Omina Technologies presented one of the six applied AI use cases that were part of AI4Growth’s 6th Session AI Biotech & Pharma.
We summarized the talks by Janssen on speeding up drug discovery through multi-task federated learning and the talk by XploData on how to automate the contract review process by using NLP. We also summarized our own use case on preventing brand switching and optimizing sales resources using explainable AI (XAI)
What if we could speed up drug discovery by co-opetition instead of competition? In co-opetition competing pharma companies cooperate.
The European project MELLODDY, in which multiple competing pharmaceutical companies collaborate, aims to boost the predictive performance and chemical applicability domain of drug discovery-relevant models, without unacceptable leaks of private information.
Sharing common knowledge that is beneficial to any drug discovery task between pharma companies could speed up a specific drug discovery task for a specific pharma company. A pharma company can benefit from training its’ specific drug discovery model starting from a more general drug discovery model that has been trained to perform multiple related drug discovery tasks on data from competing pharmaceutical companies. The specific drug discovery model will perform better and generalize better (less prone to overfitting) than when the model would be trained only on the specific pharma company’s data.
Technically, this benefit can be realized by implementing multi-task federated learning. Multi-task learning (MTL) is a subfield of machine learning in which multiple learning tasks are solved at the same time while exploiting commonalities and differences across tasks. In this case, the multiple tasks are drug discovery tasks on compound data sets owned by different pharmaceutical companies.
In multi-task federated learning, the machine learning model travels from one data center to another, while the data is not shared and remains in the data center of the specific pharmaceutical company. However, the raw dataset used for training contains private information, which can be maliciously recovered by carefully analyzing the model and outputs. To preserve privacy, the machine learning model is made of a common trunk shared between partners, while private heads are not shared among partners. The multi-task predictive machine learning algorithms incorporating an extended privacy management system, to identify the most effective compounds for drug development, while protecting the intellectual property rights of the consortium contributors.
The multi-task algorithms are fed to an application layer, which implements a traceable federated learning platform. The application is deployed using container orchestration (automated multicluster Kubernetes). The infrastructure is implemented as code; infra as code, which means that infrastructure can be set up and destroyed by code.
J&J has a lot of contracts with multiple suppliers that have to be verified for contract validity. The contracts have to be checked on the presence of certain key elements such as privacy, etc. Until now, the contract validation was done manually and it was a repetitive and time-consuming boring task.
The AI solution had to automate the contract review.
The solution pipeline contains the following steps:
The AI solution includes an NLP model and a classification model.
DistilBERT is used as transformer model. It is prefered over BERT, as it is a small, fast, cheap and light Transformer model. DistilBERT has 40% less parameters than bert-base-uncased , runs 60% faster while preserving over 95% of BERT’s performances as measured on the GLUE language understanding benchmark.
A contract review application has been built using Flask. J&J legal employees can see how past contracts score on the required contract criteria in the contract review application. They can also upload a new document and identify how the uploaded document scores on the required contract criteria.
The time to review a contract went from 60 minutes to 2 minutes, a 30X speed increase.
This use case has been implemented for a large pharmaceutical company.
The increase in competition due to the rise of generic drugs forces the pharmaceutical company of branded drugs to optimize its’ sales.
The objectives of the AI solution were two-fold:
The AI Solution has to provide accurate yet explainable brand switching predictions to facilitate legal audits and adjustments to action policies.
The pharmaceutical industry is highly regulated. Explainable AI allows assessing legal compliance.
Interpretable methods such as GA2M are combined with SHapley Additive exPlanations to provide accurate and explainable predictions on the efficiency of actions.
Marginal Effect Table (inferred from the model) provides insight into the marginal effect of increasing the action intensity with 1 unit on sales. These insights enable to re-optimize the current action plan.
Did you like the summary on AI4Growth?
If so, then subscribe to our Events Update list and receive future event updates and event summaries.
To receive future event updates, sign up for the Events list in this form.