Two people meet by chance on the footbridge that spans the lagoon in Boston’s downtown public garden. One is a distinguished biologist, the other a passionate exponent of artificial intelligence. As colleagues at Cambridge Consultants (CC) in the nearby Seaport district, their initial small talk centers on the coincidence of their lunchtime meeting. But discussion soon turns to work. And the exchange that unfolds becomes the springboard for a collaboration with significant implications for life-saving cell and gene therapies. The biologist is Karen Weisinger, the AI expert is Ram Naidu, and this is the conversation that is bridging the gap between bioprocessing and machine learning…
Karen Weisinger:
It was great to run into you that day Ram, I think very quickly we both realized that by coming at the problem from our own different angles we might spark some real progress. As a biologist, I’m excited by the emergence of cell-based therapeutics. I mean let’s be clear here, the notion that modified cells can be transplanted into our bodies to cure disease is one of the most thrilling and important developments in medicine. But like many others, I’m frustrated that demand is being hampered by the reality of high-speed manufacturing costs that are just way too high.
Ram Naidu:
You’re right Karen, it was a meeting of minds that day for sure. It’s obvious that development of these novel therapies at scale requires fundamentally different approaches than small-molecule drug manufacturing. For a start, these are living therapies! I’m really pleased that our collaboration is already showing that it’s possible to apply machine learning to speed up and improve the efficiency of quality assurance in scaled biological processing.
Karen Weisinger:
Speeding up quality assurance for scaled bioprocessing – it’s got a great ring to it hasn’t it? And you’re right to say that dealing with living therapies presents very particular obstacles. We must isolate cells from patients or donors, manipulate them and help them to grow and multiply outside the body to produce a living product at a therapeutically relevant dose. To complicate that, batch sizes for manufacturing are typically small. Live cells are not standardized or well characterized – and cultivating them is complex, labor-intensive and demands skills that take years to develop.
Ram Naidu:
All of which brings us to the nub of the challenge of course. Expanding access to cell-based therapeutics to a meaningful level that will have true impact for patients across the world depends on achieving a scalable, automated manufacturing process. We have to find a way to link the exciting new developments in the laboratory to commercial therapies that can change lives for the better.
Karen Weisinger:
For sure – and these are the themes that drove our big picture thinking early in the project. How do we reproducibly manufacture a highly complex living therapy for all of the patients who may need it? How do we confirm that that the product is safe and effective and that the manufacturing process we used didn’t encounter any obstacles along the way? It was clear to us that current processes just don’t cut it. Right now, we have skilled experts ‘in the loop’ at every stage. They’re mired in manual, impractical steps because we have yet to harness the possibilities offered by advanced technology.
Ram Naidu:
Precisely. And as with many complex problems, we can boil the thrust of the challenge down to a relatively small number of fundamental questions that need to be addressed. Vital things like, how much can we practically automate while remaining confident that quality can be assured? How can we make sure that human experts are kept informed – and know when to step in when required? And – crucially – how can we monitor progress in near real time?
Karen Weisinger:
We found that the best way to understand what automating the bioprocess for cell therapeutics actually involves was to observe our scientific colleagues as they carefully nurture cells through the journey. At each step, they characterize process sterility and assess cell viability through spot measurements – as well as assessing rates of change of various performance indicators. They characterize each step at multiple levels, including the process and cellular level. All the steps and measurements are specific to the process and have to be empirically discovered and thoroughly assessed before they can be used as established critical process parameters (CPPs).
Ram Naidu:
I find this bit fascinating – remind me how you infer risk in the process Karen. Are there different ways and specific measurements that are used?
Karen Weisinger:
Sometimes scientists infer risk to viability through measurements of the metabolites that are generated. Other times, they run nucleic acid assays, protein-based assays or use sequencing data. The whole time, they use the subtle logic and deduction to continuously determine whether they should intervene and execute preventive and corrective actions. Or not.
Ram Naidu:
Understanding how information sits in this layered problem holds the key to automating what is extracted, synthesized and actioned upon. The solution requires more than a conventional control system. Any solution must respond actively and reliably to unexpected deviations and learn from them. The solution must subscribe to being part and parcel of a human-machine team that monitors and optimizes performance under all conditions. This is easier said than done.
Karen Weisinger:
We realized that such a system requires the generation of insights at multiple levels of the manufacturing bioprocess. Enabling these insights requires the intersections of disciplines – biology, advanced in-line, contactless sensing, and artificial intelligence that generates real-time prescriptive insights in a dynamic, adaptive control system. In-line sensing allows us to track changes in near-real time. Ideally, we would like this to be as direct and as minimally disruptive as possible. But in-line sensing poses the risk of contamination. So, we need to enable contactless methods, possibly at-line.
Ram Naidu:
Non-invasive sensing sounds great but the results are a set of numbers not directly interpretable by a biologist. So, a step of translation is required to convert measurements at various levels into insights that correlate the quality of the current stage in the process with the desired outcome such as number of healthy cells. Again, AI can create the link between a comprehensive set of measured variables and a set of process parameter values that need to be updated as necessary. Putting all this together enabled us to formulate what we called our blueprint for the bioprocess of the future.
Figure 1 - The illustration compares the current approach with our vision for when AI is incorporated into the bioprocess. With no AI (above), the bioprocess is dependent on highly skilled technical expertise. This is not the best use of the scientists’ time. They can become disaffected and even leave to pursue other opportunities – to the detriment of the company. As well as depending on such expertise, the analysis is time-consuming and prone to operator error. It all adds up to increased costs, time and effort, as well as the potential for compromised quality.
Figure 2 - When the process is monitored by AI (above), human expertise is working on novel processes to train and develop ML – and to make crucial, final decisions. This is much more stimulating mentally. Scientists are liberated to devote more time to improve the company strategically, while productivity and talent retention are optimized. Bioprocess analysis and validation is faster and more accurate. AI integration minimizes the need for highly skilled individuals to perform mundane tasks, allowing them to be innovative in additional areas. Product development, product quality, productivity, working environment and company culture all benefit.
Karen Weisinger:
The key question here is what does it really mean to use AI in this process? AI and machine learning have become buzzwords that are thrown in the mix at every opportunity. Everyone wants to or claims to use AI/ML in their system. But we must demand something more tangible – what does that actually mean and what can it do to make our bioprocesses smarter and faster? More than that, how do you actually incorporate AI/ML into a system? For us mere mortals (that is, non-AI people), perhaps it’s worth thinking of AI/ML as Design of Experiments (DOE) on steroids!
Ram Naidu:
Let’s unpack that thought. Specifically, the AI system is a set of mathematical models that can analyze measurements to generate insights about the state of the bioprocess at any instant. The insights can further be interpreted to determine a set of recommendations that allow parameters of the bioprocess to be adjusted in order to keep the manufacturing process on track.
To do this, the models have to be trained to understand which data are representative of a healthy process and which are not. The data must be labelled or tagged by an expert to associate the measurements with known outcomes and corresponding process parameters. This is hard work, but it is well worth the effort. As part of the training process, the model is evaluated on these well-characterized data sets and its parameters are adjusted to allow it to correctly associate the data with input parameters and output quality. The training process relies on carefully curated data, data hygiene and sensor calibration. There is also a set of guard rails or exception handling that protects the process from incorrect interpretations when sensor measurements are noisy or corrupted, trends are unusual or erratic, all of which require higher level review by skilled experts.
Karen Weisinger:
Which brings us to a key question: what kind of experiments inform the training process? In our biology labs at CC, we are looking to answer exactly this by creating a proof of concept. The goal is to demonstrate that we can teach AI models to identify patterns in biological processes and create insights and recommendations to ensure output quality.
Ram Naidu:
Indeed. First, we grow cells (suspended/adherent) using a manual process. At each step along the way, our biology scientists measure cell properties, collect microscopic images, describe what they use as markers of a health process at a macro, micro and cellular level. They work with our machine learning scientists and engineers to label these data and identify changes that they detect and link these to insights. The data and insights are used to teach algorithms to consume sensor data, process them to extract information and classify a state of the process as being on track, borderline or high risk. The state is then used to recommend adjustments to process parameters, additions of reagents, and cell manipulations.
Karen Weisinger:
Machine learning captures information from a given bioprocess and makes use of algorithms to translate this to useful data. It does this in a massively parallel manner and therefore, at a much higher speeds than any human could. The overall result is potentially game-changing… a lot of useful data in a very short amount of time. The progress we’ve made is really reshaping our vision for the future of both ML and bioprocessing – but let’s be clear, the current state-of-the-art is only really in its infancy. Our work will continue, the bridge building has only just begun!
It looks like having AI alone is not enough. We still need to have a good characterization of our biological system to decide on the appropriate CQAs (critical quality attributes) by which the AI will measure the system. Plus, a good characterization will help define our CPPs that can ensure that we drive the process toward our CQAs.
Ram Naidu:
We currently think of CQAs as being indicative of the quality of the process at many levels. But, figuring out what quality metrics are relevant and what are not, isn’t easy by any means. Again, AI can potentially help create insight here. There are methods that can look at metrics and identify those salient ones that represent the process but are also responsive to changes in parameters. Do you think this is helpful in aiding scientific experts?
Karen Weisinger:
What this means is that part of the experimentation that we are performing to inform our bioprocess development, must then be incorporated into the ML process. These separate concepts, to be efficient and informative, must happen in parallel and inform each other. Another thought just popped into my mind. Very often we biologists use unbiased assays to characterize our bioprocess. Unbiased assays refer to a methodology where we question our system in a way that is not hypothesis driven.
One does have to have a general question, but this is a typical approach if we do not want to be confined by our current way of thinking, or if the process we are questioning is new or not well documented. This is very often the way we define our CPPs and our CQAs when dealing with a novel bioprocess. I’ve heard of supervised and unsupervised ML, is this a parallel to what we biologists call unbiased? Where does that come into play – is it relevant in this case?
Ram Naidu:
A great question! Remember how we discussed labelling the experimental data to characterize the signature of a good scientific process? AI that learns from labelled data is what we call supervised learning because we have experts telling us what is good and what is not. Alternately, it would be great if we could use AI methods to learn patterns in the data without any labels. Such unsupervised learning from unlabeled data would reduce the burden on our scientists. They are already so busy being ‘in the loop’. But they would need to trust the output of the process. I think in this case, we could start with carefully labelling a small data set and then use AI to iteratively label the rest of the data. We call this process semi-supervised learning, and it would be great to test this out on the data that you are collecting in the biology lab.
Karen Weisinger:
Wouldn’t it also be fantastic if we had a system where the ML algorithm could know enough about the bioprocess that it could trigger the autonomous tweaking of the system? That way instead of measuring the CQA at the end and seeing if it meets our standards, we could constantly be tweaking to meet that wanted CQA.
Ram Naidu:
This is exactly the direction in which we should head. It is no longer a pipe dream. The best way to progress along this journey is to start with what we have, experiment and learn as we go and refine our approach over time – an agile methodology, if you will.
Karen Weisinger:
The way I see it we need to pull together a large training set to achieve all we want. As a vision for the future, I see the decentralization of information where we all combine reliable training sets for ML applications. Just imagine the implications of that and what it would mean for the future. I told you this is one conversation that will run and run!
