New AI Model Predicts Protein-Protein Interactions with Unprecedented Accuracy
Researchers report an AI system that predicts how proteins interact with each other — a harder problem than single-protein structure prediction — with accuracy levels that could accelerate cancer biology and drug discovery research.
A new AI model that predicts how pairs of proteins interact has been reported in Phys.org, representing a meaningful advance over single-protein structure prediction tools like AlphaFold.
Why protein-protein interactions are harder
AlphaFold 2 (2021) and AlphaFold 3 (2024) solved the protein folding problem — predicting the three-dimensional shape of a single protein from its amino acid sequence — with remarkable accuracy. But many of the most medically important questions in biology involve how proteins interact with each other:
- Which proteins bind to which targets in a signalling pathway?
- How does a mutation in one protein disrupt its interaction partner?
- Which drug candidates could disrupt a protein-protein interaction implicated in cancer?
Protein-protein interaction (PPI) prediction is substantially harder than single-protein folding because the interaction surface depends on the conformational changes that occur when both proteins are present simultaneously — a dynamic that is difficult to capture from static structures.
AlphaFold Multimer (released by DeepMind alongside AlphaFold 2) addressed this for pairs of proteins, but performance on novel interaction pairs without close homologs in the training data remained a significant limitation.
What the new model does differently
The model described in the Phys.org report uses a “protein pair reading” approach — rather than predicting structure directly, it learns a joint representation of both proteins and their likely interaction geometry simultaneously. Key reported improvements:
- Higher accuracy on novel PPI pairs with no close structural homologs in the PDB
- Better prediction of interface residues — the specific amino acids that form the contact surface
- Stronger performance on disease-relevant interactions (cancer driver proteins, immune checkpoint complexes)
The research was conducted at an academic medical centre; preprint details were available at time of reporting.
Implications for drug discovery
Protein-protein interactions are increasingly recognised as a target class for drugs. Historically, PPIs were considered “undruggable” because the flat, extended contact surfaces between proteins were thought to be difficult to block with small molecules. Advances in structural understanding of PPI interfaces — accelerated by better AI prediction — have opened new approaches:
- PPI inhibitors that disrupt oncogenic complexes (e.g., BCL-2 family interactions in cancer cell survival)
- Molecular glues and bifunctional degraders (PROTACs) that co-opt PPI surfaces
- Biologics targeting PPI interfaces with antibody fragments
For researchers in cancer biology, immunology, and structural biology, better PPI prediction means faster computational screening of which interactions are tractable targets and which interface geometries to prioritise in fragment-based drug design.
Current tools for PPI prediction
| Tool | Approach | Access |
|---|---|---|
| AlphaFold Multimer | End-to-end structure prediction for complexes | Free (web + local) |
| RoseTTAFold2 | Improved multimer prediction | Open source |
| ESMFold | Fast embedding-based prediction | Free API |
| New model (unnamed) | Joint pair representation | Research preprint |
References
- (2026, April). AI model ‘reads’ protein pairs, unlocking new insights into disease and drug discovery. Phys.org.
- Bryant, P., et al. (2022). Improved prediction of protein-protein interactions using AlphaFold2. Nature Communications, 13, 1265. https://doi.org/10.1038/s41467-022-28865-w
- Evans, R., et al. (2022). Protein complex prediction with AlphaFold-Multimer. bioRxiv. https://doi.org/10.1101/2021.10.04.463034