PIPENN

Which residues of a protein actually touch a partner is the difference between a sequence and a hypothesis a wet lab can test, and the experiments that settle it are slow. Sequence-based predictors fill the gap, but each commits to one network design, and the designs disagree about what a residue even is: a fully connected net sees one residue's feature vector and nothing of its neighbours, a dilated convolutional net widens its receptive field layer by layer, a recurrent net carries the whole chain. Each captures overlapping but distinct patterns, and none dominates on every protein.

PIPENN, from the IBIVU bioinformatics group at Vrije Universiteit Amsterdam, turns that disagreement into the method. The same sequence-derived feature set is pushed through six architectures — a fully connected net, a dilated CNN, a U-Net, a residual CNN, a GRU network and a CNN/RNN hybrid — and a seventh net, ensnet, takes the six per-residue probabilities as its own input and emits the final call. The six are trained on a training split; ensnet is then trained on the predictions the six make on a held-out validation split, so it learns where each architecture can be trusted rather than averaging them blindly.

It covers three interaction classes at once — protein–protein, protein–DNA/RNA and protein–small-molecule — trained on BioDL, a corpus the group built from PDB interface geometry and the BioLip ligand-binding database. The successor PIPENN-EMB keeps the ensemble and swaps the hand-crafted features for protein language model embeddings.

#Key Features

• One feature set, six architectures, one verdict: The ensemble net beats every individual architecture on all of the paper's test sets, not only on average.
• Three interaction types with type-specific checkpoints: Weights trained on the protein, small-molecule and nucleotide subsets consistently outperform the generic all-types model on their own test sets — AUC-ROC 0.755 vs 0.733, 0.864 vs 0.826 and 0.894 vs 0.835 respectively.
• Sequence in, per-residue probability out: Inputs are one-hot amino acid identity, a PSI-BLAST PSSM, NetSurfP-predicted secondary structure and solvent accessibility, a Pfam domain flag and protein length, plus window-averaged aggregates over 3, 5, 7 and 9 neighbouring residues.
• Checkpoints for every architecture and every training set: Seven trained .hdf5 files ship per training set across five training sets, and the repository documents applying them without retraining.
• An audited design rather than a chosen one: Initialisation, regularisation, loss, activation and encoding were each ablated, which is how one-hot encoding beat ProtVec embeddings.

#Technical Details

BioDL was assembled from 138,729 PDB structures at 2.5 Å or better plus the whole BioLip database, annotating a residue as interacting when a ligand atom falls within the sum of van der Waals radii plus 0.5 Å. Residues were mapped to UniProt via SIFTS, clustered at 25% identity, filtered to lengths of 26–700, and stripped of every cluster containing a ZK448 protein. The architectures were built in Keras on TensorFlow 2.1 and trained with Adam at a learning rate of 1e-4, 20% dropout with batch normalisation, early stopping on AUC, and a cross-entropy loss carrying an extra term for the roughly 11% interface-residue class imbalance.

On the held-out BioDL protein–protein test set the ensemble reaches AUC-ROC 0.755 against 0.739 for the best single architecture, with MCC 0.249. On ZK448, the independent benchmark curated by the Kurgan lab and excluded from training, it reports AUC-ROC 0.718 for protein–protein, 0.823 for protein–nucleotide and 0.842 for protein–small-molecule interfaces, ahead of the published sequence-based predictors evaluated on the same sets — SeRenDIP on PPI, SCRIBER on small molecules, DRNApred on DNA/RNA.

#Applications

The practical use is narrowing a protein down to the residues worth mutating — picking alanine-scan targets, reading a candidate epitope off a viral surface protein, or annotating a gene product with no solved structure. A public webserver takes a pasted sequence and returns a table of predicted interface positions, with a choice of the protein, nucleic acid, small molecule or all-types model; the same checkpoints run locally, which is the route that scales past a handful of sequences.

#Impact

PIPENN's durable contributions are the BioDL corpus, a systematic account of which architectural building blocks move residue-level interface prediction, and the finding that ensembling across architectures pays off where any single design does not. The group's own follow-up sharpens the picture: PIPENN-EMB reports MCC 0.313 and AUC-ROC 0.800 on the same protein–protein test set against 0.249 and 0.755 here, crediting language model embeddings rather than the ensemble. Limitations remain: the PSSM feature comes from a PSI-BLAST search, so predictions lean on finding homologs for the query; running a checkpoint on your own sequences means matching the fixed CSV layout and editing constants in a Python module; and the weights are HDF5 files on a university download page, covered along with the datasets by the same GPL-3.0 license as the code.