Fragment Based Drug Discovery, or FBDD for short, is nowadays a well-established and common approach adopted by many pharmaceutical companies and academic groups (Erlanson et al., 2016). The rationale behind FBDD has been extensively reviewed and entire books have been written, including some with references to clinical candidates (Klon, 2015; Rees, 2016). The general concept of FBDD is straightforward; it starts with the generation of libraries of small molecules called “fragments.” The size of these libraries varies from a few hundred to thousands of molecules (for industrial cases) (Singh et al., 2018), usually prepared following the so-called “rule of three,” i.e. a molecular weight smaller than 300Da, lipophilicity as expressed by LogP smaller than 3, and a maximum number of hydrogen bond donors and acceptors less than 3 (Congreve et al., 2003). For efficiency reasons the binding against a target of interest is usually evaluated for each of these molecules in so-called “mixtures” of 5–10 compounds in a single experiment (Mercier and Powers, 2005). The strategy of using small molecules instead of large entities allows for a more efficient exploration of the chemical space, defined as the ensemble of all possible molecular conformations presenting drug-like properties, and estimated to be a staggering ∼10⁶⁰ molecules (Reymond et al., 2010). The FBDD approach provides for a great chemical variety to probe this chemical space and has many other benefits, such as cost and time reduction in the data analysis.

Using fragments as a starting point in the early stages of drug discovery has been demonstrated to be a viable approach for producing compounds that are specifically tailored to their targets (Hughes et al., 2011). This approach also increases the novelty of standard drugs and enables the path of chemical optimisation to be monitored, for example by restricting the lipophilicity issue observed in lead molecules obtained by high-throughput screening (HTS) (Author Anonyms, 2007; Erlanson et al., 2016). However, since the fragments are much smaller compared to traditional lead-like molecules, their binding affinity to a target of interest is nearly always low (μM to mM). Therefore, it requires a technique, such as NMR, that is capable of detecting these weak interactions (Ortega-Roldan et al., 2009; Vinogradova and Qin, 2012).

In this manuscript, we focus exclusively on the relevance and impact of NMR spectroscopy on the generation of new clinical drugs. Through a literature review, we establish the impact of NMR in FBDD. We performed an in-depth assessment of three case studies that establish the impact of various NMR techniques on different stages of drug development. We establish patterns in the drug-development optimization process and formulate proposals for its improvement.

To appreciate the influence on NMR in the current developments, we queried the PubMed database with the “fragment-based drug discovery” and “NMR” keywords. It revealed a steady and significant increase in the number of publications, growing from 145 in 2015 to 292 in 2020, for a total of 1,200 journal publications over this 6-year period (Figure 1A). A further search through the FDA database revealed, over the same time span, that several drugs were unmistakably obtained from fragments (Figure 1B) (Deeks, 2016; The ASCO Post, 2019). Excluding biologics (antibodies, oligonucleotides etc.) and other chemical compounds (diagnostics, combination of old drugs etc.), original small molecules make up a significant proportion, typically ∼40% or more, of the total FDA approvals in any given year (Figure 1B).

From the FDA database analysis, it appeared that only Venetoclax was developed with the help of at least one NMR technique in the early stages of the discovery. Whilst this may seem to be a low number relative to the total number of FDA approvals, this finding is likely explained by the fact that not every breakthrough in the drug development process obtained by NMR will have been shared in the public domain. Furthermore, despite our exhaustive literature search (vide infra), the origin of new molecular entities remained often obscure or very difficult to trace.

The true impact of NMR techniques in the development of original lead-molecules appeared clearer from the inspection of the origins of the clinical compounds. We performed an in-depth analysis of all FBDD-derived molecules that are or have been clinical candidates at any time in the past, up to December 2020, and for which relevant information was publicly accessible through journal articles, tables or from Dan Erlanson’s “Practical Fragments” blog (Erlanson, 2018). Our analysis showed that NMR is used in all three stages of the FBDD process (Figure 1C). For sixteen out of twenty clinical compounds (80%), we traced that NMR was used for identifying the initial binding fragments, the so called “hit identification” stage. For the subsequent FBDD stages, which include the “binding site identification” and “hit optimisation,” NMR was used less often, i.e. in 50 and 40% of the cases, respectively, (Figure 1C), with other techniques, such as X-ray crystallography, being preferred.

NMR-derived compounds were identified mostly by a number of ligand-detected, one-dimensional (1D) NMR techniques, such as Water-Ligand Observed via Gradient Spectroscopy (Water-LOGSY) (Antanasijevic et al., 2014), saturation transfer difference (STD) (Viegas et al., 2011; Meyer et al., 2004), or the so-called T_1ρ-experiment (Hajduk et al., 1997). In contrast, target-based two-dimensional (2D) NMR techniques, such as the chemical shift perturbation (CSP) experiment (Mureddu and Vuister, 2019), were used for the hit and/or binding site validation. Lastly, the so-called SAR (structure activity relationship) by NMR method (Fesik et al., 1997; Shuker et al., 1996), which employs mostly NOE-related techniques and multi-dimensional NMR experiments, was mainly used for hit-growing and linking guidance during the optimisation stage of the FBDD process (Figure 1D) (Petros et al., 2006) (see also Supplementary Table S1).

The 1D ligand-detected techniques are considered a gold standard in NMR screening, as these do not require expensive protein labelling and therefore can be used against a broad range of molecular targets (Ciulli and Abell, 2007). Furthermore, the various expression systems of the target, e.g. bacteria, insects or human-derived cells, and other common limitations, such as molecular weights, are not of critical importance for these 1D techniques (Campos-Olivas, 2011). Moreover, the inherent versatility of NMR has also allowed the detection of the binding activity of small-molecules to receptors in their native environments and in in real-time, a strategy called in-cell NMR (Primikyri et al., 2018). Many common NMR techniques have been used in or adapted for in-cell NMR (Kang, 2019). However, STD and Tr-NOESY techniques (Mari et al., 2010) have been successfully employed without the limiting step of isotopic labelling (Primikyri et al., 2018).

In addition, 1D ligand-detected techniques can also be utilised in difficult cases where expression and/or purification of the target macromolecule is a limiting factor and only nanomolar concentrations can be obtained. Most importantly, the richness of information acquired in a small amount of time (i.e. minutes per sample) allows the analysis to be performed in a high-throughput fashion (Dalvit et al., 2002). However, 1D ligand-detected experiments are not suitable for detecting the binding sites on the target, and higher dimensionality NMR techniques, including chemical-shift perturbation mapping (Williamson, 2013; Mureddu and Vuister, 2019), are often required. The latter enables the monitoring of target residues that are most likely to be interacting with the fragments, providing validation of binding as well as guidance on the next stage of development (Williamson, 2013).

Fragment optimisation is best achieved where a high-resolution 3D molecular structure of the target is available. While there are several techniques capable of resolving molecular structures, the simplicity and the rapid throughput associated with X-ray crystallography, it makes this the preferred method whenever possible (Erlanson, 2015). However, often targets of interest cannot always be exhaustively assessed by X-ray crystallography. For example, complexes displaying a highly flexible mode of interaction with the target molecule are best inspected by NMR (Teague, 2003; Leone et al., 2010), as crystal packing forces preclude the molecular adaptation required for complex formation. Moreover, the crystal lattice also might not allow the ligand to permeate through to the binding pockets (Zheng et al., 2015). In contrast, the NMR technique can provide unambiguous information on the various orientations of the ligand with respect to the target, referred to as poses. These poses can be combined with computational methods for designing drug-like compounds with improved binding and pharmacological properties. Interestingly, the emergence of Artificial Intelligence (AI)-driven structure elucidation, such as AlphaFold-2 (Jumper et al., 2021) and RoseTTA fold (Baek et al., 2021) provides an additional avenue for the fragment-optimization stage. A description of the most recurrent NMR techniques to elucidate structure information is available in the supplementary materials (Supplementary Table S1).

In subsequent sections of this manuscript, we present three case studies that employed a variety of NMR techniques and therefore can be considered templates of NMR-FBDD. The first case study explores the development of a compound denoted as AZD-3839. It originated from fragments identified by a ligand-detected primary screening using the Water-LOGSY technique. The second case study examines the history of the FDA-approved drug ABT-199, commercially called Venotoclax, which was also derived by FBDD. The development of this drug relied on a variety of target-detected NMR methods. Finally, the third case study analyses a compound known as S64315, a recent FDA-approved drug, which illustrates the role of NMR in FBDD through a combination of ligand-detected and target-based NMR techniques.

We explored the development histories of the AZD-3839, ABT-199 and S64315 compounds (Jeppsson et al., 2012; Souers et al., 2013; Szlávik et al., 2019), from the primary screening to the final lead optimisation, focusing on their target interactions and the rationale behind their optimisations. These cases highlighted the underlying role of NMR techniques during all drug discovery phases and their impact throughout each stage.

With multiple compounds in clinical phases, NMR has demonstrated a key role in the process of fragment-based drug discovery (Singh et al., 2018) (Figure 5A). In 2016, ABT-199, commonly known as Venetoclax, was the first confirmed FDA-approved drug derived largely by NMR-FBDD (Erlanson, 2012; Souers et al., 2013; Hubbard, 2016). The development of other fragment-derived drugs, e.g., Vemurafenib (approved in 2011) (Flaherty et al., 2011), Erdafitinib (approved in 2019) (Perera et al., 2017; The ASCO Post, 2019), Pexidartinib (approved in 2020) (Benner et al., 2020) and Sotorasib (approved in 2021) (Ostrem et al., 2013), were driven by both NMR, X-ray crystallography and other techniques.

Interestingly, Vemurafenib, and Pexidartinib, both developed by Plexxikon, the phase-III drug Capivasertib (Addie et al., 2013) developed by Astra-Zeneca, and many other clinical candidate molecules, all include the identical 7-azaindole fragment (Irie and Sawa, 2018; Qhobosheane et al., 2020). This unique structural composition presents both crucial hydrogen-bond acceptor and donor groups making this small molecule able potentially to interact with over 90 different kinds of kinase active sites, which has been considered encapsulating the entire human Kinome (Irie and Sawa, 2018).

In spite of this distinctive case of a generally adaptable building block, over the years, different methodologies have been explicitly developed and proven crucial for enhancing the success rate in the drug discovery process (Pellecchia, 2010). The great flexibility and adaptability of NMR provides for qualitative and quantitative insights at each point of the drug development process (Campos-Olivas, 2011; Pellecchia, 2010). From the case studies, it emerged that NMR is predominately used in the primary screening, also known as “hit-identification” (Figure 1C). The detection of weak binders, defined as small-molecules presenting transient interactions to targets or unfavourably high dissociation rates, is paramount in the early stages of FBDD (Wang et al., 2017). The identification of such compounds represents a clear advantage as it shows its natural ability to bind its target and succeed among cocktails of other sub-molar molecules. The crucial chemical nature of these molecules, i.e., their “core”, is often used as a scaffold and preserved during their evolution to final drugs (Figures 3A, 4A).

However, the process hit-identification by NMR has also shown a number of drawbacks (Davis and Erlanson, 2013). For example, the usage of only a single ligand-detected 1D technique for identifying binding fragments may prove to give erroneous results (Davis and Erlanson, 2013). Hence, the recommendation is to use at least two NMR techniques, such as STD and Water-LOGSY (Davis and Erlanson, 2013) in parallel.

An inevitable consequence of using weakly binding fragments is the so-called “non-specific” binding event (Mercier et al., 2009). A simple strategy for alleviating this issue was employed in the development of AZD-3839. By recording competition experiments using a potent known, but clinically unsuitable, ligand for the same target non-specific binding molecules were identified and excluded from further development.

Furthermore, from the case studies it appeared that the full analytical power of 1D NMR spectroscopy was often not exploited. Instead, the NMR data appear to be used solely as a binary result, probably also due to a lack of proper computational- and data-analysis tools available at the time the research was conducted (Mureddu et al., 2020). The data obtained from Water-LOGSY and STD experiments offer further quantitative information (Meyer et al., 2004; Cala and Krimm, 2015). The SAR by Water-LOGSY, for example, suggests a scoring factor to identify the most exposed portion of the molecule. Assessing all data that can be derived from 1D and multi-dimensional NMR experiments can eventually provide insights into the ligand binding pose (Raingeval et al., 2019).

Upon validation of fragment hits, the next stage is usually the exploration of potential binding sites on the target. Chemical shift perturbation, or CSP, is so far the most popular NMR technique used for this task (Williamson, 2013; Mureddu and Vuister, 2019). CSP has been widely used as a standard for molecules that progressed into clinical phases. Nevertheless, a CSP analysis potentially might drive researchers in wrong directions, and final conclusions should not be based on this approach alone. Common errors observed in practice include subjective judgments or misinterpretation of shifting peaks, especially in crowded regions of the spectra, leading to overestimating of the CSP effects. Furthermore, in some instances, compounds have been shown to change the pH of the solution, resulting in false positive CSPs (Davis and Erlanson, 2013). By performing appropriate control experiments, such errors might hopefully be avoided.

NMR has often been associated with a requirement for daunting and time-consuming data analysis. There may be a multitude of other undescribed factors, but NMR’s lack of modern, more practical, quicker, and unbiased methods, alongside with automated data analysis routines has comparatively slowed the entire NMR-FBDD process, prompting a need for improvements in all these aspects (Mureddu et al., 2020). To this end, we have recently developed a versatile and flexible data-analysis program called AnalysisScreen (Mureddu et al., 2020), part of the CcpNmr Analysis package (Skinner et al., 2016), which presents dedicated provision and capabilities for the analysis of all forms of NMR data used in the NMR-FBDD process, including an integrated CSP analysis (Mureddu and Vuister, 2019).

An additional key step before the expensive and laborious optimisation processes of candidate compounds begins is a minute hit chemical assessment; for example, by employing Pan-Assay Interference Compounds (PAINS) protocols (Baell and Nissink, 2018). Applying these filters to hits or families of hits can help identify erroneous binders. PAINS-flagged molecules can exert photo-reactivity, redox-activity and other undesirable chemical phenomena, which can lead to non-specific or unwanted biological activities. Unfortunately, it is also wise not to rely solely on PAINS filters. A recent analysis showed that many PAINS-flagged molecules had been wrongly evaluated by the applied filters, either as false negatives or false positives (Capuzzi et al., 2017).

Ayotte et al. proposed the use of a NMR CPMG series to detect potential aggregation of compounds in mixtures (Ayotte et al., 2019), another potentially complicating effect. In addition, the authors also pointed out that aggregation can be solvent-dependent, and thus a minor adjustment of the sample composition might improve or worsen the outcomes of screening experiments (Ayotte et al., 2019).

The methodological and data-analysis improvements are fundamental to design the molecules from the early stages and onward. From the analysis of the three cases, it emerged that molecular optimisations are mainly guided by multiple manual moiety substitutions, followed by their chemical synthesis and re-evaluation. Although this might generate potential leads, a meticulous use of computational approaches could likely have accelerated this process further. Molecular docking studies aided the final lead generation of the ABT-199 compound; however, combining NMR and molecular docking can still introduce mistakes and whereas docking alone can also not be fully trusted (Chen, 2015). Even the newest scoring functions, implemented using artificial intelligence (AI), likely still present significant limitations (Gabel et al., 2014), as incomplete or erroneous classification of existing experimental data can compromise and bias their validity (Uçar et al., 2020).

Despite these setbacks, we firmly expect that newer chemo-informatics and AI algorithms, together with improved high-performance computing resources for rapid parallel in-silico drug screening using Molecular Dynamic (MD) simulations, will replace the current optimisation stages of the FBDD protocols. Several algorithms can already complement the experimental data validation (Bienstock, 2011). In addition to experimental methods, such as X-ray crystallography, NMR and electron microscopy, AI-driven structure elucidation by tools such as AlphaFold-2 (Jumper et al., 2021) and RoseTTA fold (Baek et al., 2021) can also provide the 3D molecular information required for the fragment-optimization stage. Ideally, the computational approaches for fragment optimization should simultaneously take the interaction, the experimental data such as derived by NMR, as well as biological aspects into consideration. For example, the robustness of a candidate compound or potential drug resistance could also be assessed early in the development process, thus integrating multiple fields of drug-discovery and pharmacology in a holistic way that scientists alone could not have attained (Dey and Caflisch, 2008).

By analysing the molecular characteristics of the various compounds, we have identified correlations in the development patterns among the three cases studied. By comparing the Tanimoto coefficients (Bajusz et al., 2015) for the three final compounds (Figures 2C, 3C, 4C), we speculate that expanding the molecule (via growing/linking methods) did not improve the binding affinity. This suggests that just covering the conformational molecular interaction space does not necessarily lead to higher affinity drugs. Figure 5B displays the normalised Tanimoto scores for the three different fragment-to-drug evolutions. Clearly, the development of AZD-3839 and ABT-199 display a highly similar pattern, which is dissimilar from S64315. An obvious conclusion could be that they simply differ in their optimisation protocols. However, it could equally suggest that the optimization of S64315 is not yet complete and multiple additional changes might occur before “converging” to the ultimate drug.

In conclusion, although only a few drugs approved by FDA have an NMR fragment-based trackable history that is easily accessible from public domain data, an analysis of relevant publications over the last 6 years shows an impressive number of journal articles reporting on discoveries of molecules in the pre-clinical stages in which NMR had a crucial role (Figure 1A). One of these molecules, Asciminib (Schoepfer et al., 2018), has only just recently (2021) been granted a FDA Breakthrough Therapy designation (Furnari et al., 2021), and it constitutes yet another marvellous example of the crucial role of NMR experiments in every stage of its development (Schoepfer et al., 2018).

In all, the overall methodological advances in the steps leading from initial hit to candidate-drug provide great hope that, compared to the time currently taken, new potent and selective drugs will soon become much more rapidly available.