictors, and shares a code base with it: the two are selected by a one-hot-encoding flag rather than shipped separately.

#Key Features

• Interface calls from a bare sequence: The only required input is an amino acid sequence; no structure, no multiple sequence alignment, and no PSI-BLAST profile.
• Embeddings absorb the old features: Across ablations that added predicted secondary structure, solvent accessibility and PSSM back on top of the embeddings, MCC stayed between 0.290 and 0.313, so only protein length was kept alongside them.
• A deliberately trimmed ensemble: Of the six architectures in the PIPENN suite, the combination of a recurrent net, a residual CNN and a U-Net scored best, cutting roughly 6 million parameters from the ensemble.
• Flat performance across the homology gradient: Per-protein AUROC shows no significant correlation with similarity to the training set on any of the three evaluation sets.
• Checkpoints beyond protein-protein: Trained weights are released for the protein, DNA/RNA and small-molecule interaction subsets of the BioLIP-derived training data — but only the per-architecture nets. The trained ensnet that combines them is published for the small-molecule subset alone, so assembling the protein-protein ensemble means retraining that layer yourself.

#Technical Details

The six PIPENN architectures — a fully connected net, a dilated CNN, a residual CNN, a two-layer GRU network, a CNN/RNN hybrid and a U-Net — each emit a per-residue interface probability, and an ensemble net (ensnet) takes those probabilities as its input. PIPENN-EMB is the configuration that pairs the recurrent net, residual CNN and U-Net with ensnet, using ProtT5-XL embeddings plus protein length as features. Training used BioDL_P_TR, built from PDB interface annotations and post-processed by clustering at 25% identity and restricting lengths to 30–700 residues; hyperparameters were carried over unchanged from PIPENN.

On the held-out BioDL_P_TE set the embedding models reach MCC 0.313 and AUROC 0.800, against 0.249 and 0.755 for the feature-based predecessor. A head-to-head between language models, training the dilated CNN on each, favoured ProtT5-XL over ESM-2 at 3B parameters on the ZK448 benchmark (F1 0.492 vs 0.440, MCC 0.376 vs 0.312, AUROC 0.811 vs 0.791). The generalisation analysis used BLASTP to score every test protein by sequence identity times query coverage against each method's own training set: AUROC for PIPENN-EMB is uncorrelated with that score across BioDL_P_TE, ZK448 and a tuberculosis set, and its AUROC in the 0–15% band matches its AUROC in the 31–45% band.

#Applications

The practical use is triaging residues for experiments on proteins where structure-based prediction is not an option or would be circular — orphan proteins, viral and bacterial gene products, and regions that AlphaFold 2 models with low confidence. The authors demonstrate this on 25 drug-resistance-associated proteins from Mycobacterium tuberculosis drawn from a GWAS panel, where sequence-based PIPENN-EMB (AUROC 0.659, MCC 0.319) lands between two structure-based predictors that were given predicted structures to work from. A public webserver takes a pasted sequence and returns a table of predicted interface positions, and the pretrained checkpoints can be run locally.

#Impact

PIPENN-EMB's more durable contribution may be methodological rather than a leaderboard position: the paper measures how much of a method's benchmark score is explained by homology to its own training data, and reports that a competing sequence-based predictor with better headline numbers on ZK448 and BioDL_P_TE — Seq-InSite — loses that advantage as similarity to its training set drops, while PIPENN-EMB does not. Read straight, the comparison is a caution about benchmark sets that have been optimised against for years. The limitations are candid: interface prediction accuracy is plateauing overall, the webserver serves a single architecture rather than the full ensemble, and the weights are distributed as HDF5 files from a university download page with no stated license.