Author: admin

Designing novel target molecules by integrating the topological, chemical, and physical attributes of protein cavities necessitates advanced neural networks. While existing approaches like Bicyclic Generative Adversarial Networks (BicycleGANs) (Skalic et al., 2019) and Recurrent Neural Networks (RNNs) (Xu et al., 2021) have demonstrated potential, end-to-end standalone tools for GPCR-specific ligand design remain scarce. To address this, we developed the Gcoupler and provided it to the community as a Python Package and a Docker image. Gcoupler adopts an integrative approach utilizing structure-based, cavity-dependent de novo ligand design, robust statistical methods, and highly powerful Graph Neural Networks. Gcoupler consists of four interconnected modules, that is, Synthesizer, Authenticator, Generator, and BioRanker, that collectively impart a smoother, user-friendly, and minimalistic experience for the end-to-end de novo ligand design.

Synthesizer, the first module of Gcoupler, takes a protein structure as input in Protein Data Bank (PDB) format and identifies putative cavities across the protein surface, providing users with the flexibility to select cavities based on druggability scores or user-supplied critical residues. Since cavity-dependent molecule generation mainly depends on the chemical composition and geometric constraints of the cavity, it is, therefore, indispensable to select the cavity for the downstream steps considering its chemical nature (hydrophobicity/hydrophilicity) and functional relevance (proximity to the active site or residue composition), among others. Accounting for these, Gcoupler offers flexibility to the users to select either of its predicted cavities based on the user-supplied critical residue or by user-supplied cavity information (amino acids) using third-party software (e.g., Pocketome) (Hedderich et al., 2022). To enhance user experience, Gcoupler computes and outputs all identified cavities along with their druggability scores using LigBuilder’s V3 (Yuan et al., 2020) cavity module. Briefly, these druggability scores consider solvent accessibility, cavity exposure or burial, and detected pharmacophores and cavities, which are further prioritized based on this score. Post-cavity selection, the Synthesizer module generates cavity-specific ligands influenced by topology and pharmacophores, outputting SMILES, cavity coordinates, and other requisite files to downstream modules for further steps (Figure 1a). The chemical composition of the in silico synthesized ligands by the Synthesizer module is influenced by the cavity topology (3D) and its composition (pharmacophores). Noteworthy, the Synthesizer module of Gcoupler employs LigBuilder V3 (Yuan et al., 2020), which utilizes the genetic algorithm for the in silico ligand synthesis. Notably, the fragment library of LigBuilder, comprising 177 distinct molecular fragments in Mol2 format, allows the selection of multiple seed structures and extensions that best complement the cavity pharmacophores throughout multiple iterative runs. For each run, once a seed structure is confirmed, Gcoupler employs a hybrid approach of the Growing and Linking modes of the LigBuilder build module, enabling the stepwise addition of small fragments to the seed structure within the binding pocket of the target GPCR to build synthetic ligands. Gcoupler generates 500 unique molecules by default, though it can also be user-defined. The Synthesizer module of Gcoupler enhances LigBuilder V3 practical applicability through automation, dynamic adaptability, and abstraction. This allows for more efficient and targeted ligand generation, even in challenging design scenarios for GPCR ligand design. However, it lacks user-defined library screening, proposes synthetically challenging molecules, and often requires post-processing to isolate High-Affinity Binders () from a broad affinity range of synthetically designed compounds.

To address this limitation, a second module was added to Gcoupler, termed Authenticator. This module processes output files from the Synthesizer module, conducting downstream validation steps and preparing results for constructing deep learning-based classification models (third module). The Authenticator requires input protein 3D structure in PDB format, cavity coordinates, and all silico-generated molecules from the Synthesizer module. Authenticator utilizes this information to further segregate the synthesized compounds into HABs and Low-Affinity Binders (LABs) by leveraging a structure-based virtual screening approach (AutoDock Vina) (Trott and Olson, 2010) and statistically backed hypothesis testing for distribution comparisons (Figure 1a). The Authenticator module outputs the free binding energies of all the generated compounds, which further segregates the compounds into HABs and LABs by the statistical submodule while ensuring the optimal binding energy threshold and class balance. Of note, the Authenticator is also capable of leveraging the Empirical Cumulative Distribution Function (ECDF) for binding energy distribution comparisons of HABs and LABs and performs the Kolmogorov–Smirnov test (Berger and Zhou, 2014), Epps–Singleton test (Goerg and Kaiser, 2009), and Anderson–Darling test (Engmann and Cousineau, 2011) for hypothesis testing. This expanded array of statistical tests allows users to employ methodologies that best suit their data distribution characteristics, ensuring robust and comprehensive analyses. Moreover, the Authenticator module incorporates a unique feature for decoy synthesis using HABs. This functionality enables the generation of a negative dataset in scenarios where the Synthesizer module fails to produce an optimal number of LABs. By synthesizing decoys from HABs, users can effectively balance their datasets, enhancing the reliability of downstream analyses. Lastly, the Authenticator module also accommodates user-supplied negative datasets as an alternative to LABs (Mysinger et al., 2012). This feature provides users with the flexibility to incorporate external data sources, enabling robust prediction model building by the subsequent Generator module.

The Generator, the third module, employs state-of-the-art GNN models such as Graph Convolution Model (GCM), Graph Convolution Network (GCN), Attentive FP (AFP), and Graph Attention Network (GAT) to construct predictive classifiers using Authenticator-informed classes. These GNN algorithms are tailored to extract features from the graph structure of the compounds generated by the Synthesizer and apply them to the classification task by leveraging Authenticator-informed class information. For instance, the GCM assimilates features by analyzing neighboring nodes, while the GCN extracts features through a convolutional process. The AFP model focuses attention on specific graph segments, and the GAT employs attention mechanisms to learn node representations. By default, Generator tests all four models using standard hyperparameters provided by the DeepChem framework (https://deepchem.io/), offering a baseline performance comparison across architectures. This includes pre-defined choices for node features, edge attributes, message-passing layers, pooling strategies, activation functions, and dropout values, ensuring reproducibility and consistency. All models are trained with binary cross-entropy loss and support default settings for early stopping, learning rate, and batch standardization where applicable. Gcoupler provides off-the-shelf hyperparameter tuning to ensure adequate training, which is essential for optimizing model performance. After selecting the best parameters and classification algorithm, Gcoupler further ensures the mitigation of overfitting and provides a more precise estimate of model performance through k-fold cross-validation. Notably, by default, Gcoupler employs threefold cross-validation, but users can adjust this parameter.

Finally, BioRanker, the last module, prioritizes ligands through statistical and bioactivity-based tools. The first level ranking offered by BioRanker is composed of a statistical tool that encompasses two distinct algorithms, namely G-means and Youden’s J statistics, to assist users in identifying the optimal probability threshold, thereby refining the selection of high-confidence hit compounds (Figure 1—figure supplement 1a). Additionally, bioactivity embeddings computed via Signaturizer (Bertoni et al., 2021) enable multi-activity-based ranking using a modified PageRank algorithm. Briefly, the bioactivity descriptors of the predicted compounds are projected onto various biological activity spaces, including Chemistry, Targets, Networks, Cells, and Clinics, by performing pairwise cosine similarity comparisons with HABs. The PageRank algorithm is then applied for activity-specific ranking and supports multi-activity-based ranking for sequential screening based on user-defined biological properties. BioRanker also offers flexibility through customizable probability thresholds, enabling stringent or relaxed selection of compounds. Users can also input SMILES representations for direct screening, bypassing prediction probabilities. Taken together, Gcoupler is a versatile platform supporting user-defined inputs, third-party tools for cavity selection, and customizable statistical analyses, enhancing its adaptability for diverse ligand design and screening tasks. This integrated framework streamlines cavity-specific ligand design, screening, and ranking, providing a comprehensive solution for GPCR-targeted drug discovery.

To evaluate Gcoupler’s performance, we tested its modules across five GPCRs (AA2AR, ADRB1, ADRB2, CXCR4, and DRD3) using experimentally validated ligands and matched decoys from the DUD-E dataset (Mysinger et al., 2012). The DUD-E datasets contain five GPCRs alongside information about their cavity coordinates, positive ligands, and decoys (https://dude.docking.org/subsets/gpcr). We used these five GPCRs as independent samples to evaluate different modules and sub-modules of Gcoupler. We first checked whether the cavity search algorithm of Synthesizer could accurately detect a given orthosteric ligand-binding site for a GPCR. Gcoupler accurately identified orthosteric ligand-binding sites and additional allosteric cavities across all targets, validating its de novo cavity detection algorithm (Figure 1—figure supplement 1b). We next asked whether Gcoupler could also synthesize molecules similar to the reported ligands for respective orthosteric sites based on the cavity’s physical, chemical, and geometric properties. For orthosteric sites, the Synthesizer module generated ~500 compounds per GPCR. Subsequently, as per the Gcoupler default workflow, the Authenticator module conducted a virtual screening of these newly synthesized compounds, segregating them into HABs and LABs. Although the Authenticator module provides flexibility in selecting an optimal threshold to distinguish HAB and LAB, we chose the default cutoff of –7 kcal/mol for AA2AR, CXCR4, and DRD3. For ADRB1 and ADRB2, we selected a threshold of –8 kcal/mol to minimize overlap in distributions and thus avoid class imbalance, a critical parameter that could influence the downstream model generation using the Generator module (Figure 1—figure supplement 1c). Statistical validation confirmed significant separation between these groups (p < 0.0001), enabling the Generator module to construct graph-based classification models with high values of AUC–ROC (>0.95), sensitivity, and specificity (Figure 1b, c, Figure 1—figure supplement 1d, e). These models reliably distinguished ligands from decoys, demonstrating Gcoupler’s accuracy in identifying high-affinity ligands.

In addition to evaluating Gcoupler’s performance for the orthosteric sites of GPCRs, we also validated its capability to identify allosteric sites and their corresponding ligands. In this case, we first gathered information about the experimentally validated GPCR–ligand complexes sourced from the PDB database. We chose three GPCR–ligand complexes (β2AR-Cmpd-15PA, CCR2-CCR2-RA-[R], and CCR9-Vercirnon) from the PDB (Shen et al., 2023). We removed the ligands from the PDB files and executed the standard Gcoupler workflow with default parameters. Gcoupler successfully identified allosteric binding sites and generated classification models for synthetic compounds with consistently high AUC–ROC values (>0.95) (Figure 1d, Figure 1—figure supplement 2a–c). This high level of accuracy indicates the robustness of Gcoupler’s algorithms in distinguishing between true positives (allosteric ligands) and true negatives (non-binders). Projection of experimentally validated ligands onto these models further confirmed their predictive accuracy (Figure 1e), underscoring Gcoupler’s robustness and versatility for orthosteric and allosteric ligand discovery.

Next, to evaluate the efficiency of Gcoupler, we compared its run time with the biophysics-based gold standard molecular docking (AutoDock) (Morris et al., 2009). To address the runtime efficiency, we first utilized the ChEMBL31 database (Gaulton et al., 2012) to identify GPCRs with the highest number of reported experimentally validated agonists. We selected the alpha-1A adrenergic receptor (ADRA1A) since it qualifies for this criterion and contains 993 agonists (Figure 1—figure supplement 3a, b). Methodologically, we followed the conventional steps of AutoDock Tools for molecular docking while keeping track of execution time for each step throughout the entire process until completion (Figure 1—figure supplement 3a, c). In parallel, we applied the same timestamp procedure for Gcoupler, including its individual module sub-functions (Figure 1—figure supplement 3a–d). Gcoupler was 13.5 times faster, leveraging its deep learning-based Generator module and AutoDock Vina’s efficiency. Both methods provided comparable predictions for active compounds, demonstrating Gcoupler’s speed and accuracy, making it ideal for large-scale ligand design and drug discovery (Figure 1—figure supplement 3e–h).

Finally, we used Gcoupler to evaluate the ligand space conservation (functional conservation) of the GPCR–Gα interface. Specifically, we aimed to explore the possibility of direct small molecule binding to the GPCR–Gα interface to modulate downstream signaling pathways. We analyzed multiple human GPCR–Gα complexes from the PDB (Figure 1f, Supplementary file 1), identified conserved motifs (DRY, CWxL, and NPxxY) and binding pockets through sequence and structural analyses (Figure 1g). To determine the topological similarity of the GPCR–Gα-protein interface, we undertook a detailed structural analysis across a wide array of GPCR–Gα-protein complexes. This analysis involved identifying and extracting the cavities present within each complex. By focusing on these critical regions, we aimed to assess the degree of structural conservation and quantify it through normalized root mean square deviation (RMSD) values. Specifically, the normalized RMSD values, which provide a measure of the average distance between atoms of superimposed proteins, indicated a high degree of similarity. The mean RMSD value was found to be 1.47 Å, while the median RMSD value was even lower at 0.86 Å. These values suggest that the overall topology of the GPCR–Gα interface is well conserved across different complexes, highlighting the robustness of this interaction site (Figure 1h–j, Figure 1—figure supplement 3i–k, Supplementary file 2). Finally, to test whether this topological and sequence conservation also impacts the ligand profiles that could potentially bind to this interface, we performed the Gcoupler workflow on all 66 GPCRs and synthesized ~50 unique ligands per GPCR (Figure 1h). We next computed and compared the physicochemical properties (calculated using Mordred; Moriwaki et al., 2018) of these synthesized ligands and observed high cosine similarity, which further supports the functional conservation of the GPCR–Gα interface (Figure 1h, k, Figure 1—figure supplement 3l, Supplementary file 3). In summary, we used Gcoupler to systematically evaluate and analyze the ligand profiles of the GPCR–Gα-protein interface and observed a higher degree of sequence, topological, and functional conservation.

Here, we engineered BOAS that included tandemly linked, antigenically distinct HA heads as a single construct. This platform allows a mixing-and-matching of up to eight distinct HA heads from both influenza A and B viruses. Furthermore, we showed that the order and number of HA heads can vary without losing reactivity to conformation-specific mAbs in vitro, highlighting the flexibility of this platform. Mice immunized with BOAS had comparable serum reactivity to each individual component though relative binding and neutralization titers varied between immunogens; this is likely a consequence of length and/or composition. Further oligomerization for increased multivalent display was accomplished by conjugating two 4mer BOAS inclusive of eight distinct HA heads to a ferritin nanoparticle via SpyTag/SpyCatcher ligation. Similar to the BOAS, these conjugated nanoparticles elicited similar titers to all eight HA components and could neutralize matched viruses.

Thus, tandemly linking HA heads is a robust method for displaying multiple influenza HA subtypes in a single protein-based immunogen. Binding titers were elicited to all components present in the immunogen, and there was no significant correlation between HA position within the BOAS (i.e. internal or terminal) and immunogenicity. However, the relative immunogenicity of each HA varied despite equimolar display of each HA subtype. There were qualitatively immunodominant HAs, notably H4 and H9, and these were relatively consistent across BOAS in which they were a component; this effect was reduced in the mix cohort. Further studies using the modularity of the BOAS could further deconvolute relative immunodominances of HA subtypes.

Despite similar binding titers across multiple BOAS lengths, expression levels and neutralization titers were quite variable. While all 3mer to 8mer BOAS could be overexpressed, expression inversely correlated with overall length. To mitigate this, multiple BOAS (e.g. two 4mers) or conjugation to protein-based nanoparticles, as was done here, could be used to ensure coverage of each desired HA subtype. Furthermore, neutralization titers were quite variable across different BOAS lengths despite similar binding titers. This may be related to multiple factors, including homology, stability, and accessibility of neutralizing epitopes for different BOAS lengths. Notably, for longer BOAS, we observed degradation following longer term storage at 4° C, which may reflect their overall stability. Studies manipulating BOAS composition at intermediate lengths could optimize neutralizing responses to particular influenzas of interest.

Based on the immunogenicity of the various BOAS and their ability to elicit neutralizing responses, it may not be necessary to maximize the number of HA heads into a single immunogen. Indeed, it qualitatively appears that the intermediate 4-, 5-, and 6mer BOAS were the most immunogenic and this length may be sufficient to effectively engage and crosslink B cell receptors (BCRs) for potent stimulation. These BOAS also had similar or improved binding cross-reactivity to mismatched HAs as compared to longer 7- or 8mer BOAS. Notably, the 3mer BOAS elicited detectable cross-reactive binding titers to H4 and H5 mismatched HAs. This observed cross-reactivity could be due to sequence conservation between the HAs, as H3 and H4 share ~51% sequence identity, and H1 and H2 share ~46% and~62% overall sequence identity with H5, respectively (Figure 4—figure supplement 1). Additionally, the degree of surface conservation decreased considerably beyond the 5mer as more antigenically distinct HAs were added to the BOAS. These data suggest that both antigenic distance between HA components and BOAS length play a key role in eliciting cross-reactive antibody responses, and further studies are necessary to optimize BOAS valency and antigenic distance for a desired humoral response.

Potential enrichment of serum antibodies targeting the conserved RBS and TI epitopes may also be contributing to observed cross-reactivity. Both epitopes are relatively conserved across all BOAS (Figure 4C), and the two BOAS showing the most cross-reactivity, the 3mer and 5mer, elicit a significant portion of the serum response toward both RBS and TI epitopes as determined via a serum competition assay with available epitope-directed mAbs (Figure 4B). Notably, this proportion is approximate, as at the time of reporting, mAbs that bind the receptor binding site of all components were not available. RBS-directed mAbs to the H4 and H9 components were not available, and the RBS-directed antibodies used targeting the other HA components have different footprints around the periphery of the RBS. Additionally, there are currently no reported influenza B TI-directed mAbs in the literature. Therefore, this may be an underestimate of the serum proportion focused on the conserved RBS and TI epitopes. Isolated TI-directed mAbs, in particular, can engage more than nine unique subtypes across both group 1 and 2 influenzas (McCarthy et al., 2021; Watanabe et al., 2019), and our monomeric head-based BOAS immunogens have the otherwise occluded TI epitope exposed (McCarthy et al., 2021; Bangaru et al., 2019; Watanabe et al., 2019). Furthermore, we have previously shown that this TI epitope, when exposed, is immunodominant in the murine model (Bajic et al., 2019). Further studies with different combinations of HAs could aid in understanding how length and composition influences epitope focusing. For example, a BOAS design with a cluster of group 1 HAs followed by a cluster of group 2 HAs, rather than our roughly alternating pattern could influence which HAs are in close proximity to one other or could be potentially shielded in certain conformations and thus could affect antigenicity. Combining the BOAS platform with other immune-focusing approaches (Dosey et al., 2023), such as hyperglycosylation (Bajic et al., 2019; Thornlow et al., 2021; Ingale et al., 2014; Eggink et al., 2014) or resurfacing (Bajic et al., 2020; Hai et al., 2012) could enhance cross-reactive responses. Additionally, modifying linker spacing and rigidity can also be used as a mechanism to enhance BCR cross-linking and thus enhance cross-reactive B cell activation and elicitation (Veneziano et al., 2020).

BOAS can be further multimerized via conjugation to a surface of a NP. Interestingly, this only had a marginal effect on immunogenicity. The BOAS NP elicited titers of ~10⁵ (Figure 6B), whereas the best BOAS alone reached an order of magnitude greater (Figure 3C). This appears in contrast with other studies where attaching an antigen to a NP scaffold enhanced immunogenicity and neutralization potency (Kanekiyo et al., 2013; Jardine et al., 2013; Tokatlian et al., 2019; Kato et al., 2020; Marcandalli et al., 2019). One recent example designed quartets of antigenically distinct SARS-like betacoronavirus receptor binding domains (RBDs) coupled to an mi3-VLP scaffold via a similar SpyTag-SpyCatcher system and showed increased binding and neutralization titers following conjugation to the NP compared to quartet alone (Hills et al., 2024). This discrepancy may be in part due to the larger mi3 NP (Bruun et al., 2018) which displays 60 copies of the antigen rather than the 24 copies displayed on the ferritin NP used in this study. It is also possible that the difference in immunogenicity could arise from the increased molecular weight of the BOAS NP immunogen compared to the BOAS alone, leading to a difference in moles of BOAS antigen in each cohort. However, due to the large size of the BOAS, the addition of the ferritin NP does not add a large amount of mass. 20 µg of BOAS NP or an 8mer BOAS equates to ~64 and ~83 µmoles of each HA component, respectively. This ~30% greater amount of HA in the 8mer BOAS, however, does not account for the observed difference in serum binding titers. Nevertheless, HA-specific responses were similar whether the BOAS were conjugated to the nanoparticle or not, indicating that HA proximity to the NP surface did not impact responses to each component. This observation is consistent with betacoronavirus quartet NPs as well. Additionally, BOAS conjugation to the NP significantly reduced the scaffold-directed response. The addition of the large BOAS projections to the NP surface likely masked the immunogenic scaffold epitopes (Kraft et al., 2022).

Collectively, this study demonstrates the versatility of the BOAS platform to present multiple HA subtypes as a single immunogen. This ‘plug-and-play’ approach can readily exchange HAs to elicit desired immune responses. BOAS are potentially advantageous over other multivalent display platforms, such as protein-based NPs, which can produce off-target responses due to their inherent immunogenicity (Kanekiyo et al., 2013). Furthermore, when genetic fusions of the antigen to nanoparticles is not possible, SpyTag-SpyCatcher (or another suitable conjugation approach) must be used, further contributing to scaffold-specific responses as well as additional multi-step manufacturing and purification challenges. Not only does our BOAS platform circumvent these potential caveats, but because this is a single polypeptide chain, this immunogen could readily be formulated as an mRNA lipid nanoparticle (LNP) (Chaudhary et al., 2021). The BOAS platform forms the basis for next-generation influenza vaccines and can more broadly be readily adapted to other viral antigens.

Activewear brand Sweaty Betty has become involved in a new dispute over advertising slogans, which a period underwear company claims were copied.

Kelly Newton said Sweaty Betty’s use of two taglines that were very similar to her firm Nixi Body’s “seemed a little off”, and while she could not get them trademarked she felt Sweaty Betty was “taking from” other female founders.

Sweaty Betty said the “No ifs. Just butt.” line had been used by many brands and influencers, but also said it was reviewing all its marketing campaigns.

Ms Newton said she was speaking out after seeing personal trainer Georgina Cox reveal Sweaty Betty had offered her a settlement over a disputed slogan.

Ms Newton, who co-founded Nixi Body in 2019, said the company has advertised its leak-proof period underwear with the lines “Keeping you moving through menstruation, motherhood and menopause” and “No leaks. No ifs. Just butts.” for years.

But late last year a friend alerted her to a Sweaty Betty ad for its femtech range of leak-resistant and maternity leggings with the tagline “Keeping you moving through menstruation, maternity and menopause”.

At first, Ms Newton thought it could have been a coincidence, and said she was annoyed but did not feel there was anything she could do.

That tagline has since been changed on the Sweaty Betty website to “Menstruation. Maternity. Menopause. Together we’re raising the bar for every woman, for any life stage.”

Then, earlier this year, she noticed Sweaty Betty was running a campaign with the line “No ifs. Just butt.”

“I thought ‘oh hang on a second’,” Ms Newton told the BBC.

Again, Ms Newton said she didn’t reach out. She said she had tried to get the “Keeping you moving through menstruation, motherhood and menopause” line trademarked but was unable to do so, and said she knew that legally she had little recourse.

“Legally they were within their rights to use these words, but they’re words that we’ve used for quite a few years.”

It was personal trainer Ms Cox sharing her story about dealing with the brand that made her want to speak out, Ms Newton said.

Last month Ms Cox revealed that Sweaty Betty had offered her £4,000 in exchange for confidentiality over the use of the phrase “Wear The Damn Shorts” that she claims was used without crediting her in its latest campaigns.

She told the BBC she came up with the slogan alongside her sister in 2020 and it quickly went viral. In 2023, Sweaty Betty approached her to use it in a campaign and paid her £3,500 to do so.

She was also paid for another campaign using it in 2024, but was not approached about it for this year.

Ms Newton said she felt compelled to speak out about her experience after talking to Ms Cox, because she was not the only person affected.

“Your tagline can’t be empowering all women when actually all you’re doing is taking from them,” she said, adding that she was not seeking compensation but rather for Sweaty Betty to be accountable.

“It just feels like for me all the slogans I’m seeing are coming from other female founders.

“I just want them to do the right thing and just morally acknowledging what they’ve done isn’t great and to do better really.”

Sweaty Betty said it chose words “to empower women through fitness” to advertise its femtech range, something many brands try to do, which “sometimes… means the language used by different brands aligns”.

“We have been using the phrase for over a year and Nixi has not been in touch with us about our use of it. The phrase is used by many brands in various forms and that is why no individual or brand can claim exclusive rights to it or trademark it,” the company said in a statement.

“It is never our intent to take credit for the work or creativity of others, particularly from trailblazing women, and we have reached out to Nixi Body directly to convey this. We have also been undertaking a review of all our campaigns and marketing language to fully understand the origins of the phrases we use and will continue to do this.

“We note that such a review was conducted with respect to the ‘Wear the Damn Shorts’ phrase and its origins date back to at least as early as 2019, before either Ms Cox or Sweaty Betty first used the phrase. We continue to work towards a resolution with Ms Cox, as has always been our objective, however this dispute remains ongoing.”