Exploring protein hotspots by optimized fragment pharmacophores

Fragment-based drug discovery (FBDD) has emerged as a powerful lead generation strategy, focusing on the screening of small, typically low-affinity compounds. Prior research indicates that fragments preferentially bind to protein hotspots, characterized by high contributions to overall binding energy. These hotspots often exhibit conserved binding pharmacophores, suggesting a limited number of distinct interaction patterns. However, challenges remain in FBDD, including concerns about target specificity and the high frequency of inactive fragments in screening campaigns. Some studies highlight the prevalence of “privileged fragments” active against multiple targets, indicating that while specificity can be engineered during fragment optimization, understanding conserved pharmacophores can improve library design. This work aims to address these challenges by designing a minimal fragment library that broadly covers experimentally confirmed binding pharmacophores found at protein hotspots. The hypothesis is that a carefully curated library based on these frequent interactions will be highly efficient in identifying hits across a diverse range of protein targets, improving the overall success rate of FBDD.

The authors reviewed studies on fragment-based drug design, highlighting the preferential binding of fragments to protein hotspots and the importance of understanding binding thermodynamics. They cite existing research demonstrating that fragments bind to hotspots through a limited set of optimal geometric hydrogen bonds, emphasizing the critical role of the nature and quality of these interactions in selecting starting points for drug discovery. The authors also discuss the controversies surrounding fragment-based approaches, such as the frequent observation of inactive fragments and the issue of target specificity. Analysis of large screening campaigns revealed the prevalence of privileged fragments active against multiple targets, emphasizing the need for a more strategic approach to fragment library design. Previous studies on the conservation of both hotspots and pharmacophores across related protein families supported the researchers' assumption that the pharmacophore space can be efficiently represented by a well-defined set.

The methodology involved a multi-step process. First, the researchers extracted experimental fragment-binding modes from the Protein Data Bank (PDB), focusing on fragment-sized ligands (10-16 heavy atoms). FTMap analysis was used to identify hotspot-binding fragments. Schrödinger's ePharmacophore module then extracted pharmacophore models (maximum four features) from each protein-ligand complex, considering energetic contributions to binding. A two-step clustering approach reduced redundancy. First, pharmacophores with identical feature sets were grouped (level 1 clusters). Then, spatial alignment and hierarchical clustering (HCA) with a 2 Å RMSD cutoff yielded non-redundant pharmacophores (level 2 clusters). This resulted in 425 unique pharmacophores. Next, the researchers optimized a small fragment library to maximize coverage of these pharmacophores. Commercial fragment collections were filtered by size, rotatable bonds, and other properties. An algorithm simultaneously minimized pairwise fingerprint similarity (of molecules and pharmacophores) and maximized pharmacophore coverage. This process led to the SpotXplorer0 library, comprising 96 compounds. The library was then screened against established target classes (GPCRs and proteases) using biochemical assays (radioligand binding and chromogenic protease assays), and against challenging targets (SETD2, SARS-CoV-2 3CLPro, and NSP3 macrodomain) using various assays (chemiluminescence-based enzymatic assay, cell-based assay, and X-ray crystallography). The X-ray screening utilized the XChem platform at Diamond Light Source, a high-throughput technique for detecting weak binders. In each case, the percentage of known pharmacophores covered by the identified hits was calculated by comparison with ChEMBL data. Topological features (ring type, hydrogen bond patterns) of identified hits were also compared to those in ChEMBL to assess the representativeness of the library.

The SpotXplorer0 library, comprising 96 compounds, achieved high coverage of experimentally validated 2-point and 3-point binding pharmacophores (76% and 94%, respectively). Screening against GPCRs (5-HT1A, 5-HT6, 5-HT7) and proteases (Factor Xa, thrombin) yielded diverse fragment hits, with minimal overlap between targets, successfully retrieving a majority of known pharmacophores for each target (Table 1). Notably, the relatively small number of hits covered a significant portion (on average 80% for 2-point and 60% for 3-point pharmacophores) of the known pharmacophores present in much larger ChEMBL datasets (Table 1, Fig. 2). Topological analysis (Fig. 3) showed that the identified hits exhibited similar ring types and hydrogen bond distributions to those observed in ChEMBL fragments for the same targets, reinforcing the effectiveness of the library design. Further analysis of thrombin and Factor Xa showed that 86.7% and 85.0% of X-ray-validated pharmacophores were represented by SpotXplorer0 hits. Screening against SETD2 yielded two fragments (SX045, SX084) inhibiting SETD2 activity with IC50 values of 300 and 500 µM, respectively. SX045 also showed anti-proliferative effects in leukemia cells. X-ray crystallography screening against SARS-CoV-2 3CLPro identified SX013, which inhibited 3CLPro activity (IC50 of 31 µM) and showed antiviral effects. Screening against the NSP3 macrodomain revealed five fragments (SX003, SX005, SX048, SX051, SX054) binding within the ADP-ribose binding pocket, some mimicking the natural ligand's interactions and establishing additional interactions (Fig. 4, Fig. 5). These fragments exhibited antiviral activity in cellular assays.

The study's findings validate the effectiveness of the SpotXplorer approach for designing efficient fragment libraries. By focusing on experimentally validated pharmacophores at protein hotspots, the researchers created a small library that successfully identified hits across a diverse range of targets, including those known to be challenging. The high percentage of known pharmacophores covered by the relatively small number of hits demonstrates the efficiency of the approach. The success in identifying hits for recently emerged targets like SARS-CoV-2 3CLPro and NSP3 further underscores the method's broad applicability and its potential to accelerate drug discovery efforts. The approach's success stems from its protein-centric nature, in contrast to conventional ligand-based methods. The identification of novel pharmacophore arrangements in thrombin, not previously observed in the PDB, suggests that the approach can identify new binding poses. The results suggest that by explicitly considering the experimentally validated pharmacophore space, significantly better coverage of the pharmacophore space is achievable when designing small fragment libraries.

This study successfully demonstrated a novel approach (SpotXplorer) for designing fragment libraries in FBDD. The SpotXplorer0 pilot library showcased exceptional coverage of experimentally verified pharmacophores and identified hits for diverse targets, including challenging ones. This protein-centric approach offers a significant advantage over traditional ligand-based methods. Future work could involve expanding the SpotXplorer library, incorporating additional pharmacophores and exploring the identified fragments for optimization into potent drug candidates.

The current set of non-redundant pharmacophores is limited to those observed in publicly available PDB structures. While the library design accounted for several factors, such as protonation states, tautomeric forms, and protein-induced polarization effects, these considerations might not fully encompass the complexity of protein-ligand interactions. The study's success with the SARS-CoV-2 targets might be due to the high-resolution structural information available for these proteins. The applicability of the method to targets with limited structural information might require further investigation.