Cross-species protein-protein interaction predictor: an MLP on frozen ESM-2 embeddings, trained on 200 animals, applied to fungi, plants and bacteria.
No providers recorded yet. Browse all providers
Interactome maps are lopsided. Human, yeast and a handful of model organisms have been screened exhaustively for protein-protein interactions (PPIs), while most species have barely been touched, and experimental screens are too slow and costly to close that gap one organism at a time. ANIMA (Artificial Neural Interaction Model for Animals) takes the transfer route: learn what an interacting pair looks like from species where the data are abundant, then apply that fixed model to species where they are not.
The mechanism is deliberately simple. Each protein is embedded once with the 150M-parameter ESM-2 protein language model, mean-pooled to a single 640-dimensional vector, and the two vectors of a candidate pair are concatenated and scored by a small multilayer perceptron. The authors' argument is that the generalization comes less from classifier depth than from how the training species are chosen: 200 animals picked to be maximally spread across the phylogenetic tree, so the classifier sees as much evolutionary diversity as the data allow.
ANIMA was developed by Bruno Florentino and colleagues at the University of São Paulo, with Bielefeld University and the Federal University of Technology – Paraná, and posted to bioRxiv in September 2026 as a preprint that has not yet been peer reviewed.
Positives come from STRING version 12, restricted to proteins of at most 1,000 residues and to interactions with a STRING confidence of at least 0.90 for training (0.98 for testing). Negatives are random pairs of proteins not reported to interact, sampled to match the number of positives, so some undiscovered true interactions will be labelled negative in both training and evaluation. The training set holds about 1.8 million positive and 1.8 million negative pairs spanning roughly 625,000 proteins. Each pair becomes a standardized 1,280-dimensional vector, and the released classifier is a four-layer MLP (1,024, 512 and 128 hidden units with ReLU and dropout) of about 1.9 million parameters, with hyperparameters chosen by random search over 150 configurations under 3-fold cross-validation on the training data.
Test sets cover 261 unseen animal species, 663 fungi, 128 plants, 65 other eukaryotes, 457 archaea and 1,000 bacteria. Accuracy is 95.3% on held-out animals, 91.1% on other eukaryotes and 82.5% on bacteria and archaea, and within animals it falls from 96% for pairs of well-represented proteins to 85% when both proteins lack close training counterparts. On animal subsets of 5,000 interactions per representation-score class, ANIMA outperformed D-SCRIPT, Topsy-Turvy, TUnA, DeepNano-seq, PLM-interact and SENSE-PPI in accuracy in every class. All reported results are computational.
ANIMA is aimed at researchers working on species with sparse interaction data — wildlife, livestock, parasites, non-model eukaryotes — who need a first-pass interactome or a ranked list of candidate partners to prioritize for yeast two-hybrid, co-immunoprecipitation or proximity-labelling experiments. The representation score gives users a way to judge how far a given query sits from the training distribution before trusting a prediction.
ANIMA adds to a line of sequence-only, cross-species PPI predictors, and its main contribution is a counterpoint to the trend toward ever larger architectures: a frozen mid-sized protein language model plus a small MLP, trained on a phylogenetically curated species set, outperforms more elaborate models in the authors' benchmarks. Its stratified evaluation by evolutionary distance and by training-set similarity is a more informative reporting format than a single aggregate accuracy. Performance on bacteria and archaea is markedly lower than on eukaryotes, and the authors plan training sets for non-animal species.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.