IsoDDE: Generalization Leap in Biomolecular Interaction Prediction

Original: Isomorphic Labs Technical Report (February 2026)

1. Research Background

1.1 Post-AlphaFold 3 Structure Prediction Landscape

The release of AlphaFold 3 established deep learning's dominance in biomolecular structure prediction. Since then, various AlphaFold-style structure models have emerged (such as Chai-1, Boltz-1/2, Protenix), mainly improving model performance through architecture adjustments, conditional control features, and inference-time guidance mechanisms.

1.2 Core Challenges in Generalization

Despite recent significant progress, benchmark tests reveal models' continued failures in generalizing to unexplored molecular spaces:

• Antibody-antigen docking failure rates exceed 50% (even for AlphaFold 3)
• Current co-folding models mostly memorize small molecule binding patterns, leading to performance collapse on novel pockets
• Accurate structure predictions often fail to translate into quantitative interaction strength metrics

1.3 Limitations of Existing Methods

Traditional molecular docking tools rely on experimentally determined protein structures as input and cannot truly achieve de novo prediction. More importantly, existing models show significantly degraded performance when handling out-of-distribution data, limiting their applications in first-in-class target discovery and novel regulatory mechanism research.

2. Technical Architecture Overview

IsoDDE's architectural improvements represent the first substantial accuracy leap in this field since AlphaFold 3's release.

2.1 Architecture Optimization

The model retains AlphaFold 3's core elements—triangular operations acting on pair representations—but makes several optimizations:

• Improvements to triangular attention
• Signal flow rearrangement to improve information transfer between single and pair representations

2.2 Introduction of Physical Priors in Generative Modeling

The model re-examines the mathematical formulation of generative modeling, introducing physical prior knowledge. This includes improvements to the diffusion process and joint optimization of local structures and global conformations at different noise levels.

2.3 Inference-Time Guidance Mechanism

Similar to recent other models (such as Boltz-2), IsoDDE also has inference-time guidance capabilities for correcting physical inaccuracies (such as incorrect chirality). However, unlike these models, IsoDDE's unconditional prediction accuracy itself has been significantly improved.

3. Performance Evaluation

The following evaluation results are based on benchmark tests reported in the paper. IsoDDE uses the same training data cutoff date as AlphaFold 3 (September 30, 2021), ensuring fair comparison.

3.1 Protein-Ligand Structure Prediction

On the Runs N' Poses benchmark (used to evaluate co-folding models' generalization to samples far from the training set), IsoDDE demonstrates significant performance improvements:

• In the most difficult similarity range (0-20%), IsoDDE achieves 50% success rate, more than 2× improvement over AlphaFold 3
• 17 of these successful predictions were cases AlphaFold 3 failed to predict correctly

3.2 Antibody-Antigen Interface Prediction

On a low-homology antibody-antigen test set containing 334 structures:

• High-fidelity range (DockQ > 0.8): IsoDDE achieves 39%, 2.3× improvement over AlphaFold 3 (17%), and 19.8× improvement over Boltz-2 (2%)
• Correct prediction range (DockQ > 0.23): IsoDDE achieves 63% success rate with single seed
• With 1000 model seeds, high-fidelity prediction rate reaches 59%

3.3 CDR-H3 Loop Prediction

The CDR-H3 loop is the main source of antibody variability. IsoDDE predicts CDR-H3 loops for 70% of antibodies with ≤2Å backbone RMSD, 1.2× improvement over AlphaFold 3 (58%), and 1.6× improvement over Boltz-2 (43%).

3.4 Binding Affinity Prediction

In predictions of newly released activity data from ChEMBL 35, IsoDDE maintains high performance across target categories. On a curated dataset suitable for physics-based simulation, its affinity prediction surpasses gold-standard free energy perturbation methods such as FEP+.

3.5 Novel Pocket Identification

On general test sets, IsoDDE's pocket identification capability AUPRC is 1.5× better than P2Rank. This capability enables the model to discover novel pockets previously identifiable only through experimental methods.

4. Typical Cases

4.1 Protein-Ligand Cases

PDB 8EA6: Ligand opens a new cryptic pocket at the interface of two NKG2D protein chains, forming a ternary complex. The ligand acts as an allosteric protein-protein interaction inhibitor.

PDB 8E23: Allosteric inhibitor of Pol θ causes helix movement, opening a binding site in a pocket not seen in the training set.

4.2 Antibody-Antigen Cases

PDB 9FZD: Nanobody binding to bacterial outer membrane protein A (OmpA). No OmpA-nanobody complex structures in training data.

• IsoDDE prediction: DockQ 0.943, CDR-H3 backbone RMSD 0.94Å
• AlphaFold 3 prediction: DockQ 0.00, CDR-H3 RMSD 4.98Å (nanobody positioned on wrong side of OmpA)

PDB 8Q3J: Antibody Fv binding to mouse IL-38.

• IsoDDE heavy chain-antigen interface: DockQ 0.876, CDR-H3 RMSD 0.78Å
• AlphaFold 3 flips heavy and light chain orientations: DockQ 0.060

5. Limitations and Discussion

5.1 Impact of Training Data Cutoff Date

IsoDDE uses the same training data cutoff date as AlphaFold 3 (September 30, 2021), ensuring fair comparison, but also means the model cannot utilize structure data published after this date.

5.2 Computational Resource Requirements

Antibody-antigen prediction with 1000 seeds can significantly improve performance, but this brings corresponding computational overhead. For industrial-scale drug discovery workflows, the feasibility of this computational cost needs further evaluation.

5.3 Real-World Validation Needs

There may be gaps between benchmark performance and value in actual drug projects. Additionally, the model's performance in stereochemistry (chirality, atom clashes) requires more independent evaluation.

5.4 Competitive Landscape

Open-source models (such as Boltz-2, Chai-1) continue to improve in controllability and efficiency. Although their current unconditional accuracy does not match IsoDDE, their openness and rapid iteration capabilities may narrow the gap.

6. Conclusion

IsoDDE has achieved substantial progress in the generalization capability of biomolecular structure prediction. In multiple tasks including protein-ligand, antibody-antigen, and novel pocket identification, its performance shows significant improvement over existing methods, with some capabilities (such as affinity prediction) potentially surpassing traditional physics-based methods under specific conditions.

References:
[1] Isomorphic Labs Team. Accurate Predictions of Novel Biomolecular Interactions with IsoDDE. Technical Report, February 2026.