Abstract
In structure-based drug design (SBDD), the "lock and key" model often fails to account for the dynamic nature of proteins. Most existing AI models assume a rigid protein pocket, which can lead to suboptimal ligand design when starting from unbound (apo) structures. We introduce Apo2Mol, a full-atom hierarchical graph-based diffusion model that jointly generates 3D ligands and predicts the "induced fit" rearrangements of the protein pocket. Trained on a curated dataset of over 24,000 experimental apo-holo pairs, Apo2Mol demonstrates superior performance in generating high-affinity ligands and realistic protein conformations compared to existing rigid-pocket models.
The Challenge of Protein Flexibility
Traditional SBDD typically relies on holo structures—proteins already bound to a ligand—to guide the design of new chemical entities. However, in novel drug discovery, researchers often only possess the apo (unbound) structure. Upon binding, proteins frequently undergo significant conformational changes to accommodate the ligand, a process known as "induced fit."
Designing a drug for a static apo pocket often misses the optimal binding pose that only becomes available after these structural shifts. While Molecular Dynamics (MD) simulations can model flexibility, they are computationally intensive and sensitive to force-field inaccuracies.
Methodology: Joint Diffusion for Induced Fit
Apo2Mol addresses this by treating the ligand and the protein pocket as a co-evolving pair within a diffusion framework. The model employs a full-atom hierarchical SE(3)-equivariant Graph Neural Network (GNN) to process structural information.
Joint Diffusion Process
The core innovation is a coupled diffusion process. During training, the model gradually adds noise to a ligand's atomic coordinates while simultaneously interpolating the protein pocket from its bound (holo) state back to its unbound (apo) state.
In the reverse (generative) phase, the model starts with a cloud of random atoms within an apo pocket. It then iteratively denoises the atoms into a valid molecule while "refining" the pocket into the predicted holo conformation.
Hierarchical Modeling and Parameterization
To maintain physical plausibility, Apo2Mol operates at two levels:
- Atomic Level: Captures precise chemical interactions between the ligand and the pocket.
- Residue Level: Models the protein's backbone and side-chain movements using translations, rotations (represented by quaternions), and side-chain torsion angles ($\chi$ angles).
Experimental Setup and Datasets
The researchers developed a high-quality benchmark using the PLINDER resource, extracting over 24,600 experimental apo-holo pairs from the Protein Data Bank (PDB). This dataset is significantly larger than previous benchmarks like ApoBind and is filtered for high resolution ($\le$ 2.5Å) and 100% sequence identity to ensure that observed movements are strictly due to ligand binding.
Results and Discussion
Apo2Mol was evaluated against state-of-the-art baselines, including IPDiff and TargetDiff. Key findings include:
- Superior Affinity: Apo2Mol generated ligands with an average Vina min score of -6.79, outperforming the best baseline (-6.40).
- Higher Success Rate: The model achieved a 42.7% "High Affinity" rate, significantly higher than the 30-34% range of competing models.
- Chemical Realism: Generated molecules exhibited realistic bond lengths and geometries, indicating a deep understanding of 3D chemical space.
- Pocket Accuracy: The predicted holo conformations closely matched experimental RMSD distributions, capturing the nuances of induced fit.
Conclusion
Apo2Mol represents a significant shift in AI-driven drug discovery by moving beyond the rigid-pocket assumption. By learning the dynamic relationship between ligands and proteins directly from experimental data, it provides a more realistic and effective tool for designing drugs against novel targets where only unbound structures are available.
Summary updated