The mechanistic functional landscape of retinitis pigmentosa: a machine learning-driven approach to therapeutic target discovery

Retinitis pigmentosa (RP) is a rare hereditary disease that mainly affects the photoreceptors in the retina, resulting in progressive vision loss. RP is typically diagnosed in childhood and rather than a single entity, RP is better described by a group of disorders characterized by night blindness, narrowing of the visual field, and ultimately, legal blindness [1]. Due to RP’s relatively high prevalence among rare diseases, affecting more than 1.5 million patients worldwide (from 1/2500 to 1/7000) [2], it is considered the most common inherited disease of the retina [3]. It is a complex disorder with multiple genetic causes and diverse clinical manifestations, counting mutations in over a hundred genes known to lead to the development of the disease [4]. However, the genetic cause of RP remains unknown in 50% of the patients [5]. In addition, current therapeutic options for RP are limited, with only a few treatments available that can slow down the progression of the disease or improve the quality of life of patients.

Increasing the knowledge of the molecular basis of the disease is a crucial step in the search for new therapeutic options to treat RP. The exponential growth of genomic data, fueled by sequencing technologies [6], has revolutionized the field of rare diseases with foreseeable achievements such as being able to diagnose all Mendelian diseases in the near future [7]. The knowledge generated in the last decade has expanded by more than two orders of magnitude the number of variants with known phenotypic effects [8]. However, the study of diseases has undergone a conceptual shift, moving from a focus on individual genes to exploring how they interact within cells [9]. This shift has led to the use of network-based models that rely on accumulated biological knowledge to identify causal relationships between genes, as defined by biological pathways [10]. The so-called “mechanistic models of human diseases” were developed by combining these novel approaches with biological human data.

Mechanistic models can be defined as mathematical representations of biological systems that aim to capture the underlying mechanisms that drive the behavior of these systems. Interestingly, these models can provide a causal link between gene expression and functional cell behavior [11]. However, the real potential of mechanistic models relies on their capability to reduce the complexity of the data while reinforcing interpretative power, unveiling specific interactions and functional outcomes [10]. Indeed, mechanistic models have helped to understand the disease mechanisms behind different cancers [12‐14], rare diseases [15], complex diseases such as diabetes [15], or obesity [16], as well as the mechanisms of action of drugs [17], including drug repurposing [18] or gender-specific effects of drugs in cancer [19]. A relevant property of mechanistic models is that they model causality in a quite accurate manner [9]. Therefore, they can predict the potential consequences of gene perturbations (e.g. knock-out, drug inhibition effect, etc.) over pathway activities, as well as the downstream functional consequences in the cell [10, 20]. They have been used to predict candidate essential genes in cancer cell lines, whose further inhibition validated the predictions [12]. Repurposable drug candidates were proposed for Fanconi Anemia [17] which were further experimentally validated [21], and for COVID-19 [18], also validated using a cohort of around 16,000 patients of the Andalusian healthcare database [22, 23]. These computational methods represent a leap in the generation and testing of hypotheses guided by a specific functional rationale, especially useful in scenarios with limited knowledge like rare diseases such as RP.

The development of novel therapies is a slow and costly process, particularly in rare diseases, where orphan drug research accounts for limited resources. Drug repurposing, the process of discovering new therapeutic uses for existing drugs [24], emerges as a potential solution to this problem in this context [25], reducing the time and expenses involved in drug development [26].

In this work, we propose a data-driven approach to construct actionable models that describe the disease mechanism of RP. Moreover, by making use of explainable machine learning models, trained with publicly available gene expression data, we have estimated the impact that targets of drugs already in use have in the signaling mechanisms of RP, allowing us to infer and select the most suitable therapeutic target genes that will lead us to drug candidates for its repurposing in RP. Indeed, the role in RP progression of six of the proposed targets has been experimentally validated. The comprehensive analysis of the results enabled us to identify drug targets that hold significant potential in impacting disease mechanisms able to revert cell functional processes to a healthy state.

The overall methodology developed for RP modeling is divided into three main stages depicted in Fig. 1. Briefly, the process entails collecting data on RP-related mechanisms from existing open-access databases (DBs). We ensured the precision and fidelity of RP-phenotypes and RP-gene associations by using standardized definitions from OrphaNet [27], OMIM [28], and the Human Phenotype Ontology DBs [29]. The information retrieved, together with the accumulated knowledge on functional gene interactions provided by the KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways DB [30], is used to produce a map of functional interactions among RP-associated genes and genes upstream and downstream in the pathways, thus defining a set of sub-pathways, the so-called signaling circuits, which can be considered a reasonable representation of the RP disease map [17, 18]. This RP mechanistic map can be used to build a signal transduction model that allows the estimation of RP-related cell functions from gene activities measured as gene expression levels [12]. The model of the RP mechanistic map is then integrated into an explainable machine learning (ML) model, which infers potentially causal relationships between genes of interest, in our case known drug targets (KDTs) extracted from the DrugBank DB, and the RP-related functional activities as described in the RP mechanistic map. The ML methodology prioritizes KDTs with a high predictive score, degranulating the results into the specific influences that each KDT has over the different parts of the mechanistic map. The results highlight KDTs that are predicted as relevant in regulating the functional landscape of RP mechanistic map and could potentially be further studied as promising therapeutic targets.

In this work, we propose a unique approach that, starting from the set of disease-affected genes, provides a comprehensive landscape of the molecular mechanisms of the RP disease along with its druggable space. The method establishes the RP mechanistic disease map as an actionable environment and employs an explainable machine-learning model to assess the influence of druggable molecules, like KDTs, over the disease environment. Our approach merges information from transcriptomics, pathway graphs, biological/clinical DBs, and drug-target interactions, to generate an in-depth view of the disease. The novelty of this workflow lies in the integration of multiple data sources, reinforcing interpretability with biological knowledge while reducing the dimensionality of the datasets.

Our methodology has enabled us to create a comprehensive yet straightforward map of RP, representing its functional landscape. The utility of this method relies on its interpretability. Artificial intelligence is often used in repurposing methods to compensate for the lack of prior knowledge. However, rational target discovery and drug repositioning can be a useful addition when prior knowledge is available, as it can help fill in the gaps and limit the unknown variables. By employing mechanistic models grounded in solid biological knowledge, we have managed to reduce dimensionality by providing a causal link between gene expression and functional cell behavior, and assess the impact of specific targets on different parts of the mechanistic map. This is particularly beneficial for diseases with diverse genotypic presentations and complex pathophysiology, such as RP [5].

Retinitis pigmentosa (RP) is a heterogeneous genetic disorder that leads to the progressive degeneration of photoreceptor cells in the retina, eventually leading to blindness. Although the pathogenesis of RP is complex, as part of this work, we have identified key features of the disease. Based on the signaling circuits identified by our methodology, and supported by the scientific literature, we have defined nine hallmarks of RP: “Apoptosis”, “Necrosis”, “Oxidative stress”, “Inflammatory response”, “DNA integrity”, “Neuronal processes”, “Fatty-acid–lipid metabolism”, and “Sensory and stimuli response”.

Photoreceptor cells, like other neuronal cells, have a postmitotic nature that allows for extended cell life beyond normal physiological conditions, but it may also induce abnormal photoreceptor cell death [45, 46]. We have indeed defined cell death (“Apoptosis” and “Necrosis”) as RP hallmarks within the proposed RP actionable map.

In RP, apoptosis is the major mechanism of photoreceptor cell death, involving the activation of caspases or certain members of the TNF receptor family [47], triggered both by intracellular stress signals, such as DNA damage or oxidative stress, processes defined as RP hallmarks, or by the binding of extracellular ligands to death receptors on the cell surface [43, 48].

Necrosis also has been linked to retinal degeneration as the main mechanism responsible for massive uncontrolled photoreceptor cell death [49], necrosis-like processes can happen rapidly and are linked to a failure in bioenergetics [46]. Interestingly, receptors that initiate cell death are part of the TNF superfamily, a family of cytokines involved in inflammation (another defined hallmark of RP) and in several pathways within our defined RP map [50, 51]. Inflammation is usually mediated by TNFα, IL-6, IL-1α, and IL-1β, among other mediators, all of which are present in RP patients and murine models of the disease [52‐54], regulating the production of chemokines and cytokines [55‐58]. Moreover, TNFα can promote, via RIPK3-MLKL signaling, the activation of NLRP3 inflammasome, a major multiprotein that induces inflammatory-mediated immune cell infiltration, accelerating necrosis-like photoreceptor cell death [58‐60].

Oxidative stress is another critical component of the pathogenesis of RP that is present in both the RP map and the predicted relevant KDTs. In RP, oxidative stress is initiated by the release of ROS from dying photoreceptor cells and microglia cells. The accumulation of ROS leads to the activation of additional stress pathways, including the unfolded protein, leading to cell death [44, 60‐62].

We also observed the implication of energetic and glucose-related functions, including HIF1 and HK1 pathways. Both HKI and HKII isoforms are expressed in the retina and target genes of the HIF-1 pathway, part of the RP map [58]. In fact, several cases of mutations in the HK1 gene were reported to cause RP in humans [4, 5]. This could affect glucose metabolism and energy supply to retinal cells, given that photoreceptor cells have very high energetic requirements.

Since neurons are highly excitable and sensitive to ischemia-induced death, due to their high ATP turnover rate [48, 63], the presence of voltage-gated sodium or calcium channels makes neurons susceptible to overload and swelling, while glutamate-gated channels make neurons susceptible to excitotoxicity and rupture [64‐66]. Thus, stimuli and sensory transduction also plays a role in RP progression, making photoreceptor cells and other neural cells susceptible to excitotoxic neurotransmitters, such as glutamate, or to the action of other neurotransmitters, like GABA or glycine [67]. As mentioned above, RP leads to a progressive loss of photoreceptor cells, but other neural cells suffer morphologic and metabolic changes resulting in retinal remodeling including loss of retinal ganglion cells (RGCs) or network rewiring [68].

As commented above, gene defects primarily lead to the dysfunction and death of photoreceptor cells and many therapeutic approaches are focused on these cell types (outer retina). However, as RP progresses a remodeling of the inner retina also occurs including formation of abnormal neural circuits and changes in the electrophysiologic properties of the retinal network [69‐71]. After photoreceptor cell death, the inner retina changes their synaptic connections affecting the transmission of information [72]. The communication between retinal neurons is mediated electrical but also chemical signals, the former includes the neurotransmitters glutamate, and GABA and their receptors. Glutamate mediates vertical communication between photoreceptor-bipolar cells and bipolar-ganglion cells. While GABA mediates lateral communication via horizontal and amacrine cells. The maintenance of the integrity of the visual systems beyond photoreceptors is a limitation for the successful outcome of any retinal treatment. In this study, we selected as KDT some subunits of the receptors associated to the glutamatergic, GABAergic and Glycinergic neurotransmission because changes in these receptors would have a significant impact in retinal cross-talk. In addition, the understanding of the time-course of the retinal remodeling may be important for achieving a proper window of therapeutic intervention [69].

In this study, we predicted and experimentally validated that RP is accompanied by alterations in neurotransmitter subunits, such as glutamate, GABA, and glycine. In many types of horizontal and amacrine cells, GABA serves as the main inhibitory neurotransmitter, in addition to glycine, which is known to also have an inhibitory role [73]. Retinal activity functions required a balance between excitation (glutamate) and inhibition (GABA and glycine). By increases in chloride conductance, GABA causes membrane hyperpolarization. Active chloride extrusion is achieved via the coupled transport of potassium and chloride ions via K–Cl cotransporters (KCC), with KCC2 being the primary active chloride extrusion system responsible for GABAergic systems. Hence, the changes in KCC2 expression (predicted and validated) could affect the GABAergic driving force [74]. Indeed, the ionotropic GABA A receptors (consisting of five subunits) are predicted as relevant KDTs, and found in almost all retinal cells [75, 76]. Inhibitory neurons seem to influence how the excitatory neurons integrate information through synaptic interactions. Under physiological conditions, retinal inhibition modulates the threshold responses of RGCs under ambient light, selectively masking certain signals to ensure the appropriate ones cross the optic nerve. GABA A receptor seems to be involved in the dynamic control of this masking inhibition and serves as a mechanism for neuronal adaptation. The upregulation of the alpha 1 and epsilon subunits of the GABA A receptor (Fig. 5C, D) observed in rd10 retinas could potentially change the threshold of RGCs activities leading to modifications in the perceived visual environment [77].

RP leads to photoreceptor degeneration triggering changes in the retinal morphology in human and animal models of RP. Studies on rd1 mice, carrying a mutation on the exon 7 of the Pde6b gene encoding the beta subunit of cGMP-PDE, with a greater rate of degeneration than rd10 mice, show that cone-mediated GABAergic amacrine cells displayed functionally altered NMDA receptors (such as Grin1) and high levels of GABA in rod bipolar cells, processes predicted by our model and validated as significantly affected in rd10 mice (Figs. 5E–J, 6 and Additional file 14: Fig. S4). In other studies, the use of GABA receptor antagonists in P23H rats (RP animal model carrying a missense mutation in the rhodopsin (rho) gene) reduces the observed stimulation thresholds of RGCs [78, 79]. Interestingly, the use of picrotoxin, an antagonist for GABA A/C and glycine receptors, also results in the blockade of the inhibitory mechanisms and ameliorates visual function in rd10 mice [80].

Although the role of GABAergic systems in RP is not completely established, our data suggest that the GABAergic system may be overactive during RP progression. We observed a significant increase in Gabra1 and Slc12a5 genes expression, their respective proteins GABARα1 and KCC2 as well as a similar distribution of them in the outer and inner plexiform layers (OPL and IPL) in retinal sections at P23. Functionally, GABA A receptor and KCC2 are both membrane proteins, closely related, which modulate GABAergic inhibition. An upregulation of Slc12a5 expression could increase Cl- extrusion, and subsequent Cl− gradient through the membrane facilitating the entry of Cl− through the ionotropic GABA A receptor. In addition, we showed an upregulation of Gabra1, which could exacerbate the inhibition induced by GABA through its binding to GABA A receptors. The higher staining of KCC2 and GABARα1 observed in the inner retina of rd10 mice (Fig. 6B, C) would support this idea. In addition, it has been suggested that the remodeling of the inner retina is accompanied by upregulation of glycine and GABA A receptors in rd1 mice [81]. Our findings suggest that therapeutic approaches aimed at reducing GABAergic inhibition and increasing the excitation of RGCs may be beneficial for RP. It has been suggested that anticonvulsants, such as tiagabine and vigabatrin, have a protective role in light-induced retinal degeneration in Abca4^−/−Rdh8^−/− mice, and that this protection is mediated by their role as GABA modulators [82]. Novel antiepileptic drugs act selectively through the GABAergic system and, although the underlying mechanism relating antiepileptic drugs with retinal degeneration is not yet fully explored, our model highlights the role of anticonvulsants in the mechanistic map of RP, probably through GABA-related processes [83]. Digging into the drugs targeting relevant predictive KDTs, the overrepresentation of anesthetic, hypnotics, and sedative drugs such as Pentazocine, Ketamine, Taurine, Flumazenil, and Phenobarbital among other drugs predicted by our model, coherently remark the relation between the anti-excitotoxic effects of certain anesthetics and their neuroprotection that has been long studied [84], whilst its role (and potential use) in RP remains undiscovered [85]. As a matter of fact, selected predicted drugs shown in Fig. 7C share a neuroprotective effect, also appearing to be related to fatty acid processes. Omega 3-related compounds, also predicted by our model, appear to have an impact on the progression of RP as discussed in Olivares-Gonzalez et al. [44].

The effectiveness of applying mechanistic models and ML methodologies to drug repurposing has been undoubtedly demonstrated in several works [17, 19, 86]. These approaches allow for the virtual testing of hundreds to thousands of scenarios with very minimal resources. While many ML-based repurposing strategies focus on the physicochemical properties of the drugs [87], they often overlook the overall impact on diseases or cells. This is where mechanistic models of disease come into play, offering insights not only into disease mechanisms but also into drug actions in specific disease contexts [88]. Indeed, supervised ML has been successfully used to boost how these tools model cellular behavior, improving the accuracy of mechanistic models in characterizing cell types [14, 49], identifying targets for disease treatments [17, 42], or predicting the effect of single or combined drugs on cells [50, 51].

The main purpose of this work was to detect actionable target genes that could have an impact on RP, providing an explorable resource of knowledge. Nevertheless, this methodology can also be used to model all cellular mechanisms to evaluate potential undesired off-target and side effects as well as to the identification of new disease-related genes, such as GABRA, GRIN1, KCC or SLC12A5 whose mutations can lead to the development of retinal dystrophies. Due to its exploratory nature, the lack of ground truth in the validation process comes as a limitation of the method, we have demonstrated the robustness of the results experimentally validating the role of selected KDTs, with no previous association with RP, in RP murine models. However, further research is needed to assess the in vitro and in vivo impact that the drugs targeting relevant KDTs would have on the disease progression or initiation.

Nevertheless, given that RP has more than 250 genes identified, but only 50% of RP cases are attributable to identified genes, this application on its own has tremendous potential for the clinical community, as a tool to evaluate the impact of genes of uncertain significance [5].

In this work, an actionable map of Retinitis Pigmentosa (RP) is presented, reflecting the mechanisms and dynamics of the functions involved in disease initiation and progression. The map serves as a resource for further analysis, validations, and computational modeling, such as network or graph analyses. The methodology identifies potential targets impacting RP processes, providing a list of relevances and pinpointing the mechanism, function, and circuit affected as well as the KDT responsible for the predicted impact. These relevances can guide and aid in finding possible clinical intervention points. Although focused on signaling pathways, the model can be expanded to other types of pathways using directed interaction networks, and various datasets representing gene/protein abundance, bearing in mind that limitations in sample numbers and balanced dimensionality must be considered. The model’s strength lies in its applicability to scenarios with limited knowledge, such as rare or emerging diseases like COVID-19 [19, 21, 86]. It can be adapted to infer interactions between genes and mechanisms, providing a sophisticated way to model cell behavior under disease conditions. Combining mechanistic models with ML provides a useful tool to explore disease druggable space, mechanisms, and pathogenesis aiding in genetic diagnosis and identifying genes with a key role in diseases. The possibility of using the available genomic repositories to formulate hypotheses and extrapolate conclusions in the rare diseases field constitutes an attractive opportunity to guide and accelerate the experimental validations in order to focus and prioritize target candidates.