Phenotypic Screening in Early Drug Discovery: Opportunities and Challenges

Despite the prevalence of target-based discovery, phenotypic screening is emerging as a popular technique in early drug discovery. Challenges with target-based delivery have shown obvious flaws with predictive validity. The latest application of AI in phenotypic drug discovery has shown significant potential to optimise current techniques and the drug development process. 

In the search for drug candidates, target-based drug discovery (TDD) has been the method of choice for decades. TDD relies on methods with defined molecular targets, enabled by genomics and recombinant technology. In TDD, molecular biological techniques identify the expression of pathological genes which are then tested against a library of large compounds. 

Unfortunately, one of the main concerns with TDD is poor translatability. In other words, targets tested in simple cell-based assays do not present a similar profile in complex organisms, like humans. This particular challenge is suggested to be due to the inability of TDD to comprehend the widespread action of a drug on multiple targets. Selective serotonin  reuptake inhibitors (SSRIs) is a commonly prescribed anxiety medication, and an example of a drug which interacts with many targets, hence a number of side effects that arise. 

Phenotypic screening (PDD) is an alternative method of drug discovery that aims to address some of the challenges with TDD. Phenotype simply means the physical presentation of genetic code (genotype). For example, replication of chromosome 21 (genotype) results in Down syndrome (phenotype). 

Phenotypic assays test drugs within relevant biological systems or pathways to identify active biological compounds. Testing a compound within a biological pathway will reveal any target interactions which will support drug discovery in understanding potential effects of the drug within the human body. This is a point echoed in an article which infers that phenotypic assays aim “to improve the translation of drug discovery to the clinic”. Biological systems within the human body are inherently complex, hence the ability of PDD to test drugs within this environment and allow researchers to potentially identify new targets. 

The fundamental basis for the potential success of PDD is something known as the chain of translatability. This chain of translatability has been described as the “presence of a shared mechanistic basis for the disease model, the assay readout and the biology of the disease in humans”. The significance of this is that it increases the likelihood of strong predictive validity, a key part in predicting the clinical therapeutic response of a drug within humans. 

The chain of translatability is dependent on a deep understanding of the disease at a molecular level in order to best select and validate an experimental cellular system. It is important to identify the pathological molecular mechanisms within a system as this is replicated in vitro for drug screening assays.  

Examples of phenotypic screening 

For PDD at the whole animal level, phenotypic screening is modelled within organisms like fruit flies, zebrafish or mice. This helps researchers to identify the phenotypic changes on a whole body scale that arise from the molecular interaction with drugs. In addition, these in vivo approaches can highlight toxic side effects at very early stages within clinical studies, potentially saving time and money. 

Zebrafish larvae and embryos are a popular choice for phenotypic screening due to the convenience of manipulating the genetic code in high throughput experiments using a method known as automated microscopy. High throughput experiments allow scientists to test vast numbers of molecules, and then profile them against a mass of biological targets in a short timescale. The entire zebrafish genome has been sequenced in its entirety, and so is open to analysis and genomic manipulation. In terms of translatability, >80% of human disease-associated genes have equivalent representation in zebrafish. 

Genetically-modified zebra fish are used to represent human genetic kidney disorders. Kidney disease in humans is a therapeutic area with a relatively solid understanding of pathology, but limited therapeutic options. Hence, zebrafish mimicking the human genetic code for kidney diseases allows researchers to implement chemical screening for potential therapeutic drugs. This phenotype-based, whole-organism screening in zebrafish has a multitude of advantages, one of which enables the identification of potential drug candidates without prior knowledge of a validated target, which TDD relies heavily upon. In addition, it allows the simultaneous assessment “of compound efficacy, toxicity, biodistribution, and pharmacokinetics within a vertebrate model system.”

One possible limitation of PDD over TDD is the sustainability of the discovery pipeline. The extensive process from lead-finding to clinical candidate typically requires greater time and resources in PDD due to the complex screening assays. The development of these challenging assays is one of the causes for a lower probability of technical success in primary screening in comparison to TDD. Other technical risks with phenotypic screening include a high false-positive rate, and often poor results in identifying a molecule suitable for “in vivo proof-of-concept validation”. In other words, a potential drug candidate which was successful in vitro, but does not demonstrate the same effective therapeutic action in in vivo models. 

Latest innovations in phenotypic drug discovery: AI and ML

New computational methods in drug discovery have seen a substantial development in  phenotypic screening. Next-generation phenotypic screening has become an important part of early drug discovery for infectious diseases, introducing deep-learning models already popular in neuroscience. 

Deep-learning models are a class of machine-learning (ML) algorithms that utilise multiple layers to present higher-level features from raw input. One example of an ML-software platform known as HRMAn (Host Response to Microbe Analysis). HRMAn is capable of learning from host pathogens on how to differentiate and assess protein recruitment by host cells during infection. In a 2019 article it states that the “open-source image analysis platform is based on machine-learning algorithms and deep learning, and is highly flexible, as evidenced by its capacity to learn phenotypes from the data without relying on researcher-based assumptions.” This is a prime example of the advantages to using AI-driven approaches in drug discovery. The capacity of HRMAn to classify and quantify pathogenic mechanisms shows potential to enhance early drug discovery for infectious diseases. 

Convolutional neural networks are an example of a deep-learning model in neuroscience. More recently however, it has shown promising applications in oncology. In a 2020 study, a team investigated whether the AI system could classify the “sensitivity of anticancer drugs, based on cell morphology during culture”. The study aimed to determine whether cell features could be classified at a single-cell level using ML. The CNN-based model was constructed to predict the efficacy of antitumor drugs at the single-cell level. ML revealed the model identified the effects of antitumor compounds with an accuracy of 0.8. This is a huge step forward in understanding the potential of computational-based phenotypic screening in precision medicine for oncology.

Charlotte Di Salvo, Former Editor & Chief Medical Writer
PharmaFEATURES