C34 An AI drug discovery case study establishing new neuroprotective compounds for treating ALS
27-11-2019
MNDA SYMPOSIUM PERTH DEC19 PRESENTATIONS
Session 4 Therapeutic strategies and machine learning
BenevolentAI, London, United Kingdom, SITraN, Sheffield, United Kingdom
Background: Previous work has shown Nilotinib protects motor neurons in co-culture with patient-derived iAstrocytes, likely due to the inhibition of multiple kinases. Nilotinib has negligible CNS exposure, so achieving sufficient target engagement in patients with this molecule will be a significant challenge - novel compounds with improved properties are desirable.
Objectives: Predictive and generative machine learning models were used to design novel compounds which demonstrated robust protection of motor neurons, and with desirable properties for an ALS therapy. These compounds were characterised in vitro and in vivo, with the aim of identifying a clinical candidate for ALS.
Methods: Compounds identified were synthesised and tested in an assay which measured the survival of induced motor neurons in the presence of astrocytes induced from the fibroblasts of ALS patients. The lead molecule was also tested in in vivo models of target and mechanism engagement.
Results: A curated training set of experimental compound data was generated, and used to train predictive models. This model was used to power a generative algorithm, which designs novel drug-like compounds which are predicted to bind key kinases of interest, are in unexplored regions of chemical space and have significant CNS exposure. Key compounds were synthesised and tested, and lead molecules were discovered in Key compounds were extensively profiled, and shown to rescue motor neurons to an equivalent extent to Nilotinib in coculture. These compounds demonstrate improved pharmacokinetic properties, including improved solubility, brain:plasma ratios of ~0.7–1.0 and robust target engagement in mouse brains following a single oral dose. We show that although inhibition of protein phosphorylation downstream of ABL1 correlates with motor neuron rescue, selective inhibition of ABL1 is not sufficient for motor neuron rescue.
Discussion and conclusions: We have shown the strategy described above is highly effective for the discovery of novel, optimised compounds for CNS drug discovery. By optimising multiple parameters in parallel, algorithms are able to quickly identify optimal regions of chemical space - by surfacing these to experienced medicinal chemists, we have demonstrated the potential to increase the efficiency of the drug discovery process. In under 6 months we discovered novel, efficacious and highly brain penetrant compounds, which have been characterised in vitro and in vivo. This series is currently progressing towards clinical candidate, with the aim of entering clinical trials in 2020. The utility of this approach has wider implications for drug discovery in ALS, and neurodegenerative diseases in general.
C35 Machine learning to accelerate drug discovery: a novel small molecule rescues ALS phenotypes in preclinical models
Verge Genomics, South San Francisco, CA, USA (shortened version)
Background: There has been little progress made in developing effective ALS therapies. Although there are promising treatments in clinical research today, most are limited to narrow subsets of ALS patients and likely cannot be applied to the broader ALS patient population. Leveraging machine learning (ML) on human patient transcriptomic data, Verge Genomics has developed a platform to evaluate ALS in a patient-centric manner, allowing for the dis for the discovery of novel gene targets that could lead to meaningful disease-modifying therapies for patients.
Objectives: Evaluating gene expression between sporadic ALS patients (no known mutation) compared to healthy individuals, we identified a gene network that was robustly down-regulated across multiple ALS patient cohorts, we then predicted gene targets that can reverse this network, and validated disease models in our own labs. After conducting a comprehensive evaluation of 10 gene targets and 200 compounds, we identified two classes of compounds that demonstrate rescue in ALS patient-derived motor neurons and rodent models. Novel chemical structures were synthesized based on these compounds towards the discovery of lead candidates for clinical development.
Results: VRG50014 is a novel brain penetrant small molecule lead candidate discovered at Verge Genomics that rescues in ALS phenotypes. Using human-derived cell lines, we evaluated survival in iPSC MN and HEK-293/TDP43 lines. In both models VRG50014 demonstrated significant rescue in cell survival compared to untreated controls. Likewise, in the rodent model, chronic treatment with VRG50014 demonstrated measurable rescue in both CMAP and pNFH, promising ALS biomarkers. Additionally, we observed transcriptomic rescue with changes towards healthy or wild-type groups in the ALS-patient derived motor neurons and murine TDP43 model.
Discussion and conclusions: By leveraging ML for gene target discovery, Verge Genomics progressed from target identification to nomination of a novel lead, VRG50014, in under 18 months. More importantly, VRG50014 rescues the dysregulated ALS patient gene signature in ALS in-vitro and in-vivo models making significant progress towards a clinical candidate, with plans to file an IND in the coming year.
Source: Abstract Book symposium Perth
![Share](/sites/all/modules/addtoany/images/share_save_171_16.png)