HPInet

Gram-negative bacteria use ten distinct secretion systems to inject effector proteins into host cells, and those effectors drive infection by hijacking specific human proteins. Mapping which effector binds which host target is central to understanding bacterial pathogenesis, but interaction assays are slow and the known effector-host interactome remains tiny — fewer than two thousand validated pairs across all secretion system types. General interaction predictors transfer poorly to this setting, and the few host-pathogen-specific tools are mostly locked to a single pathogen.

HPInet, developed by Yueming Hu and colleagues in Ming Chen's group at Zhejiang University with collaborators at Shenzhen University and posted to bioRxiv in August 2025, is a binary classifier for bacterial effector-human protein pairs built on frozen ESM-2 embeddings. Its architectural contribution is modest but deliberate: convolutional layers and global response normalization are inserted into the feed-forward block of an otherwise standard transformer, sharpening extraction of local sequence motifs that matter for interfaces.

The trained checkpoint is deployed as a free public web server, so the model can be applied to new sequence pairs without local installation or retraining. The self-attention weights and gradient-based attributions also expose which residues drove each prediction, giving a sequence-only route to candidate interaction sites.

#Key Features

• Effector-host specialization: Trained specifically on interactions between secreted bacterial effectors (T1SE through T10SE) and human proteins, a domain that general PPI predictors handle poorly.
• Siamese transformer with local feature extraction: Convolutional layers plus global response normalization inside the transformer feed-forward network improve capture of local sequence patterns over a vanilla transformer block.
• Residue-level interpretability: Attention scores and SHAP GradientShap attributions nominate interaction-site residues from sequence alone, without docking or a structure model.
• Secretion-system-specific variants: Alongside the general model, dedicated checkpoints for T3SE, T4SE, T5SE, and T7SE interactions are trained and served.
• Class-imbalance handling: Focal loss with a 1:10 positive-to-negative ratio and dissimilarity-based negative sampling address the realistic scarcity of true interactions.

#Technical Details

The positive set comprises 1,855 experimentally validated effector-human PPIs from HPIDB, MINT, IntAct, PHISTO, and literature, reduced to 1,792 unique pairs and then to 1,552 non-redundant pairs after CD-HIT clustering at 40% identity with cluster-crossing pairs removed. Negatives were drawn from 10,059 human-bacteria interactions by randomly pairing effectors with human proteins and discarding any pair above 30% sequence similarity to a known interaction, held at a 1:10 positive-to-negative ratio. The data was split 9:1 into training and independent test sets with balanced class distributions. Six input representations were compared — ESM-2 650M, ProtBert, ProtT5-XL-U50, PSSM, BLOSUM62, and one-hot — all used as frozen features with no fine-tuning of the backbone. Training used AdamW with a cosine warm-up schedule, initial learning rate 6e-5, and focal loss with gamma 2.

ESM-2 embeddings performed best across every metric. On the independent test set HPInet reached accuracy 0.882, F1 0.532, and AUPRC 0.522; against deepHPI (Faster), the only cross-species comparator, accuracy rose from 0.502 to 0.891, F1 from 0.161 to 0.543, and AUPRC from 0.095 to 0.522. HPInet beat deepHPI on every secretion system type except T2SE. For interpretability the authors docked pairs with AlphaFold 3 and used PDBePISA-derived interfaces as ground truth: HPInet's nominated residues correlate positively with AlphaFold 3 ipTM (GradientShap R = 0.35, p = 2.94e-20; attention R = 0.22, p = 5.80e-09). Notably, of all 1,792 pairs docked, only 17 met AlphaFold 3's pTM > 0.5 and ipTM > 0.8 reliability thresholds, underscoring how poorly structure-based docking handles this class of interaction. Benchmark numbers come from the authors' own curated dataset rather than an external suite. The preprint carries a CC BY-NC-ND license; no source code or weights are distributed, and inference is available only through the hosted server.

#Applications

Microbiologists and infection biologists can use HPInet to prioritize candidate host targets for a newly characterized bacterial effector, or to screen a pathogen's predicted secretome against human proteins before committing to pulldowns or two-hybrid assays. The web server accepts FASTA input for both partners, offers the general or a secretion-system-specific model, returns downloadable CSV results with an interactive network view, and reports importance scores, transmembrane helices, signal peptides, and subcellular localization to help judge whether a hit is plausible in vivo.

#Impact

HPInet's value lies less in architecture than in showing that a frozen protein language model plus a modest task-specific head outperforms both structure-based docking and prior host-pathogen tools on effector-host interactions — a domain where data is scarce, interfaces are unusual, and AlphaFold 3 produced confident predictions for under 1% of pairs. Its reach is bounded: performance rests on one curated benchmark restricted to Gram-negative effectors and human proteins, absolute F1 remains near 0.53 under realistic class imbalance, and with no released code or weights the fixed public server is the only way to run it.