Beyond the buzzwords
From AlphaFold's Nobel-winning protein-structure predictions to popular generative AI tools like ChatGPT, artificial intelligence (AI) has powered some of the most amazing technological advancements of our time. As it reshapes industries worldwide, executives are facing pressure to “implement AI” or risk being left behind. But what does implementing AI actually mean?
AI is not a single monolithic technology, nor is it a silver bullet. It is not even a well-defined set of capabilities. In fact, the term 'AI' has become so overused that it risks losing its meaning entirely and obscuring the real value it can provide. To navigate this fast-changing landscape, companies need to look beyond marketing buzzwords, and focus on solutions with a demonstrated ability to solve clearly defined business problems.
AI is not a single monolithic technology, nor is it a silver bullet. It is not even a well-defined set of capabilities. In fact, the term 'AI' has become so overused that it risks losing its meaning entirely and obscuring the real value it can provide. To navigate this fast-changing landscape, companies need to look beyond marketing buzzwords, and focus on solutions with a demonstrated ability to solve clearly defined business problems.
The evolution of AI techniques
Artificial intelligence refers to machines doing tasks that mimic human intelligence, such as recognizing images, interpreting language, automating processes, and making decisions. It encompasses a vast set of tools and technologies, each suited to solving different types of problems. Understanding the evolution of these technologies - how they work, as well as strengths and limitations - is useful in helping biopharma companies understand how AI might best work for them.
Over the last several decades, AI science has progressed through several key phases:
Over the last several decades, AI science has progressed through several key phases:
- Expert Systems and Heuristics
Many early AI systems relied on explicit rules and logic designed by human experts and programmed directly into the system. A classic example is Dendral, developed in the 1960s to help chemists identify organic molecules based on mass spectrometry. Dendral used pre-programmed heuristics, such as the conservation of mass during molecule fragmentation, to replicate and automate chemists’ decision-making processes. Its success demonstrated that machines could match human experts in specific, knowledge-intensive tasks. But expert systems depended on pre-programmed “knowledge”, limiting their ability to handle broader and more complex problems. - Machine Learning (ML)
By the 1990s, AI began shifting to models that “learn” directly from data, instead of relying solely on pre-programmed rules. ML techniques like linear regression, K-means clustering, and support vector machines (SVMs) work by finding patterns in their training data and making predictions based on those patterns. For example, linear regression can be used to correlate environmental factors with product yield in biomanufacturing, or to understand how different risk factors affect health outcomes in epidemiology, while SVMs, which can identify subtle patterns in high-dimensional data, are used to diagnose cancer through analysis of gene expression profiles.
These early ML methods (including many that are not mentioned here) are mathematically well-understood, computationally efficient and handle structured, numerical data very effectively. One way to loosely visualize how they work is to think of them as plotting their training data as points in a multi-dimensional space and then finding relationships between those points - whether it is drawing a best-fit line through them, segmenting them into clusters of close data points, or dividing them into two groups in a way the maximizes the distance between them. From fraud detection to recommendation engines to cybersecurity, these ML algorithms are broadly used across many different industries. - Neural Networks (NN) and Deep Learning
A major ML breakthrough came with neural networks, whose structure is inspired by the human brain. NNs are made up of layers of connected nodes (neurons) where the strength of the connections between the nodes (synapses) changes with each piece of training data. One can think of the input layer nodes as representing different features or characteristics of the data, and the output layer nodes as different classifications or predictions for the data. Often, layers of intermediate nodes are also used, to capture different types of patterns in the data. The term “deep learning” simply refers to the use of these hidden, intermediate layers. Neural nets are able to handle data that is less structured and more complex than the numerical data that earlier ML models excel on - data such as natural language, images, and video. AlphaFold is powered by deep neural nets, as are medical imaging programs that help detect and diagnose lesions. - Generative AI and Large Language Models (LLMs)
Over the past two decades, advances in neural network architectures, along with the availability of massive datasets, have enabled NNs to scale by many orders of magnitude, greatly boosting their ability to recognize patterns at many different levels of abstraction. Large Language Models (LLMs), which are neural networks with hundreds of billions of nodes, now power generative AI (GenAI) services like ChatGPT or Midjourney that can converse intelligently in natural language or generate images from natural language prompts in a way that feels incredibly similar to creative human thinking. And yet when simplified to the most basic level, LLMs, like the ML models before them, are fundamentally just taking data sets (as digitized text, audio, and video are all simply strings of 0s and 1s) and finding patterns in them.
Different techniques for different challenges
As AI continues to evolve, looking at it as a diverse toolbox is critical to leveraging it successfully. Just as mRNA is not the right technology for every vaccine or cancer treatment, generative AI is not the right technology to solve every business problem.
Each AI technique has its unique strengths and weaknesses. LLMs, for example, work well with unstructured data. But they require tremendous amounts of data and compute power, and the resulting models are so large and complex, it is difficult for scientists to understand exactly how they work or how well they work. They shine in creative applications but struggle with tasks requiring precision or detailed domain knowledge, as they.are known to hallucinate, botch simple calculations, and confidently spew logical inconsistencies. In contrast, techniques like K-means clustering and linear regression require highly structured data with well-defined features, but can work effectively with small data sets, making them ideal for attacking problems such as batch process optimization or supply chain optimization.
In addition, when it comes to solving real-world business challenges, implementation details matter. The choice of parameters, the type and quality of training data (including biopharma domain-specific data), and the use of relevant heuristics are just a few “details” that can have a huge impact on model efficacy. The user interface, data integration touchpoints, and security capabilities of an AI-driven solution are also critical factors.
Each AI technique has its unique strengths and weaknesses. LLMs, for example, work well with unstructured data. But they require tremendous amounts of data and compute power, and the resulting models are so large and complex, it is difficult for scientists to understand exactly how they work or how well they work. They shine in creative applications but struggle with tasks requiring precision or detailed domain knowledge, as they.are known to hallucinate, botch simple calculations, and confidently spew logical inconsistencies. In contrast, techniques like K-means clustering and linear regression require highly structured data with well-defined features, but can work effectively with small data sets, making them ideal for attacking problems such as batch process optimization or supply chain optimization.
In addition, when it comes to solving real-world business challenges, implementation details matter. The choice of parameters, the type and quality of training data (including biopharma domain-specific data), and the use of relevant heuristics are just a few “details” that can have a huge impact on model efficacy. The user interface, data integration touchpoints, and security capabilities of an AI-driven solution are also critical factors.
How to best leverage AI: key takeaways
For companies looking to integrate AI, two key takeaways stand out:
- Go beyond the buzzwords
Instead, focus on evaluating how well the solutions address your specific business challenges and impact your bottom line. This is far more important than which AI technologies may or may not be found under the hood. - Data is key
The power of AI lies in its ability to find patterns in data, which means its results are only as good as the data it has access to; even the best AI technology is not as effective without timely access to comprehensive, high-quality data. In biopharma manufacturing where data is traditionally “trapped” - on paper, in spreadsheets, or across various siloed systems - building a robust data infrastructure that consolidates, cleans, and provides timely access to operational data is the most critical investment a company can make to move towards successfully leveraging the full potential of AI.
