AWS re:Invent 2020: Accelerating regulatory assessments in life sciences manufacturing
![AWS re:Invent 2020: Accelerating regulatory assessments in life sciences manufacturing AWS re:Invent 2020: Accelerating regulatory assessments in life sciences manufacturing](/pic/aws_re_invent_2020_accelerating_regulatory_assessments_in_life_sciences_manufacturing/R1NJQjlJZFdHWWM_.jpeg)
welcome everyone my name is usha nandigala i've been with aws for three years and in my current role i help our healthcare and life sciences customers innovate with amazon and aws technologies i'm here to kick off the session with an introduction to aws manufacturing in life sciences followed by our customer merck who will be presenting their solution on how they are able to accelerate regulatory assessments in life science manufacturing using aws what we hear from our customers is the operating model for manufacturing and supply chain is centered around data not just historic data to predict demand but also real-time patient demand sensing data patient sentiment data and probabilities of several input conditions that feed into manufacturing planning and other downstream processes aws is constantly innovating on our analytics iot and ai ml frameworks to better enable such a data-driven model also as many of you know aws was born from amazon a recognized leader over 20 years experience in logistics fulfillment and supply chain we commercialized our expertise from amazon in the form of various service offerings like amazon connect which is a managed call center offering amazon forecast which leverages ml in delivering highly accurate forecasts and amazon neptune which is a graph database for highly connected data sets all these services among many others are additional enablers in creating solutions for the different aspects of the operating operating model you see here now taking a step back we understand the trends and our customers tell us their major challenges in manufacturing and supply chain are in three areas uh opportunity to improve operational efficiencies new products demand new approaches and regulatory expectations for operational efficiencies there's an opportunity to improve the production area by reducing waste and making use of data that's already been collected for improvements for new products like carti we have arrived at a batch of one model many of the production approaches used in traditional pharma will not suffice for a precision medicine or biologic drug product now regarding regulations the industry regulate regulatory bodies have enacted new regulations addressing concerns around drug shortages recalls and counterfeit drugs so fda has indicated an expectation for manufacturers to have consistent control and real-time understanding of critical quality metrics along with the ability to immediately report any changes to those metrics as we dive deep into improving operational efficiencies we see the concept of a successful smart factory is one that uses data to achieve transparency agility and continuous improvement data driven outcomes require unlocking your data through desperate factory protocols oem machines and production systems like erm and mes once the data is readily available you now have the opportunity to implement smart factory use cases such as computer vision quality detection real time failure diagnostics with iot predictive maintenance with aiml edge analytics are just a few to name another unique challenge as we improve operational efficiency is staying aligned with compliance aws complies with number of regulatory domains that relate to healthcare and life sciences we are compliant with compliances that directly regulate aws where we are not directly regulated like the gxp compliance we have number of tools and resources that support you in building your gxc compliant environment as noted on this slide there is your account team professional services number of documents publicly available or under nda quick start guides and there are of course other customers you can talk to we are thrilled to be joined by one such customer mark who is on their journey to accelerate operational efficiency and how they access regulatory compliance in life science manufacturing at this time i'm excited and honored to invite invite john baker executive director of machine learning and analytics and kyle hayes director of regulatory affairs from merck to share with you their innovative solution they built on aws thank you usha hello everyone my name is kyle hayes and i'm a director in merck's regulatory affairs cmc organization the cmc organization helps to ensure merck's medicines and vaccines are safe and effective for our patients in the pursuit of our mission merck has collaborated with aws to develop the change assessment knowledge engine or what we affectionately refer to as cake as a cmc professional the cake application is transforming what has historically been an arduous process to assess changes into a piece of cake some of the statements that we make during this presentation may be considered forward-looking statements within the meaning of the safe harbor provision of the u.s private litigation reform act of 1995. merck is one of the world's largest multinational biopharmaceutical companies with a diverse portfolio of medicines vaccines and animal health products at merck we are driven to making a difference through our commitment to research and development with upwards of 10 billion dollars invested in r d annually we discover develop manufacture and market vaccines and medicines that are distributed to patients in more than 140 countries merck is headquartered in new jersey with facilities located around the globe at merck we are inspired by a shared vision and mission to save and improve lives for more than 125 years merck has been inventing for life bringing forward medicines and vaccines for many of the world's most challenging diseases in pursuit of our mission to appropriately manufacture a medicine or vaccine specific manufacturing processes product characteristics and product testing must be defined in order to ensure that the product is safe effective and consistent between batches these activities are known as cmc short for chemistry manufacturing and controls the cmc regulatory dossier is a critical module of clinical trial and marketing authorization applications small molecule biologic and vaccine products have similar yet different cmc regulatory requirements but each share the same purpose to ensure that the manufacturing process consistently produces a high quality product that is both safe and efficacious to that end it is critical that product quality is maintained throughout the product life cycle such that the attributes that are critical to the quality of the drug product remain consistent with those in the clinical studies ultimately regulators need to be assured that there is consistency between the product tested during the clinical trials and the commercial product manufactured years later merck operates in a highly complex regulatory environment governed strictly by more than 100 different national regulatory authorities around the world such as the fda health canada and european medicines agency to name a few the mission of these national regulatory authorities is to protect and promote public health in pursuit of their mission each national regulatory authority establishes their own regulations and guidelines and they are constantly evolving changing and being updated furthermore regulations both within an agency and across agencies vary for small molecule biologic and vaccine drug products as well as by the type of change that is being requested or required for instance the regulatory requirements for a modification to a manufacturing site's water system may be different for small molecule drug products compared to biologic and vaccine drug products since our business is highly dynamic it results in thousands of change proposals being generated annually each of which needs to be reviewed and assessed by a cmc scientist these changes can be initiated from our external suppliers health authority regulators or internally driven such as process improvements the current process requires cmc scientists to research across multiple systems repositories and documents to gather information and adjudicate an assessment for regulatory reportability for example a cmc scientist needs to read and interpret the change understand what markets the product is registered in and review the applicable health authority regulations to determine potential reportability for the change this process is highly manual cumbersome and very challenging to consistently assess changes in conformance with diverse and dynamic health authority guidelines and regulations what makes it challenging to recruit and retain cmc scientists therefore this process was an ideal candidate to explore and develop a more automated solution to streamline it for cmc scientists once the potential for regulatory reportability is determined the change proposal is progressed into our gmp system for execution and final evaluation we recently launched cake into three pilot sites in a cohort of external manufacturing where cmc scientists are conducting real world real world change proposal assessments in parallel with the manual legacy processes early data indicate up to a ninety percent reduction in overall duration and a thirty to seventy percent reduction in effort over the next six months we will incorporate learnings from a pilot and industrialize the cake tool for a full roll out to all of merck's human health manufacturing sites my colleague john baker executive director in machine learning and analytics at merck will now talk about the technology and architecture that forms the basis for cake thanks kyle hello everyone my name is john baker and i'm the project lead for cake my role at merck is executive director of machine learning and analytics where i lead a portfolio of valuable delivery projects in the area of ai machine learning and advanced analytics and i will be speaking to some of the technology and architecture that forms a basis for cake as well as some of the cultural elements that were critical to our speed of delivery and overall success the conceptual solution for cake has three key components first the back end is an extensive aws neptune graph database model that contains a wealth of information about merck's supply chain and our regulatory operating environment including our supply chain manufacturing recipes products health authority regulation health authority regulatory guidelines product registration information merck regulatory knowledge and much more and this is really the core of what makes the cake engine work second we have a natural language processing model that has been trained on thousands of historical change requests to recognize patterns and classifications of change requests with a fairly high degree of accuracy when a new change request arrives we apply the nlp model to make predictions about that change request the prediction is then confirmed with the cmc scientist amended as needed and then used together with the graph database to identify potential relevant regulatory reporting requirements this model is continuously learning from the results of each new change request prediction with the cmc scientist either confirming that prediction is accurate or amending that prediction and feeding it back into the model and third we have an elegant custom-built front end that the cmc scientist uses to interface with the system they log in ingest the change request confirm key information about the request and there and then nearly instantaneously receives a recommendation about potential regulatory reportability along with clear rationale and supporting detailed information now obviously there's tremendous value here in that end result or output being the recommendation and supporting rationale but what we heard unanimously from our user community is that the cake tool is really easy to use and intuitive following the natural flow and train of thought of a cmc scientist at the beginning of the project we invested many hours and days to literally sit with merck cmc's data scientists to learn about how they conduct these assessments and then we embedded actual merck cmc data scientists together with the agile team as product owners and established a diverse group of real world practice practitioners to help test the system functionality and provide feedback which we then rapidly built back into the system in a very systematic way so this is the logical architecture for cape building off what we saw on the previous slide the backbone of our aws network graph sorry aws neptune graph database is seeded and regularly updated through an etl data ingestion process from various merck enterprise systems including our sap implementation using the aws neptune graph database allows the system run extremely fast and the platform is very very scalable the machine learning models have been trained on over 1500 historical change requests and are constantly improving as they receive intelligence from new requests and all these things are integrated together leveraging api gateways and lambda functions and presented to the users via their desktop web browser cake runs in a web browser like a regular web application so to provide some functionalities the team decided to implement the web application using reach and typescript of course running on top of all these things are the typical core processes that manage security single sign-on activity logging monitoring and error handling and we use a combination of key value and object store services to collect session data during the assessment that can be used for auditing purposes and reporting with the metrics around application usage performance metrics or business data like what choices were made by the cmc scientist across different assessments here we see more specific about the suite of aws microservices that are leveraged in cape the aws network grap aws neptune graph database plays a key role in cake but there are many others amazon dynamodb for state management amazon cognito for security management aws step functions and the aws lambda function pulls everything together all of this is built on aws server-less technology from site hosting and s3 to microservices built on api gateway and lambda minimizing ongoing operational overhead and costs so here's a sample subset of the cake graph model this is probably a pretty typical graph model which is highly effective on connecting various disparate data elements in a highly flexible and organized fashion we see extensive information here about merck's products our recipes sites and materials we also see information about where our products are registered and health authority guidelines so let's take a short sample interactive walk through the simplified cake graph model to see how it works there are many different kinds of change requests but let's take the simple example of a change to a given merck product so as a cmc scientist we need to answer questions such as which sites manufacture this product and in which countries what materials are used to make that product which markets is the product registered in and what health authority guidelines are relevant for changes to this merck product so we'll start with the merck product or what we call product family because a merck product such as genuvia actually consists of several different genuvia products so we start to look at what flavors dosage forms psf's and other variations that are associated with this product they could be different strengths packaging forms or market specific packaging and then we start to drive down into the bills of materials also known as recipes and sub materials that are used to build the merck product this is an iterative cycle pharmaceutical manufacturing is very complex there are many layers to traverse in order to peel back the onion in that case we can use a recursive traversal using the repeat step in gremlin query until we find all the materials in all levels so we start with a merck product family from the change request form traverse the graph to the product of that family from the product we move to the bill of materials and then discover the materials because the bill of material can be a multi-level bomb it could be necessary to traverse the subgraph many times and then we can also filter the material related to products that are produced at an impacted site or number five as you see in this change request once we have a comprehensive understanding of the materials and recipes for this product we can then traverse into the sites to see where this product is manufactured these could be internal merck manufacturing sites or external suppliers either cmo contract manufacturers or the more traditional raw material suppliers the same type of query using sql with tables and joins can be very difficult to implement especially without knowing the number of levels in the product material hierarchy so once we have all this information about how and where the product is manufactured we can then start to look at all the markets where this portfolio of products are registered that's number six this is very important because the cmc scientist needs to evaluate potential reportability in every market the product is registered many of which are subject to different rules and interpretations often we have the markets and then we can then traverse over to the specific health authority guidelines or regulations for those markets that apply for this specific change or number seven so what you see here very clearly laid out is the power of the aws neptune graph database that is able to seamlessly and quickly link disparate information together leveraging nodes and edges that are simply just not possible in traditional sql based tabular queries so this is a sample snapshot of the final output of cake the cmc scientist receives a high level yes no recommendation for potential regular regulatory reportability the list of markets where the product is registered and potential reportability in each of those markets along with specific supporting rationale this comes in two forms first the specific health authority regulation the text of that regulation is shown when the cmc scientist hovers over the rule or a hyperlink can be clicked which takes the user directly to the actual health authority regulation regulation documentation and then second there are linkages and references to merck's internal regulatory knowledge reference database which contains supplemental information about the nuances and historical learnings for change requests that might not be straightforward or health authority or where health authority regulations are vague so before we leave this slide it's very important to point out that the human has not been removed from this process this is only a recommendation and it's not an automated decision making tool the final decision about potential regulatory reportability squarely sits with the cmc scientist the cake tool is all about helping the cmc scientists to more quickly consistently and effectively make that assessment so that wraps up the technical and architectural review but before we close i think it's important to highlight and spend some time on the more cultural elements that were key to our success on cake we literally went from an idea and a blank white board to a real world working system where cmc scientists today are using cake to conduct regulatory change assessments in less than 10 months of development time we are working very hard to improve and evolve our ways of working at merck and together with aws our project team embraced working in a different way much of which was very new for a lot of us organizationally we purposely kept our team small autonomous and nimble we affectionately called them two pizza teams teams small enough that you could feed the entire team with two pizzas we embraced an obsessive focus on customer value and made decisions quickly and even if they were not right the first time the teams learned from that experience and quickly course corrected the merck aws collaboration was tremendous and aws has been a fantastic partner for us our teams work seamlessly as an integrated unit to quickly and effectively deliver results without a lot of drama or internal politics we started with our vision and work backwards visualizing what the product would look like at the end like a press release in the future and then we made really difficult and sometimes unpopular decisions to ruthlessly prioritize our scope leveraging the agile methodology to deliver short three to four month bursts of minimum viable products or mvp deliverables and building on our success so there you have it cake provides a clear rationale and consistent recommendations in conformance with global regulatory requirements and that supports merck's commitment to providing a safe and reliable supply of medicine to our patients in august 2020 we launched into three merck sites and a cohort of external manufacturing early data shows that cake is showing significant improvements in duration effort and turn around time and we have a lot of work ahead to expand the tool and to industrialize solution for a full global human health rollout to all of our human health sites in 2021 and we're also evaluating opportunities to further improve cake with aws services for machine learning and document translations as well as leveraging the capabilities of cake and other potential use cases in the areas of regular regulatory manufacturing and supply chain so thanks so much for watching i hope you enjoyed the small taste of cake have a great day at aws re invent
2021-02-16 01:05