ppIRIS

Bacterial interactomes are far less completely mapped than eukaryotic ones, and the gap is widest exactly where it matters most for infection biology: secreted and surface-anchored effectors, and the cross-species contacts a pathogen makes inside host fluids like plasma. Affinity purification mass spectrometry covers only a fraction of this space, while most sequence-based predictors are tuned to eukaryotic training sets or are too computationally heavy to sweep a whole proteome.

ppIRIS (protein-protein Interaction Regression via Iterative Siamese networks) is a deliberately lightweight answer to that problem, developed at Université de Lorraine with CNRS, Inria, and LORIA, alongside Lund University's Division of Infection Medicine, and posted to bioRxiv in September 2025. Rather than building another large backbone, ppIRIS is a compact interaction head over frozen per-sequence embeddings from two existing protein language models: ESM Cambrian at 300M parameters for evolutionary context, and ProstT5's 3Di track for structure-aware summaries. Both proteins in a candidate pair are pooled to a single vector, so scoring never touches residue-level attention.

The design also addresses a specific obstacle in host-pathogen prediction: a model trained only on intra-species interactions tends to learn species identity rather than interaction propensity. ppIRIS counters this with a pooling-based domain shift, applying different ProstT5 pooling schemes to bacterial and human proteins during training and unifying them at inference, pushing the head toward cues that survive the swap.

#Key Features

• Dual evolutionary and structural embeddings: Concatenating ESM-C and ProstT5 3Di representations proved complementary rather than redundant — ESM-C alone maximized precision, while adding ProstT5 raised recall, F1, and MCC.
• Sequence-level cross-attention: A two-head attention layer over the pooled pair acts as a parameter-efficient learned bilinear form, adapting to the heterogeneous subspaces the two encoders contribute without the cost of a residue-level cross-encoder.
• Proteome-scale throughput: An exhaustive all-versus-all scan of the Streptococcus pyogenes M1 proteome completed in under two minutes.
• Cross-species transfer without cross-species labels: Trained only on human-human and bacteria-bacteria pairs, the model recovered 56.2% of a curated GAS-human plasma interactome.
• Prospective wet-lab validation: Two AP-MS assays against plasma, saliva, and GAS protein fractions confirmed 53% and 50% of the model's nominated pairs.

#Technical Details

Each Siamese branch applies a shared MLP encoder to a 1,984-dimensional fused embedding (attention-pooled ESM-C concatenated with a ProstT5 3Di summary), stepping through two fully connected blocks with batch normalization, ReLU, and dropout of 0.5 down to a 256-dimensional representation. Six vectors — the two encodings, their two cross-attention outputs, their absolute difference, and their element-wise product — are concatenated and passed to a combiner MLP ending in a spectrally normalized linear unit that emits a single interaction logit. Training used Adam with weight decay, binary cross-entropy loss, and gradient clipping.

Positives came from STRING v12 gram-positive bacterial interactions filtered at experimental score above 600 and combined score above 990, across 794 taxa curated through BacDive; negatives were randomly sampled undocumented pairs at 5:1 and 10:1 ratios, with 5:1 performing substantially better. MMseqs2 clustering enforced a 40% sequence-identity ceiling between training and validation proteins. On the Bernett et al. human gold-standard benchmark run under the Reim et al. evaluation framework, ppIRIS reached accuracy 0.661, precision 0.675, recall 0.620, and F1 0.647, ahead of TUnA (0.645/0.672/0.580/0.622), D-SCRIPT with ESM-2 embeddings, and Richoux-ESM-2. On the D-SCRIPT cross-species sets it obtained the highest AUROC and AUPR for D. melanogaster, C. elegans, S. cerevisiae, and E. coli. On the host-pathogen panel of 16 GAS virulence factors against 62 plasma proteins, a 0.5 threshold recovered 63 of 112 AP-MS-supported interactions at precision 0.163 and F1 0.253.

#Applications

ppIRIS is built for triage rather than exhaustive enumeration: it narrows an intractable search space to a shortlist worth spending mass spectrometry time on. In the paper this surfaced under-annotated GAS M1 clusters spanning two-component signaling, phage response, metal scavenging, and peptidoglycan synthesis; candidates were then annotated with DeepGO and modeled with AlphaFold-Multimer for structural plausibility. Groups studying bacterial pathogens or host immune evasion can run the released checkpoint on a new proteome without retraining.

#Impact

ppIRIS is a preprint awaiting peer review, and its scope is narrower than its throughput suggests: it is an interaction head, not a standalone backbone, so predictions inherit the coverage biases of the upstream language models, and embedding extraction is bound by ESM-C's separate license terms even though the ppIRIS code and shipped checkpoint are Apache-2.0. The authors are candid about failure modes — precision stays modest under realistic class imbalance, apolipoproteins and metabolic enzymes were poorly recovered because they are underrepresented in training, and randomly sampled pseudo-negatives carry label noise. Its contribution is a practical demonstration, alongside models such as PLMDA-PPI, that a thin head over frozen protein language model embeddings can match heavier residue-level predictors while staying cheap enough for iterative experimental cycles.