TL;DR
This paper introduces STRIFE, a novel deep generative model for fragment elaboration that significantly improves drug discovery by incorporating protein target-specific information. Unlike previous methods, STRIFE leverages fragment hotspot maps (FHMs) to guide the generation of elaborations that are more likely to bind effectively, outperforming existing structure-unaware approaches in generating highly ligand-efficient molecules.
What Problem Does This Paper Address?
Fragment-based drug discovery (FBDD) is a powerful approach where small molecular fragments are used as starting points to build high-affinity drug candidates. A key step in FBDD is "elaboration," where functional groups are added to a fragment to improve its binding to the target protein. However, current computational methods for fragment elaboration often struggle to account for the specific structural environment of the protein's binding pocket or a user's design intentions. This can lead to the generation of elaborations that are not optimal in terms of size, shape, or interaction potential, making the drug discovery process less efficient. The authors aim to bridge this gap by developing a generative model that can intelligently use protein structural information to guide fragment elaboration.
Methods
The core of STRIFE lies in its ability to extract and utilize "fragment hotspot maps" (FHMs). These maps are generated by analyzing the protein's binding pocket to identify regions that are likely to contribute positively to binding affinity. STRIFE uses these FHMs to define "pharmacophoric points," which represent desirable locations for specific functional groups (like hydrogen bond donors or acceptors) within the binding pocket.
The STRIFE algorithm operates in two phases:
- Exploration Phase: A coarse-grained pharmacophoric profile derived from the FHMs is used to generate a broad range of potential elaborations. These are then filtered and docked to the protein.
- Refinement Phase: The most promising elaborations from the exploration phase ("quasi-actives") are used to derive a more fine-grained pharmacophoric profile. This profile guides the generation of a second set of elaborations, which are then ranked by predicted ligand efficiency.
A key innovation is STRIFE's customizability. Users can either let the model automatically extract pharmacophoric information from FHMs or manually specify their own design hypotheses by defining pharmacophoric points, allowing for exploration of specific structural movements or interactions.
Datasets
- Training Data: A dataset of approximately 427,000 fragment-ligand pairs derived from a subset of the ZINC database was used.
- Evaluation Data: The CASF-2016 dataset was used for large-scale evaluation. The authors filtered this dataset to create a test set of 101 examples where the ground truth molecule was suitable for fragment elaboration and FHMs could be reliably generated.
Key Insights and Ideas
- Protein Structure Matters: Explicitly incorporating protein structural information (via FHMs) into generative models for fragment elaboration leads to significantly better results than structure-unaware approaches.
- Interpretable Pharmacophores: FHMs provide an interpretable way to understand where functional groups are likely to be beneficial, making the model's suggestions more understandable to chemists.
- User-Guided Design: The ability to manually specify pharmacophoric points allows STRIFE to explore user-defined design hypotheses, such as potential side-chain movements, which is crucial for real-world drug discovery.
- Ligand Efficiency is Key: Ranking elaborations by ligand efficiency (docking score divided by heavy atoms) is a more robust metric for assessing potential drug candidates than raw docking scores, especially when comparing molecules of different sizes.
Key Results
- Improved Ligand Efficiency: STRIFE significantly outperformed baseline methods (Scaffold-Decorator, CReM, DeepFrag, STRIFENR) in generating elaborations with higher ligand efficiency, as measured by ASLE (average standardized ligand efficiency improvement). For example, STRIFE achieved an ASLE20 of 0.512, compared to 0.1 for Scaffold-Decorator and -0.029 for CReM.
- Novelty: Nearly half (49%) of the elaborations proposed by STRIFE were novel, meaning they were not present in the training set, indicating good generalization capabilities.
- Case Studies: STRIFE successfully generated elaborations that were strikingly similar to known potent inhibitors in two real-world FBDD case studies, including one involving exploring side-chain flexibility.
Limitations and Future Work
- Simultaneous Pharmacophore Satisfaction: The default implementation of STRIFE does not simultaneously satisfy multiple pharmacophoric points within a single elaboration, which might limit its ability to generate highly efficient molecules in certain scenarios. However, the authors note that this can be addressed by manually specifying points.
- Computational Cost: While the upfront cost of FHM calculation is moderate, the docking of generated molecules for ligand efficiency estimation is computationally intensive.
- Fragment Prioritization: The model requires a user to specify the fragment and exit vector. Future work could involve developing a scheme to prioritize promising fragment-exit vector pairs for more efficient resource allocation.
- Broader Pharmacophore Features: The framework could be expanded to explicitly account for other properties like hydrophobicity and aromaticity, beyond hydrogen bond donors and acceptors.
- Alternative Pharmacophore Fields: The authors suggest exploring alternative pharmacophore interaction fields (e.g., GRAILS, T2F) for extracting structural information.
Summary updated