Artificial intelligence has already established a small but growing presence in the life sciences industry today, starting with drug discovery and development and now in emerging applications across the product life cycle.
While the life sciences industry is chock full of data-rich processes (which is great news for AI compatibility), AI is only just beginning to be applied to gather, manage and intelligently use all the structured and unstructured data in the domain.
Concurrently, we have seen consistent interest among our business executive readers with respect to today’s common AI trends and terms in the life sciences industry and a consensus among our interview respondents giving weight to the expectation of disruptive changes that AI will bring to their sector.
In partnership with NextLevel Life Sciences (and some of the speakers at their upcoming event), we set out to ask life sciences industry experts two questions that matter the life sciences business leaders today:
- What are the AI-related trends that life sciences executives need to know to prepare for the future and why should they should understand it?
- What AI-related terms do life sciences executives need to know to understand compatible applications?
AI-Related Trends for the Life Sciences Industry
Going in, our goal was to establish the most impactful trends affecting the AI market in the life sciences industry. We at Emerj expected a few common responses for trends from the interviewees related to AI-powered drug discovery and lack of AI-talent. We found that the experts’ ideas on the most impactful trends varied, among which include trends like commodification of data, a switch in business strategy from partnerships to acquisitions from big pharma firms and the leading role that China will play in the future.
Jan Sagal, Senior Conference Producer at NextLevel Pharma
“What Can Be Automated, Will be Automated – There are many stages of early drug discovery that are expensive to conduct, take time and are labor demanding. If pharma companies proceed through these various stages faster with automation (i.e. hypothesis to target, target to hit, hit to lead…), they will eventually save money, time, and increase success rates.
Unified Machine Learning Models – Building unified ML models can help us to predict future behavior of compounds and their performance in randomized clinical trials. There are two areas where this is achievable: predicting toxicity, predicting efficacy of combination therapies.
Lack of AI-skilled labor – Not enough talented and highly skilled AI developers are interested in working in life sciences. The industry is not being viewed as sexy, and the talented AI experts who are working in the life science space eventually become their own entrepreneurs.
Pharma companies need to become more flexible in their HR strategies and be more creative when attracting and retaining AI talent. Possible solutions include flexible working hours, pay for performance /projects or creating intramural companies on separate budgets and at different locations, especially the locale of the talent – e.g. Silicon Valley”
Michael Frank, Director, Strategy- World Wide Research and Development, Pfizer
“Commodification of Data- There are a handful of common, public data sets, such as PubMed, Clinicaltrials.gov and US patent office filing data. These have been mined and curated by many AI companies, making the data set, essentially interchangeable. Companies are differentiating themselves with their investment in the quality of curation of these data sets, as well as in their ability to stitch together orthogonal data sets such as genomic or expression data.
Table Stakes – It is now almost expected that a vendor will be applying some aspect of machine learning to their data or services. We are seeing this more and more in pharmaceutical research as well, in particular in computer vision and automated tasks. The differentiation will not be whether AI is being applied, but how well it is being applied.”
Alex Zhavoronkov, Co-founder, Insilico Medicine, USA
“Startups and tech companies take the lead in AI-powered drug discovery – Growing internal expertise in deep learning and reinforcement learning is very difficult as the top scientists in AI are interested in achieving the results quickly and there are very few deep learning experts with expert knowledge in biomedical sciences. Top AI scientists are interested in rapid publication on ArXiv and talks at major conferences. And the rate of progress in modern machine learning is outpacing the rate of progress in the pharmaceutical sciences.
It is important for executives to understand the startup landscape and learn about the most recent trends in order to adjust the partnership strategies.
Big pharma companies switch from partnerships to acquisitions – In 2017 almost every large pharmaceutical announced a partnership with one or more AI startup. The recent acquisition of Flatiron Health heralds the new trend. It is important for executives to explore the strengths and weaknesses in the startup environment.
China takes the lead – China is developing much more rapidly in AI than other countries primarily due to the heavy investments in the field and enormous interest among a very large population. Hiring AI talent in China is now more expensive than in the many other areas and there are multiple pharmaceutical companies looking to incorporate AI into drug discovery.
It is important for executives to get a more global and international outlook of the AI industry looking beyond the most common hubs like Boston and Silicon Valley.”
Take-Aways on Trends
The pharma industry is currently said to spend well over $1 billion per drug for development and it can take upto 15 years to translate a drug discovery idea from initial conception to a final product. From our interviews, it seems as though the advice coming in from the industry experts is that drug discovery and development will remain at the frontier of AI-integration in the pharma sector.
In short, the lack of success faced by many companies in drug discovery, sometimes even after years of research efforts, has meant that leaders in the industry are looking at AI technology to improve efficiency and save costs. Experts agree that we can expect a lot of startups to be active in this sector and concurrently a trend of partnerships and acquisitions by bigger pharma companies.
One of the biggest challenges for AI implementations is that the complexity involved has meant that there exists a shortage of technically skilled AI talent (our previous article with NextLevel Life Sciences covered the “AI talent gap” between life sciences and computer science). This is especially true in the life sciences industry which for many years now has not had a reputation as being relatively more ‘exciting’ than other fields such as banking, finance and marketing.
All said and done, one commonality that the experts agreed upon for life sciences executives was that the the next decade will see AI by itself not being a the great disrupter due to its prevalence, rather, how effectively companies use AI will determine the biggest winners.
Common AI-terms that Business Executives in the Life Sciences Industry Need to Know
Grasping the intricacies of AI implementation in any industry requires a basic level of know-how for business executives. This is especially true of the life sciences industry since innovation in AI-technology seems to be outpacing technology awareness levels. We asked our industry experts to share some of their thoughts on some commonly used AI-terms that business executives in life sciences need to know.
Jan Sagal, Senior Conference Producer at Next Level Pharma
“Data Standardization – The process of reaching agreement on common data definitions, formats, representation and structures of all data layers and elements (Source)
Data received in various formats is transformed to a common format that enhances the comparison process, enables use across platforms and interoperability.
An innovative pharma company generates and captures terabytes of data throughout the discovery and development of new molecular entities. When this data is combined with other sources it creates challenges for data quality and format. Companies need to clean and prepare data before it is used for further analysis.
Senior execs need to understand the importance of data standardization, as it is often the biggest obstacle to developing an effective big data analytics strategy.
Big Data Analytics – Big data analytics refers to the strategy of analyzing large volumes of data, or big data. This big data is gathered from a wide variety of sources, including social networks, videos, digital images, sensors, and sales transaction records. The aim in analyzing all this data is to uncover patterns and connections that might otherwise be invisible, and that might provide valuable insights about the users who created it. (Source)
Once data is prepared for analysis, it can fuel big data analytics engines. Companies sometimes own valuable data which are still unmined resources for potential targets and leads in drug development. Big data analytics combines different data sources and brings to light new relationships and insights which might otherwise have been left unknown.
This data strategy is often costly and takes time to deliver results and companies often fail to realize its importance for an effective R&D process.
Machine Learning – Machine learning is an artificial intelligence (AI) discipline geared toward the technological development of human knowledge. Machine learning allows computers to handle new situations via analysis, self-training, observation, and experience.(Source)
ML is a great tool for pharma R&D, assuming that the most advanced date infrastructures are in place, the right questions are asked, important goals are set, and enough training data is provided. If all these conditions are met, trained ML models can speed up exponentially the R&D process.”
Michael Frank, Director, Strategy- World Wide Research and Development, Pfizer
“Validation – With any application of AI in life sciences, a way to measure performance is essential. A commonly used tool is the Receiver Operate Characteristic Curve. This is a graphical display of the output of a model, showing the balance between the True Positive Rate (TPR) and the False Positive Rates (FPR) at every possible decision boundary. By taking the area under the curve of this ROC, you get a single number to represent how well the AI model is performing relative to a benchmark. This provides a relatively straightforward means to assess how well the AI model is functioning on the test data.
Training and Test Data – The heart of any machine learning effort, or more specifically deep learning effort, is the availability of appropriate and sufficient training data which aligns to the problem you are trying to solve. In life sciences, there is frequently a need for deep domain expertise in constructing the training set. Once the model is trained, it is validated with a set of data that has been held to the side. This sequestered data is then used to validate the machine learning, and demonstrate the degree to which it can be trusted to perform its analysis against a larger data set.
Predictive Analytics – The ability to identify trends and features before they become common knowledge. For example, a gene-target link to a disease may be identifiable via common associations long before they first occur together in a published paper. Getting at these associations and applying them to current problems can yield benefits to decision makers.”
Alex Zhavoronkov, Co-founder, Insilico Medicine, USA
“Next-generation artificial intelligence/Modern artificial intelligence – Machine learning and AI are heavily overused terms and there are many companies working in this area for decades. Most of the recent hype in AI is due to the real progress in deep learning and reinforcement learning, the modern techniques that allow for superhuman accuracy in classification, recognition, and prediction. These techniques allow for something that was previously impossible using the traditional techniques – generation of new objects with “machine imagination” using generative adversarial networks (GANs). (Source)
Generative Adversarial Networks (GANs) – Generative adversarial networks (GANs) are a class of neural networks that are used in unsupervised machine learning. They help to solve such tasks as image generation from descriptions, getting high-resolution images from low-resolution ones, predicting which drug could treat a certain disease, retrieving images that contain a given pattern, etc. GAN architecture usually involves a competition between the two-deep neural networks, hence the term adversarial. GANs provide a cornucopia of meaningful drug leads.
Reinforcement Learning – Reinforcement learning (RL) is an area of machine learning inspired by behavioral psychology and involves the reward and punishment to learn to achieve a specific objective. in RL software agents ought to take actions in an environment so as to maximize some notion of cumulative reward. This is the technique utilized by AlphaGo to learn the strategy needed to beat the human Go champion.”
Take-Aways on Terms
In recent times, AI and ML have been overused into becoming marketing buzzwords. The responses of the interviewees when they were asked to name a few of the ‘must-know’ AI-terms for life sciences leaders, we witnessed a large deviation. This is derivative of the complexity involved in creating an AI platform today.
The experts agreed that clearing away the marketing jargon and understanding the basic terms used in the AI-domain is probably the first step for life sciences executives to lay a foundation for visualizing potential AI applications that can help companies. Some of the common terms to start here would be machine learning, natural language processing, deep learning, big data and predictive analytics
What executives looking to understand AI in this sector can expect going in is that most AI implementations find data structuring very challenging. The core of most real-life AI use-cases in the life sciences industry lie in the availability of relevant training data for the configuration of the platform and useable-test data to fine-tune its accuracy.
The bottom-line here is that integration of AI technologies require an inherent knowledge of how AI can help in order to define the objective of where AI can help. Understanding how to structure or standardize enterprise data and having a capable team handling the configuration would be key steps to achieving success here.
This article was written in partnership with NextLevel Life Sciences. For more information about content and promotional partnerships with Emerj, visit the Emerj Partnerships page.
Header image credit: Adobe Stock