2023-03-16

In the past decade, the adoption of artificial intelligence (AI) and machine learning (ML) has been one of the most rapidly expanding trends across all industries. With the continual developments in technology, access to more powerful computers, increased availability of clinical and research data, and the quick development of novel algorithms that evaluate and exploit those data, the interest in applying AI and ML to healthcare is at an all-time high.

Until now, the majority of this interest has been on utilizing ML to improve healthcare delivery. AI/ML-assisted medical procedures; digital medical information management solutions that leverage AI to improve hospital processes and streamline patient experiences; and AI/ML-powered imaging systems and biometrics technology to assist clinicians in diagnosing medical disorders are a few examples. The use of these technologies to facilitate drug discovery and drug development has not received as much public attention; yet, AI/ML has the potential to revolutionize these fields as well.

Traditional drug research and development techniques are labor-intensive, time-consuming, and costly. Because AI/ML can manage huge, heterogeneous, multi-dimensional data sources, recognize patterns, and forecast outcomes, it has the potential to significantly improve the efficiency and precision of drug research and development. In turn, this will drastically cut the time and expense required to bring more effective and safer medications to market.

Realizing the potential of this technology, however, will require overcoming a number of obstacles, including issues with data quality and access, transparency of underlying development and validation processes, the possibility of bias inherent in the source data as well as the algorithm's implementation, and the absence of definitive regulatory guidance from the relevant government agencies.

The advantages and disadvantages of utilizing AI/ML in drug discovery and development are elaborated on in the sections that follow.

Possibilities for Implementing AI/ML

In large part, the success of clinical trials depends on intensive preclinical inquiry and planning – i.e., finding the most promising candidate molecules and pharmacological targets and then designing the exploratory method most likely to win regulatory approval. Traditionally, scientists must discover new drug targets through fundamental research, frequently depending on chance and trial-and-error to select the best candidate for development. Due to this, very few medication candidates that ultimately reach clinical trials are successful. It is predicted that just five out of every 5,000 drug candidates pass preclinical testing and move on to human testing, and only one of those tested on people reaches the market.

Selecting the incorrect target molecule or concentrating on the incorrect area of inquiry can cause significant delays and impede a trial from the outset, squandering valuable time and money.

Machine Learning can assist researchers in minimizing these errors by:

• Streamlining the procedure for identifying candidate molecules and drug targets and enhancing the likelihood of success with those discovered by analyzing huge volumes of existing research data and employing predictive modeling to determine probable drug target interactions.
• Synthesizing enormous quantities of biomedical data into information that is easier for people to interpret.
• Generating synthetic chemicals with desirable biological characteristics that were previously unknown
• Effectively analyzing data to elucidate the mechanism of action of a drug and increasing the likelihood that drugs will be tested on groups most likely to benefit from them.

The capacity of a corporation to identify superior therapeutic targets earlier in the R&D process can expedite the remainder of a new medicine's development. A recent analysis projected that the application of machine learning in drug discovery may save between $300 million and $400 million each successful medicine in R&D costs.

Suggestions For Overcoming AI/ML Obstacles

Despite the potential advantages of machine learning approaches in drug discovery and development, these technologies are not without obstacles.

In the case of de novo drug design, for instance, the utilized modeling techniques (i.e. generative adversarial networks and reinforcement learning) are susceptible to "mode collapse," a phenomenon in which the model generates a small number of comparable answers. The optimal outcome would be for the generative adversarial network to generate a diverse set of outputs. Nevertheless, if a generator generates a particularly plausible output (and, in fact, the generator is always searching for the output that seems most plausible), it may learn to create only that output. If this occurs, the modeling methodologies used to generate new compounds may be artificially limiting the findings, and the full potential for creating novel compounds with desirable biological features will not be fulfilled. When working with deep generative models, it is crucial to evaluate an algorithm's capacity to generate a wide range of novel structures. Adaptive multi-adversarial training (in which additional data classifiers called discriminators are spawned during training) has been shown to be effective in addressing mode collapse; however, additional work is required to develop standardized approaches that can reliably overcome these types of problems in computer-aided drug design.

Moreover, if allowed unrestricted, generative models may generate chemicals that are excessively complex or difficult to manufacture. The expected biological or physical features of a novel molecule must ultimately be confirmed in the laboratory. Researchers must take measures to ensure that their ML-generated compounds are plausible (chemically stable and synthesizable) without compromising their efficacy.

Insufficient training data and poor model calibration might result in bias in model application at a more fundamental level. All ML models' usefulness is highly dependent on the data sources employed. This is especially significant in the application of ML models in healthcare, where data is often abundant but typically derived from large academic referral centers in industrialized western nations (where people have better and more access to the healthcare system and therefore have a higher chance of their being used). In terms of illness severity and demographics, these groups are frequently unbalanced, which can lead to similarly distorted model predictions. Unsupervised learning approaches can be negatively impacted by skewed data, necessitating more investigation to establish the reach of such models. To prevent AI and ML from exacerbating the inherent bias and lack of diversity in the healthcare system, corporations must incorporate health equity into the algorithms used for drug research and development. Specific mechanisms for analyzing and resolving bias in the algorithms must be devised and implemented on a routine basis.

Inference

There is a great deal of promise for ML to improve the efficacy and integrity of medical trials, but there are still significant obstacles to overcome. Current enthusiasm for the prospective applications of machine learning in clinical research exceeds its actual utilization.