Abstract
Despite recent interest in deep generative models for scaffold elaboration, their applicability to fragment-to-lead campaigns has so far been limited. This is primarily due to their inability to account for local protein structure or a user’s design hypothesis. We propose a novel method for fragment elaboration, STRIFE, that overcomes these issues. STRIFE takes as input Fragment Hotspot Maps (FHMs) extracted from a protein target and processes them to provide meaningful and interpretable structural information to its generative model, which in turn is able to rapidly generate elaborations with complementary pharmacophores to the protein. In a large-scale evaluation, STRIFE outperforms existing, structure-unaware, fragment elaboration methods in proposing highly ligand-efficient elaborations. In addition to automatically extracting pharmacophoric information from a protein target’s FHM, STRIFE optionally allows the user to specify their own design hypotheses.
Introduction
Fragment-Based Drug Discovery (FBDD) is a powerful strategy in modern medicinal chemistry, yet the transition from a weakly binding fragment to a high-affinity lead remains a manual, labor-intensive process. While deep generative models have shown promise in automating chemical design, most current models are "structure-unaware." They excel at generating chemically valid molecules but often ignore the specific geometric and chemical constraints of the protein binding pocket, leading to designs that may be physically incompatible with the target.
Methodology
STRIFE (STructure Informed Fragment Elaboration) bridges this gap by integrating Fragment Hotspot Maps (FHMs) into a generative framework. The process begins by calculating FHMs to identify regions within the protein pocket that are energetically favorable for specific atom types, such as hydrogen bond donors or acceptors.
These maps are then distilled into discrete 3D pharmacophoric points. STRIFE utilizes a two-phase generation approach: an exploration phase that uses coarse-grained profiles to sample a wide variety of chemical candidates, followed by a refinement phase. In the refinement phase, the model uses "quasi-active" candidates from the initial pool to derive fine-grained profiles, allowing it to home in on precise structural matches that satisfy the protein's requirements.
Results and Discussion
In large-scale evaluations using the CASF-2016 dataset, STRIFE demonstrated superior performance compared to structure-unaware baselines like Scaffold-Decorator and CReM. A key metric was the Standardised Ligand Efficiency Improvement (ΔSLE), where STRIFE significantly outperformed all baselines, indicating that its designs were better optimized for the pocket relative to their molecular size.
Furthermore, STRIFE achieved a novelty rate of 49%, compared to just 4% for Scaffold-Decorator. This suggests that by following structural cues, the model effectively explores new chemical space rather than simply memorizing its training data. Case studies, such as the elaboration of fragments against N-myristoyltransferase (NMT), confirmed that STRIFE can recover known potent motifs and handle complex design hypotheses, including those involving protein side-chain flexibility.
Conclusion
STRIFE represents a significant advancement in making generative models practical for fragment-to-lead campaigns. By providing an interpretable bridge between protein structure and chemical generation, it allows researchers to incorporate specific design hypotheses into the automated design loop. While current implementations focus on single-point interactions, future iterations of the framework aim to satisfy multiple pharmacophoric points simultaneously and automate exit vector selection, further streamlining the drug discovery process.
Summary updated