March 12, 2018 PAP-Q1-18-CL-011
As the hype around artificial intelligence (AI) has been growing, its promise to change healthcare has started to materialize: medical imaging classification algorithms have bettered a panel of experienced radiologists; new in-silico developed drugs are moving through clinical trials; wearables and smart phones produce data helping with mental disease treatment; and medical decision-support systems pave the way to revolutionize the entire healthcare provider model.
In biopharmaceutical manufacturing, however, we need to search deeper to recognize the victorious arrival of AI.
Some low-hanging AI, or rather Machine Learning, fruit are ripe enough to fall in the hands of COOs and Directors of Manufacturing; machine learning-backed visual inspection and bacterial culture yield optimization are good examples. As with many other potential opportunities for the application of machine learning to achieve a predictable, high-quality, flexible, low-cost manufacturing process, we first need to look at the two integral components: the data and the data scientists.
Too often we collect manufacturing data with the hope of never seeing it again, i.e., for regulatory compliance purposes. The structure and content of this data is typically not designed for manufacturing optimization. As a result, the data generally does not help data scientists to clearly see how machine learning classification and predictive models can produce practical improvements.
Mixed paper-electronic formats are another obvious issue. Even if the API and finished dose batch records, exception reports and development and QC analytical testing results are all in electronic form, however, there is a lot of semantic/relational database preparatory work needed before Natural Language Processing (NLP) algorithms can produce actionable insights.
Great value would come from combining data sets from different manufacturing organizations within a larger biopharma company, but business unit and facility-level data silos and the lack of clear data ownership do not make this job an easy one.
The lack (or complete absence) of data scientists with dedicated machine learning training within Ops/manufacturing groups is another challenge, but one that is at least easy to explain. The AI field is now too hot, paying too well and advancing careers very rapidly. Talented candidates with machine learning-focused Ph.D. and MS degrees are being snapped up by well-financed tech startups and businesses like Google’s Deep Mind.
To make machine learning effective in biopharmaceutical manufacturing, gargantuan internal mindset and business process change around data collection, analysis, and use is necessary. There are, however, some quicker Band-Aid solutions that can be initially deployed.
Machine learning automation tools, for instance, perform reasonably well the activities of data cleaning, feature engineering, model/algorithm selection and hyper-parameter optimization. In-residence drug development and manufacturing scientists, who might not have deep machine learning expertise but surely possess robust statistical analysis experience, can use these tools to take an initial dive into the data to determine if any practical insight can be gained.
Once researchers have initial success with such tools, the Ops/manufacturing group should be in a better position to justify hiring an external data science consultancy or tap into in-high-demand corporate data-science talent. Whichever path is chosen, bringing in the experts now might be very efficient, for they can initially work on optimizing existing models, rather than starting from scratch.
Progress in NLP has also been accompanied by the introduction of text analysis automation packages, which are particularly effective when applied to documents written using limited-vocabulary, industry-specific terminology and where the relationships among the documents can be easily described using a graph.
An interesting AI approach for the biopharmaceutical industry, that requires less structural change in data collection routine might be application of Internet-of-Things (IoT) sensors, depth cameras for collecting and analyzing visual object movements, sound and speech recognition systems — all of the tools that allow shifting data collection in the background of manufacturing operations. In addition, new developments in deep neural network science, specifically Capsule Nets, have significantly improved the effectiveness of object visual recognition and classification from different angles. These neural nets can be trained using much smaller data sets, which is critical for environments with unique manufacturing processes.
The abovementioned solutions, although lacking the effectiveness of data-driven re-engineering that “smart manufacturing” might eventually require, are not disruptive to a high-quality cGMP-compliant manufacturing organization. They should be of the most interest to those companies that see manufacturing as an important value-creating activity, such as contract manufacturing organizations, generic drug producers, and biologic drug substance manufacturers with high costs of goods.
There is one other recent event that brings optimism for AI use in biopharmaceutical manufacturing, albeit one that is likely grabbing more attention from AI/machine learning aficionados. In December 2017, one of the most respected AI scientists and promoters, Dr. Andrew Ng of Stanford University, launched a company targeting specifically and exclusively the application of AI in manufacturing. For those of us, who see biopharmaceutical manufacturing as a passion and a trade, it is a precursor to the opening up biopharmaceutical manufacturing to AI - and the race is now on.
Over the next year, MILS-Group will dive deeper into the most practical cost-cutting, quality enhancing, predictability-improving applications of AI in biopharmaceutical manufacturing. We will explore applications being pursued by Big Pharma, analysis of data generated by automated production systems and much, much more.
Mr. Loghinov is the founder and managing director of MILS - Group LLC (stands for Machine Intelligence for Life Sciences), a research-driven company that supports innovation in the BioPharma Industry through application of Artificial Intelligence. Previously, Constantin worked at a leading biopharmaceutical contract manufacturing organization, as well as a business growth consultant for biopharma and medical device companies. Mr.Loghinov earned an MBA degree from Duke University with concentration in Health Sector Management.