On the comprehensive FoldBench benchmark, IntFold’s performance is comparable to AlphaFold 3 and significantly exceeds other contemporary models (Boltz-1, 2, Chai-1, HelixFold 3 and Protenix) across a diverse range of biomolecular interaction tasks.
Use it on Tamarind Bio today:
In antibody-antigen (Ab-Ag) complexes, commercially useable alternative lagged significantly behind AlphaFold3, until now.
Performance Results
IntFold achieves performance comparable to AlphaFold 3 on the comprehensive FoldBench benchmark:
Protein monomers: 0.88 LDDT score (matching AlphaFold 3)
Protein-protein interactions: 72.9% success rate (matching AlphaFold 3)
Significantly outperforms other contemporary models like Boltz-1/2, Chai-1, HelixFold 3, and Protenix
Antibody-Antigen Complexes
A particularly challenging area where IntFold shows strong improvement:
Base model: 37.6% success rate
Enhanced IntFold+: 43.2% success rate (approaching AlphaFold 3's 47.9%)
Protein-Ligand Interactions
Base model: 58.5% success rate
IntFold+: 61.8% success rate (vs AlphaFold 3's 64.9%)
Structural Constraints: Dramatic improvements when incorporating known binding sites:
PoseBusters dataset: 79.5% → 89.7% success rate
Antibody-antigen interfaces: 37.6% → 69.0% success rate
Binding Affinity Prediction: Outperformed existing methods on DAVIS and BindingDB benchmarks
Technical Advances
Custom FlashAttentionPairBias kernel: Faster and more memory-efficient than standard industry implementations
Model-agnostic ranking method: Training-free approach that improves success rates by ~3% through structural similarity consensus
Practical Impact
IntFold demonstrates the potential for controllable foundation models in drug discovery, successfully addressing limitations of general-purpose models by enabling:
Prediction of functionally critical conformational states
Integration of prior structural knowledge
Accurate binding affinity estimation for virtual screening