How Can Pharma Maximize AI Opportunities While Mitigating Risks?
“Generative AI (GenAI) models are prone to error, because they have no concept of truth or facts with the content they generate.”
A lot is possible with AI, and capabilities of AI models are expanding exponentially. By some very focused measures, we’re even seeing AI models outperform humans. (See Figure 1)1
Life sciences companies are embracing AI to accelerate the development of innovative therapeutics. Seeing an opportunity, commercialization leaders are exploring ways to harness AI to enhance their ability to deliver therapeutics in the marketplace. But most leaders have unanswered questions and concerns about how to deliver on the misunderstood promise of AI—namely “How do I get more with less?”
The AI Risk-Reward Profile
Looking beyond the possibilities, we need to understand the risks associated with intelligence that’s, frankly, artificial. These are some of the biggest risks:
• The models are only as good as their training: For machine learning models to work, they require comprehensive data sources and training led by experts. Well-known issues and errors2 have occurred in healthcare that life sciences can’t afford to replicate.
• Even when trained, they make stuff up: Generative AI (GenAI) models are prone to error, because they have no concept of truth or facts with the content they generate.
• Copyright issues: Just because Midjourney—a GenAI program that creates images from text descriptions—can produce an image, doesn’t mean you own it. Content that is created by GenAI has copyright and trademark issues that must be addressed.
• Be careful with what you feed AI: If you create a free account with these models, you should assume that anything you type or upload will become part of the training data.
Figure 1
Evaluating AI
How should commercial leaders evaluate solutions to deliver the benefits of AI while minimizing the risks?
• The AI basics
Before exploring how to apply AI, let’s start with a foundational set of definitions:
• Machine learning: A type of AI that enables systems to learn and improve from experience without being explicitly programmed. Examples include content recommendations from your Netflix app and autocomplete that suggests words while you’re typing a text message.
• Deep learning: A subset of machine learning that uses neural networks3 to learn complex patterns from data. Examples include the AI used in autonomous vehicles and AlphaFold4 that predicts protein folding.
• GenAI: A type of AI that has learned to create content in formats such as text, images, audio, video, and code.
What are some of the strengths and challenges with these types of AI? (See Figure 2)
“Active research, experimentation, piloting, and hands-on learning to deliver on what’s here and what’s coming is a requirement.”
• Where could you apply these types of AI?
This is where you could apply AI, not necessarily where you should. (See Figure 3) There are solutions out there that use AI to perform these tasks, but that doesn’t mean they work solely because they use AI.
• Trust but verify
Brand teams are under a lot of pressure to deliver in competitive marketplaces with time and cost pressures. Given these dynamics, it’s easy to want to believe that AI can perform magic. But like most things, it’s not that easy. There are many companies that have invested in AI solutions that missed expectations or failed to deliver.
However, companies will put themselves at risk if they choose to sit on the sidelines and watch as competitors put together the pieces and forage ahead. Active research, experimentation, piloting, and handson learning to deliver on what’s here and what’s coming is a requirement. Here are some principles for assessing AI solutions:
• Demonstration: This won’t be applicable for all solutions but, where possible, look for demos and hands-on access to platforms that allow your teams to perform experiments with the technology. This is possible for most gen AI platforms. Admittedly, it’s a little trickier for solutions that don’t have a customer-facing user interface.
• Validation: This requires solutions providers to provide verifiable proof points with the capabilities they’re selling. If their wares are so good, then it should be easy for them to provide evidence to support their pitch.
• Verification: Review and scrutinize the results generated by AI solutions for quality and accuracy. How you do this will vary depending on the proposed solution, but you should consider live demos, pilots, and other methods that allow you to vet the technology before you scale it.
Figure 2
Figure 3
It’s Time for Real AI
The life sciences industry is doubling down on AI to enhance its ability to bring innovative products to market that will change patients’ lives. Commercialization leaders have an opportunity to find focused ways to implement this technology to ensure they bring effective solutions to market.
To support our clients, Real Chemistry is focused on something we call Real AI. It’s not a platform nor is it something we sell to clients. It’s a commitment that we’ve made to improve the lives of patients by discovering and proving new ways to use AI in our work. We’ve taken a page from life sciences with our own version of clinical trials to deliver innovation supported by proof points.
Our Phase 1 trials are ideas we have on where we could apply AI. Our Phase 2 trials are internal experiments to prove or disprove our ideas. When we see a signal in Phase 2, we enter Phase 3 trials that are conducted transparently with clients. Not every trial ends perfectly, but step-by-step, we’re working toward perfecting our capabilities with each and every trial.
So where are you on this journey and what do you see as the opportunity for your company? Connect with us at realchemistry.com/contact to learn more about Real AI.
References
1. https://ourworldindata.org/artificial-intelligence
2. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients | Critical Care Medicine
3. Neural network (machine learning) – Wikipedia
4. AlphaFold – Google DeepMind
