Main Challenges of Using AI for Drug Development
AI can be harnessed to speed up the processes of gathering and accessing information to dramatically shorten drug development time and keep the prices of new treatments low, which becomes critical as the cost of discovery rises. and drug development is skyrocketing, according to Michel Galarnyk, AI Evangelist on cnvrg.io.
AI can speed up the processes of collecting and accessing information to dramatically shorten drug development time and control the price of new treatments. This is becoming increasingly critical as the cost of drug discovery and development skyrockets.
How much does the development of a new drug cost? A study published in 2020 concluded that the median research and development cost of a new treatment was $985 million. A significant part of this cost was the high rate of trial failure. About seven out of eight compounds that enter the clinical trial pipeline are never developed.
This is where the AI comes in. AI has the potential to find previously unexplored patterns, not immediately, which can lead to new understanding of diseases and the drugs designed to treat them.
For example, Astra Zeneca is using machine learning models to find out more quickly which genes can cause resistance to cancer treatments and Samsung has built an app to detect early Covid-19 infection.
What stands in the way of AI
Despite the high number of successes, pharmaceutical (pharma) companies must overcome unique challenges to realize the benefits of AI. Here are some examples:
- Smaller data sets – Most AI algorithms need large data sets to learn. Due to the large number of diseases and conditions, and the relatively low number of incidences of each, creating large data sets for each type of medical condition is very difficult. Typically, for machine learning models in the pharmaceutical industry to work effectively, they need a minimum of 2-3 years of historical data. The high number of mergers and acquisitions can make this objective difficult, especially when the original source of the data may no longer be available. Because pharmaceutical data sets are generally smaller with fewer patients and fewer observations per patient, it is more difficult to obtain meaningful information.
- Complex data – At the same time, there may be fewer datasets, there may be many more features for each dataset. Patient data may include information relating to their past and current health or disease, treatment history, lifestyle choices and genetic data. It can also include biometric data, which is any measurable physical characteristic that can be measured by a sensor or wearable device. Therefore, patient data may include alphanumeric data, radiology, pathology and clinical test report images in multiple formats such as JPEG/JPG and Digital Imaging and Communication for Medicine (DICOM) format. AI systems for drug development must be able to handle varied and complex data.
Learn more: What is the impact of artificial intelligence on health?
- Complicated data labeling – Data labeling is more complex and requires highly specialized input. Consider all the different types of expertise needed to identify skeletal, internal organ, nervous system, and vascular abnormalities from X-ray images. It’s not just finding the necessary expertise that’s difficult, labeling can be cumbersome and time-consuming. Every brain scan used for cancer screening must be reviewed by doctors (often three or more), and each inspection can take 5-15 minutes.
- Data Bias – Several groups of the human population have long been missing or poorly represented in medical data sets. If the training data does not represent the entire population, there may be diagnostic errors and fatal outcomes. For AI ethics and transparency, MLOps processes must be in place, and machine learning (ML) model scoring must be established, to monitor and detect drift with a continuous feedback loop. A diverse ML team should continually test models to increase transparency and eliminate data bias from machine learning models.
- Lack of data standards – Industry needs to develop its definition of what constitutes a good dataset. Organizations may collect and use different data, have different ways of encoding information into their systems, use null or dummy data when required information is missing, and document demographic data inconsistently. Without well-defined parameters, companies always wonder how to reproduce the results of studies. Once it is clear how to construct a valid dataset, it can be more easily used by other groups to advance research even further.
- Lack of biological data – Biological databases play a central role in bioinformatics. They provide scientists with the ability to access a wide variety of biologically relevant data, including the genomic sequences of an increasingly wide range of organisms. There are many ways to use biological data, for example, by comparing sequences to build a theory about the function of a newly discovered gene, by examining known 3D protein structures to discover patterns that can help predict how protein folds or by studying how proteins and metabolites in a cell work together to make the cell work. AI for data discovery depends on these types of data to understand how patients respond to drug treatments.
- Regulation – The pharmaceutical industry is also highly regulated, requiring full disclosure and transparency for each step of the drug development process. This requirement often makes the development of pharmaceutical AI more time-consuming and expensive. Pharmaceutical companies should work with regulators to streamline this process for the benefit of all. Regulators and businesses can adopt AI and other digital transformation initiatives to increase the value efficiency of regulatory operations.
In addition to these challenges, pharmaceutical companies also face the classic barriers to implementation. For example, the need for a flexible infrastructure to collect data, verify data, run applications, provide data governance, and scale. New innovations can alleviate these challenges, including better use of transfer learning and MLOps platforms that can not only help train models, but also streamline the process of bringing machine learning models to production.
Increased competition to find treatments faster will increase demand for accelerating new drug discovery by leveraging AI. The challenges of analyzing medical data are becoming more complex, but the pace of innovation is also accelerating to develop the AI tools and technologies that will enable pharmaceutical companies to discover more relevant information faster.
What technologies can be used to simplify drug development processes? Share with us on Facebook, Twitterand LinkedIn. We would like to know!
Image source: Shutterstock