Phenotypic Drug Discovery: Leveraging Computational Tools

Identifying essential genes in cancer cell lines is a straightforward way to classify targets and their associated biomarkers for repurposing existing therapies or developing novel ones. CRISPR/Cas9 screens in hundreds of cancer cell lines have offered hundreds of candidate targets. However, there is often a disconnect between a cell’s gene essentiality profile and its sensitivity to targeted small molecules. This brief talk will discuss one reason for this disconnect: functional buffering by paralogous genes. Small molecules often inhibit two or more members of a protein family, and only in their joint inhibition is the desired phenotype achieved. In contrast, now-standard CRISPR screens targeting one gene at a time likely have LOF phenotypes masked by paralog functional buffering.

A loss of beta-cell mass and biologically active insulin is a central feature of both type 1 and type 2 diabetes. A chemical means of promoting beta-cell viability or function could have an enormous impact clinically by providing a disease-modifying therapy. Phenotypic approaches have been enormously useful in identifying new intracellular targets for intervention; in essence, we are allowing the cells to reveal the most efficacious targets for desired phenotypes. A key step in this process is determining the mechanism of action to prioritize small molecules for follow up. Here, I will discuss how my group is identifying small molecules that promote beta-cell regeneration, viability, and function. We have developed a number of platforms to enable the phenotypic discovery of small molecules with human beta-cell activity, providing excellent starting points for validating therapeutic hypotheses in diabetes. We developed an islet cell culture system suitable for high-throughput screening that has enabled us to screen for compounds to induce human beta-cell proliferation, as well as to perform follow-up studies on small molecules that emerge from other efforts. This work identified DYRK1A inhibition as a relevant mechanism to promote beta-cell proliferation, and 5-iodotubercidin (5-IT) was shown to induce selective beta-cell proliferation in NSG mice.

The clinical relevance of in vitro assays is widely disputed, largely based on anecdotal evidence that phenotypic and target-based discovery efforts are commonly misled by the idiosyncratic nature of in vitro models or by the failure to account for target interactions. However, less attention is paid to the possible inadequacy of the quantitative models commonly used to analyze these preclinical data. Using two case studies, we transform the clinical utility of in vitrodata by analyzing them with systems-based models. Specifically, we designed algorithms for assessing: i) the diversification of cell subtypes in a tumor ecosystem (BooleaBayes); ii) the drug-induced rate of cell proliferation (DIP Rate); and, iii) the synergy of both potency and efficacy for drug combinations (MuSyC). These algorithms reduce error propagation at critical steps in a drug discovery pipeline,i.e., from model selection through assay design, execution, and data analysis opening the door for clinically impactful machine learning applications based on in vitro data. BooleaBayes, DIP Rate, and MuSyC are the centerpieces of our precision oncology platform designed to effectively leverage systems biology thinking and machine learning in drug discovery. Applied to Small Cell Lung Cancer and mutant-BRAF cancers, the Parthenon platform predicts clinically relevant therapeutics from preclinical studies, thereby reducing cycle times and increasing the probability of success.

Phenotypic drug discovery is emerging as an important present and future approach in pharmaceutical development. The bottleneck in phenotypic drug discovery is often target deconvolution after a promising hit compound, which modulates the phenotype, is found. Historeceptomics is a new, under-exploited technological approach to capturing simplified drug mechanism of action (MOA) information from big genomics and systems pharmacology data on specific drugs or drug candidates. The approach emphasizes investigation of the entire polypharmacologic ensemble of any drug candidate, specifically by taking into account the expression level of the drug’s targets in specific tissues and cells of interest. We applied historeceptomics to the problem of target deconvolution in phenotypic drug discovery. We found that re-ranking phenotypic screen hits according to the historeceptomics score representing the affinity of each hit for targets multiplied by the expression of those targets in the cells used in the screen resulted in unrecognized, significant drug targets being unveiled. In addition, the use of historeceptomics enables statistical significance of the top ranked hits to be calculated. Application of the approach to a published repurposing screen for drugs preventing Zika virus neuro-injury revealed CDK5 as a new, proprietary drug target for preventing this disease.

Large-scale screening in larval zebrafish has revealed neuroactive compounds and protein targets for multiple behavioral phenotypes. Conventional efforts encode zebrafish behavioral videos by bulk motion over time and compare these “motion index” drug profiles by correlation distance metrics. Using a nearest-neighbor logic, a researcher may then use landmark drugs to infer the targets of phenotypically similar compounds. Here we explore two enhancements to this discovery pipeline. In the first, we develop a learned distance metric to relate drug-like molecules to each other by their behavioral profiles using Siamese Neural Networks. The learned distance metric substantially outperforms correlation distance and generalizes well to datasets collected months apart. In the second enhancement, we combine cheminformatic target prediction with statistical enrichment calculations to generate mechanism of action hypotheses directly from phenotypic screening datasets. This procedure reveals sensible but also unexpected protein targets, prospectively confirmed in vitro and in vivo. Together, these techniques demonstrate a means to extract otherwise hidden pharmacological signal from existing and archival phenotypic screens.

The current pandemic highlights an acute need to develop fast therapeutics against health threats. Traditional approaches to drug development are expensive and slow to react to pandemics. AI tools on the other hand have the potential to accelerate and transform this effort, enabling rapid, large scale search and identification of effective drugs. In this talk, I will showcase two of my recent efforts in addressing this challenge:

First, I will present application of deep learning to antibiotic discovery (Stokes et al., Cell 2020). We trained a deep graph convolutional network to predict antibacterial activity of compounds. We performed predictions on multiple chemical libraries and discovered eight new antibacterial compounds that are structurally distant from known antibiotics and display bactericidal activity against wide spectrum of pathogens.

Second, I will discuss our recent effort on developing effective property prediction methods to accelerate the search for COVID-19 antivirals. In particular, I will introduce a novel approach to learn property predictors that can generalize or extrapolate beyond the heterogeneous data (e.g., different COVID-19 screens). To test the method, we use a combination of three data sources: SARS-CoV-2 antiviral screening data, molecular fragments that bind to SARS-CoV-2 main protease and large screening data for SARS-CoV-1. Our predictor outperforms state-of-the-art transfer learning methods by significant margin.

The COVID-19 pandemic has created unprecedented stresses for therapeutic development. With hundreds of thousands perishing in only a few months, there is great interest in discovering pharmacological treatments active against the SARS-CoV-2 virus; however, extant preclinical systems to study the virus and disease have a high risk of failure in human translation, with most in vitro work being done in an immortalized African green monkey cell line (Vero E6) and few-to-no in vivo models in the early phase of the pandemic.

In this talk I will describe the work done at Recursion to discover therapeutics for COVID-19. Recursion applied deep learning-based image AI to rapidly develop a human-cell-based assay for SARS-CoV-2 infection, applying phenotypic screening to deliver value even with the paucity of knowledge regarding virus-host interaction biology and targets. We subsequently deployed this assay to screen thousands of approved and clinical compounds for therapeutic benefit with minimal staff in only a few weeks. The utility of the platform will be demonstrated with recent results from our screening work.