- Protein-ligand structure prediction: Achieves more than twice the accuracy of AlphaFold 3. Records a 50% success rate (AF3 23.3%) even for novel protein pocket/ligand combinations with extremely low similarity (0-20%) to the training data. Models complex structural changes such as induced fit and cryptic pocket opening. Accurately predicts the cryptic pocket structure of the NKG2D homodimer, a case where AF3 failed. - Antibody-antigen interface modeling: Achieves a 39% rate of high-precision predictions (DockQ > 0.8) among 334 low-similarity protein pairs. This is 2.3 times better than AlphaFold 3 (17%) and 19.8 times better than Boltz-2 (2%). Achieves 70% accuracy (RMSD ≤ 2Å) for the CDR-H3 loop, which is the most variable and difficult to predict in antibodies, suggesting the possibility of de novo antibody design. - Binding affinity prediction: Achieves the best performance among deep learning-based methods in three public benchmarks: FEP+ 4, OpenFE, and CASP16. Achieves accuracy equal to or better than the physics-based gold standard, the FEP+ method, without experimental crystal structures (FEP+ 4 benchmark r=0.85 vs FEP+ 0.78). Reduces computation time and cost to a fraction of the original. - Novel binding pocket identification: Identifies new pockets that ligands can bind to based solely on amino acid sequences, within seconds. Significantly reduces time and cost compared to experimental fragment soaking. Successfully predicts all known pockets and recently discovered novel allosteric sites in the Cereblon protein based solely on sequence. Achieves accuracy close to experimental techniques at computer speed (seconds). - Multimodal drug design: Supports the generation of candidates across various drug modalities, including small molecules, antibodies, peptides, and molecular glues (confirmed in collaboration with J&J).