Cross-species comparative in silico platform (CRISP)

CRISP, created by Moleculomics Ltd, defines the potential of a small molecule compound to interact with metabolically crucial liver enzymes and receptors across a range of different species. The platform comprises structural models for all liver proteins; around 1,100 each of mouse, rat and human liver (mouse and rat account for 84% of all animal tests in the UK) (Annual statistics of scientific procedures on living animals, 2013). CRISP facilitates the identification of target proteins that share high structural homology across the three organisms and those that share broad structural homology but may possess critical structural differences affecting activity/specificity. Incorporation of an in vitro knowledgebase aggregating five established databases (IntAct, BioGRID, Reactome, KEGG, CTD) provides downstream signalling and metabolic pathway information, cementing CRISP as a leading biosimulation tool for the identification of the most suitable animal model for a given NCE development program.

Alternative protein structure-based in silico solutions do not offer the scale or breadth of CRISP, which spans the entire liver set of multiple proteomes. CRISP can predict the binding of a compound with a protein, the Molecular Initiation Event (MIE), and return all proteins implicated in the pathway of the MIE. A structural similarity search is performed to identify the structural and functional similarities/differences of every protein implicated in the MIE pathway for different species. Current in silico models are approached from a cheminformatics perspective and although they link chemical structure to biological activity, do not include specific protein-ligand interactions at the whole organ scale. CRISP combines high throughput structural modelling of proteins, molecular docking, in silico pathway analysis and the capacity for inter-species comparison. This facilitates rapid decision making within the R&D pipeline, enables informed selection of the most appropriate animal model, and identifies problematic compounds earlier in development by recognising protein-ligand interactions that evoke undesirable pathway outcomes.

Following further validation, CRISP will provide a comprehensive tool for early identification of protein-mediated efficacies and toxicity issues of compounds in relation to the liver, on a large scale and comparatively across human and two commercially and scientifically important model systems. Moleculomics plan to scale the technology to whole proteome analyses for human, cross referenced to all species used in animal testing, to realise the true potential of the technology.

References

• Home office Report –Annual statistics of scientific procedures on living animals – Great Britain (2013).

CRISP has the potential to be a valuable tool for efficient development of safe and innovative products in several high impact sectors such as agro-industries, pharmaceutical, biotechnology and defence. However, it is appreciated that industry will only adopt our tool if utility is demonstrated.

Moleculomics seek collaborators to assist in validation by providing information on a small number of “blind” small molecule test compounds for which in vitro protein assay and in vivo liver toxicity endpoint data is available. This will supply Moleculomics with valuable information required to both train and test the prototype to enable commercialisation of the technology. It will provide a robust foundation upon which the technology can be scaled in order to facilitate whole proteome analysis across a broad range of animal species frequently involved in preclinical trials.

In addition to the proposed validation using in vitro and in vivo hepatotoxicity data, long-term partnerships are sought to apply CRISP for the purpose of decision making in toxicology risk assessment. Partnerships are sought with companies whose work involves extensive compound discovery and/or appropriation, and for whom in the longer term, the platform would form an integral part of their compound -selection and/or risk assessment.

Information about IP

Moleculomics Ltd confirm the IP required to deliver the predictive molecular interaction components of work is proprietary and that other information used in this project exists in the public domain.

Moleculomics Ltd seek to overcome poor animal/human correlation, a key contributor responsible for the 90% of drugs that fail in human trials despite passing traditional toxicology tests involving rats (Sankar U, 2005). By reliably predicting similarities and differences between human and animal models it is possible to select more appropriate tests and animal models. This will eliminate the use of unnecessary animal models that are a poor representation of the human biochemistry and are of limited or no scientific value, reducing the number of animals required whilst simultaneously improving the efficiency of the compound development process.

Additionally, new molecular knowledge can be assimilated through CRISP to broaden scientific understanding of the mechanistic actions of NCEs and could negate the need to observe some systemic symptoms. For example, in the case of the Draize test, CRISP has the potential to predict which proteins of the human eye may be implicated through exposure to a candidate chemical compound and compare the structure/ pathway data to that of rabbit, to evaluate the relevance of the rabbit model. The aspiration is to develop the technologies to support the objectives of the NC3Rs and place in silico screening at the centre of how new compound based products are developed.

Reference

• Sankar U (2005). The Delicate Toxicity Balance in Drug Discovery. The Scientist 19:32.

Overview | Impact

Overview

With the support of CRACK IT Solutions funding Moleculomics Ltd have now established a project with Dow Agrosciences to validate CRISP against ten hepatotoxic compounds and ten non-toxic compounds (supplied by Dow AgroSciences), the identity of which will be blinded to Moleculomics. Data from the CRISP model will be compared to existing in vitro and in vivo data provided by Dow AgroSciences to evaluate the effectiveness of the model as a cost-effective compound screening method. The ability to predict/assess liver toxicity across a broad range of compounds will greatly facilitate the selection of suitable candidate(s) and reduce the need to test in vivo to discriminate.

Impact

CRISP uses publicly available information from five data repositories (UniProt, PubChem, Livertox, Reactome and KEGG) and focusses specifically on liver proteins. The structures of over one thousand liver proteins, with known homologues in each species (human, rat, and mouse), were modelled using homology modelling and protein threading (Table 1). Table 2 shows the number of human proteins that had an equivalent gene product modelled in both rat and mouse, only mouse, and only rat, highlighting significant species differences that may impact on the data generated.

Table 1- The number of 3D models generated, which demonstrate equal to or greater than 80% sequence alignment with the template(s) used and equal to or greater than 95% confirmed secondary structure using Procheck.

Table 2 - number of human proteins that have had an equivalent gene product modelled.

Ligands were collated from the LiverTox database and the FDA approved list, to derive a collection of hepatoxic and “safe” compounds. The compounds were divided into three broad sets:

• Least toxic- those in the FDA-approved set only and not listed in Livertox: 715 compounds.
• Medium toxicity- those that were listed in both databases: 656 compounds.
• Most toxic- those found only in the Livertox database and without FDA approval: 481 compounds.

Using the 3000 modelled liver proteins and nearly 2000 ligands, an extensive in silico docking programme was carried out, to identify protein- ligand interactions. A set of protein similarity scoring metrics was produced to identify and rank similar proteins across the three species. These measures of similarity include protein sequence identity, 3D structure of protein, and docked ligand conformation (including affinity and position of ligand binding). A comprehensive database was created from the docking and similarity data to enable cross-referencing of all the liver proteins in mouse, rat and human and generate meaningful comparative output.

An automated reporting framework was developed to produce a report on a query compound, detailing potential interactions with human proteins, metabolic and signalling pathways the compound is involved in, and similarities and differences with proteins in rodents. The report is delivered as an HTML document that includes ‘traffic light’ classifications to give a quick overview of similarity and can provide four different suggested outcomes: both rodent models are unsuitable, both rodent models are (equally) suitable, the mouse provides the best model or the rat provides the best model. See Appendix 1 for an example.

Dow AgroSciences was instrumental in providing feedback about the software and aided progression with the interface, but due to merger-related changes Dow AgroSciences was unable to provide the compounds for validation. Instead, CRISP was validated using ten hepatoxic and ten non-toxic compounds sourced from the published drug induced liver injury (DILI) list (Ekins S, et al., 2010). This dataset was selected to challenge the in silico model (designed to predict toxicity from protein-ligand interactions), as it contains compounds with toxicity caused by non-receptor mediated mechanisms. The DILI list was processed by considering only compounds not found in the FDA lists or LiverTox, and those with non-ambiguous PubChem identifiers. Natural metabolites, natural compounds, and other endogenous compounds like hormones were also removed.

Nine out of ten compounds, classed as non-toxic on the DILI list, were predicted by CRISP as being non-toxic also (Table 3). The exception, Zomepirac, was identified as toxic by CRISP which corresponds with the drug having approval withdrawn at a later stage. This suggests the model is capable of reliably predicting safe drug-like compounds as being non-toxic.

Table 3- DILI non-toxic compound results. The CRISP percentage toxicity indicates how much each compound behaves like a member of the known toxic group or non-toxic group of compounds, e.g. it binds to the same proteins in a similar way to those identified as tox/ non-toxic.

50% of the compounds classed as toxic on the DILI list were predicted as toxic using the CRISP platform. Of those predicted as non-toxic, two are currently available on the market (Bicalutamide and Carbendazim) and two have been withdrawn from market due to toxicity (Furazolidone and Phenacetin), but not liver toxicity. However, Thioacetamide is known to induce acute or chronic liver disease and the CRISP platform unexpectedly predicted this as non-toxic. However, Thioacetamide requires metabolic activation (via two oxidations) to elicit its toxicity and the un-oxidised form of Thioacetamide, as tested in silico in CRISP, has been found to be non-toxic to isolated hepatocytes at concentrations up to 50 mM for 40 hours (Hajovsky et al., 2012). This suggests the strength of the CRISP platform lies in predicting the toxicity of compounds for which there is a direct protein-mediated mechanism for the tested compound, and not in situations where compounds are converted into other products or intermediates by liver enzymes. These limitations are associated with the design of the platform and improvements can be made by incorporating clustering methodologies that search in chemical space for derivatives and evaluate these for toxicity alongside the test compound entered. Alternatively, complementary approaches can be used to predict metabolites or if the downstream intermediates/products are known, these can be run through the CRISP platform to assess toxicity/ rodent model applicability.

Table 4- DILI toxic compound results. The CRISP percentage toxicity indicates how much each compound behaves like a member of the known toxic group or non-toxic group of compounds, e.g. it binds to the same proteins in a similar way to those identified as tox/ non-toxic.

The free version of the CRISP interface and database is available at https://moleculomics.com/CRISP/, where it is possible to view the results for over 1800 compounds. A commercial version for testing proprietary compounds can be accessed from the Moleculomics Services page (https://moleculomics.com/service-offering/) and from the Hit-to-Lead Portal (https://h2lportal.com/CRISP/) and will cater for wider use by industry and testing of new compounds of commercial interest. The platform enables users to screen a candidate compound against proteins of the human liver to identify predicted molecular interactions and compares these to interactions in other organisms, for the purpose of selecting the most appropriate animal model for experimentation. The extension of the CRISP approach to other species used as testing models, such as Caenorhabditis elegans and rabbits, is now easily achievable and could further increase the impact of this work.

CRISP will enable the elimination (or cautionary advising) of tests where the animal model used would provide a poor representation of the human biochemistry and allows the selection of more appropriate models. This simultaneously negates unnecessary animal studies and improves the efficiency of the compound development process. CRISP also identifies where liver toxicity is highly probable to be the critical hazard and provides evidence to justify removing the compound from the drug development pipeline before vertebrate testing. Removing compounds before vertebrate testing, prevents the compound going through numerous in vivo screens including kidney, liver and reproductive toxicity, saving potentially thousands of animals. For example, a two-generation reproductive toxicity study (OECD TG 416), which requires 2000 – 3000 animals (usually rats) is currently required for all agrochemical active ingredients. Each compound identified by CRISP as liver toxic and removed from development before OECD TG 416 testing, will reduce the number of animals used in reproductive toxicity assessments by approximately 2000.

CRISP will no doubt serve to further enrich interactions with large industrial companies. For example, work is ongoing with Unilever plc (toxicity screening), Air Liquide (therapeutic actions of noble gases) and with pharmaceutical companies (lead discovery). The CRISP platform initiative has also helped drive the 3DProteome platform of two key proteomes, yeast (Saccharomyces cerevisiae) and human, carried out in conjunction with the Genome and Structural Bioinformatics Research Group at Swansea University Medical School.

Moleculomics Ltd quote:

This CRACK IT Solutions award has been of great value to Moleculomics and we are grateful for the opportunity it has provided to focus time and effort advancing a technology of direct application to industry, which we have been able to disseminate immediately.

We are grateful too for the collaboration with Dow AgroSciences, and the increased exposure of our in silico platforms to relevant companies and teams in industry. Working and connecting directly with scientists and organisations, to realise the great potential of computational aids to model selection and toxicity screening is a very valuable experience for a small university spin-out company.

References

Ekins S, Williams AJ, & Xu JJ (2010). A predictive ligand-based Bayesian model for human drug induced liver injury. Drug Metabolism and Disposition 38(12):2302-8.

Hajovsky L. et al., (2012) Metabolism and Toxicity of Thioacetamide and Thioacetamide S-Oxide in Rat Hepatocytes. Chem Res Toxicol. 2012 Sep 17; 25(9): 1955-1963.