
gPPIpred
Matiolli CC†, Marques J†, Abreu IA. gPPIpred: A User-Friendly PPI Predictor Based on Protein Molecular Graphs. (2026). DOI: 10.17912/micropub.biology.001796.
Open article →The strongest machine-learning work in this trajectory doesn't flatten biology into arbitrary spreadsheet rows and columns. It builds the relationships directly into how the model "sees" the data — as protein shapes, interaction networks, and organized biological categories.
Years after running Y2H experiments by hand in the lab, gPPIpred answers the same kind of question — will these two proteins interact? — computationally instead. Each protein is represented as a 3D graph, where the building blocks (amino acids) are dots and physical closeness between them is the connecting lines. A machine-learning model then compares two of these protein graphs side by side and predicts whether they'll stick together, with attention-based views that can guide inspection. These computational attributions do not establish a binding mechanism or experimentally confirmed interface.
The publication reports an independent validation set of 72,358 protein pairs, a Matthews correlation coefficient (MCC) of 0.4641, and 96% sensitivity at the selected threshold. These results belong to the paper's evaluation protocol; MCC is not an accuracy percentage, and the numbers do not establish performance on every new biological setting.

Matiolli CC†, Marques J†, Abreu IA. gPPIpred: A User-Friendly PPI Predictor Based on Protein Molecular Graphs. (2026). DOI: 10.17912/micropub.biology.001796.
Open article →Aid2Go connects several kinds of protein information — ESM-2 sequence representations, a protein-similarity graph, the Gene Ontology hierarchy, and experimentally supported function annotations — into a single network. A protein's amino-acid sequence becomes just one input feeding that network, instead of being asked to explain everything on its own.