AWS Partner Network (APN) Blog

Quartz Atlas AI for Drug Discovery

By Manish Patel, Global HCLS Solutions Architect – Amazon
By Kuldeep Singh, Global PDM Data and AI – Amazon
By Paul Prohodski, Specialist Leader – Deloitte
By Daniel Ferrante, Managing Director, AI & Data Strategy and Analytics – Deloitte
By Chris Hayduk, Lead Machine Learning Engineer – Deloitte

Deloitte
Deloitte
Connect with Deloitte

The process to successfully bring a new drug to market is difficult and costly. Deloitte estimates that it requires nearly $2 billion and approximately 10 years according to their 2024 report, Measuring the return from pharmaceutical innovation. Another recent study analyzed 183 oncology clinical trials targeting the Insulin-like Growth Factor-1 Receptor (IGF-1R), focusing on 16 different drugs. These trials involved over 12,000 patients and an investment exceeding $1.6 billion. Despite these significant investments, none of the drugs received approval for clinical use.

To avoid expensive and unproductive trials, the industry should focus on improving preclinical models. It should also enforce rigorous target validation and enhance early-stage decision-making in drug development. By optimizing preclinical research, pharmaceutical companies can better facilitate higher-quality drug candidates reaching the clinical trial stage, reducing the probability of late-stage failures.

As a response to assist with these challenges, Quartz Atlas AI™ for drug discovery was created. This innovative, AI-powered workbench is designed to revolutionize pharmaceutical research and discovery. Atlas AI is hosted on Amazon Web Services, built by Deloitte, and powered by the NVIDIA BioNeMo AI platform for drug discovery. It’s is a tool crafted by scientists for scientists.

In this post, we will discuss how AWS and Deloitte help customers accelerate drug discovery using Atlas AI.

Deloitte is an AWS Premier Tier Services Partner and managed service provider (MSP). Through a network of professionals, industry specialists, and an ecosystem of alliances, Deloitte assists clients in turning complex business issues into opportunities for growth, helping organizations transform in the digital era.

Solution Overview

Atlas AI is designed to help researchers address three main obstacles they commonly face. First, it tackles the lack of data connectivity between modalities and departments. Second, it simplifies insight derivation from increasing volumes and velocities of data. Third, it enhances data and AI model accessibility for bench scientists, reducing dependence on IT support. Atlas AI helps researchers accelerate R&D and address these challenges through the following key pillars:

Connectivity

  • Connects over a dozen open source pharmaceutical datasets, including proteins, chemicals, antibodies, academic literature, and more. Users can access this data in a single query in a matter of seconds, rather than searching through all these disparate datasets over the course of hours.
  • Provides functionality to ingest a research group’s proprietary data, semantically linking it to Atlas AI’s open source data offerings and augmenting its value.
  • Allows departments within an R&D group to easily access information and results from other groups, accelerating organizational learning and institutional knowledge.

Insights

  • Applies domain-specific large language models (LLMs) to each modality ingested into the knowledge graph, allowing Atlas AI to connect molecules based on their properties, labels, and subject matter expert (SME) annotations.
  • Surfaces interesting insights to the researcher through a combination of LLM-based and experiment-based edges.

Accessibility

  • The platform features a no-code interface, allowing researchers to freely interact with the knowledge graph and AI models without requiring any IT support.
  • The platform features an AI research assistant chatbot, enabling users to engage in a “dialogue with the data” to tackle complex, multi-step reasoning challenges.

Figure 1 illustrates how to accelerate drug discovery using Quartz Altas AI.

Figure 1 – Atlas AI Architecture

 The workflow steps are as mentioned below:

  1. Atlas AI users connect to the website frontend to access the underlying data, LLMs and AI scientific pipelines in no-code, accessible interface.
  2. Atlas AI use a React frontend hosted in Amazon Simple Storage Service (Amazon S3) and delivered through Amazon CloudFront to ensure that website performance is fast, reliable and secure.
  3. Amazon API Gateway allows access to all of Atlas AI’s backend services and supports a flexible microservices architecture, allowing new services continuously added, improving tools functionality.
  4. Amazon Cognito authenticate users. Prior to accessing any backend service, ensuring safe and secure access to data and AI models.
  5. Amazon Elastic Load Balancing(ELB) distributes requests over number of backend servers, ensuring that application remains performant even with the large number of inbound requests from users.
  6. Amazon Elastic Container Service (Amazon ECS) supports scalable containerized services that allows users to interact with the underlying AI models and databases.
  7. Amazon Bedrock hosts Atlas AI’s powerful large language models (LLMs). These LLMs act as the brain of the application, offering a seamless natural language chat interface. Through this intuitive platform, users can effortlessly manage AI-driven scientific workflows, access and query various data sources, and spark innovative hypotheses—all through simple conversation.
  8. Atlas AI harnesses the power of Amazon Neptune to underpin its cutting-edge system that combines generative AI, large language models (LLMs), and knowledge graph technology. This innovative platform currently incorporates 13 publicly available datasets, with more being added regularly. These datasets can be seamlessly merged with a client’s proprietary information, creating a comprehensive data foundation. This rich, integrated dataset serves as a robust resource for generating hypotheses and facilitating AI model inference and training.
  9. Atlas AI’s scientific pipelines seamlessly integrate cutting-edge AI models into comprehensive workflows, effectively simulating large-scale research experiments. These pipelines are highly adaptable and can be tailored to meet specific use case, enabling precise and targeted outcomes.
  10. NVIDIA BioNeMo on AWS is a generative AI platform for drug discovery that simplifies and accelerates the training of models using client’s data and scaling the deployment of models for drug discovery applications, NVIDIA BioNeMo provides the foundation for Atlas AI’s scientific pipelines.

Unveiling Atlas AI’s Cutting-Edge Features

Rich Knowledge Graph

At its core, Atlas AI is an ever-growing dynamic knowledge graph, thoughtfully mapping some of the most resourceful open source, peer-reviewed databases:

Deloitte-Quartz-AtlasAI-2

Figure 2 – Summary of Knowledge Graph Statistics Mapping Open Source Peer Reviewed Data

Besides the embedded databases shown in Figure 2, Atlas AI can be customized with the client’s private data safely in their own private environment, wherever their data resides. Together, they form an R&D accelerator — a powerful tool that allows researchers to execute complex queries against the dynamic Knowledge Graph without writing SQL or cypher queries. The platform is equipped to search for a range of entity types, such as proteins and chemicals, each with an associated search type. These search types are fully customizable, pre-written queries that answer research questions concerning proteins and chemicals.

Intuitive and user-friendly interface

Atlas AI provides the easy-to-use interface supports queries like ‘Protein-Protein Interactions’ and ‘Protein Structural and Functional Role.’ Other options include ‘Similar Proteins or Compounds,’ ‘Compounds Bindings to Proteins,’ ‘Reaction Pathways,’ and ‘Supporting Literature.’ These queries can be customized to fit your specific needs. They retrieve data related to proteins and/or compounds quickly, all in a single pane.

Researchers can navigate across different degrees of connectivity from the initial protein or chemical of interest, enabling them to observe directly connected neighbors or more distant connections. For instance, as shown in Figure 3 a “two-hop” search for the protein ‘Bet v 1-a’ (found in birch pollen) would return a network of related chemicals, proteins, antibodies, protein variants, mutations, CATH domains, diseases, and patents. This semantic integration of multiple different data modalities is returned in a single query, significantly accelerating the research process

Figure 3 – Atlas AI creating knowledge and connections from data

Atlas AI employs chemical and protein language models to generate the relevant connectivity and network map of the data, connecting the ingested data to the Knowledge Graph. This provides both a causal experimental view of the data and a correlational intuitive view derived from the appropriate language models, thus establishing the searched piece of data in its broader semantic and multimodal contexts.

Atlas AI can execute a ‘similar protein’ query, which uses special LLM-derived edges to identify proteins like a target protein. The search results return a large network of proteins connected by two separate edges: edges derived from protein language models, and edges derived from experimental data.

As illustrated in Figure 4, The connection of experimental information with language model data from all the different modalities absorbed into the knowledge graph enables Atlas AI to suggest comprehensive hypotheses and conjectures to researchers, which can help them enhance their work processes.

Deloitte-Quartz-AtlasAI-7Figure 4 – Atlas AI’s depth of data

Empowering Researchers with Conversational AI

When it comes to accessibility, Atlas AI features a chatbot as indicated in Figure 5 simulates a research assistant, enabling researchers to query the knowledge graph conversationally. This removes the need to learn any specialized querying commands. The combination of the query builder and the chatbot empowers users to harness the full power of the knowledge graph to execute complicated queries and answer complex research questions without needing IT or software engineering help.

Figure 5 – Atlas AI easy to use interface with AI Research Assistant Chat

Conclusion

Quartz Atlas AI™ for Drug Discovery is a revolutionary AI-powered workbench that aims to transform the pharmaceutical research and discovery landscape. Hosted on Amazon Web Services (AWS), this innovative tool provides researchers with semantically enhanced data connectivity, accelerated time to insights, and increased accessibility for AI models and multimodal data. By harnessing the power of generative AI, it aims to revolutionize the scientific research landscape and foster innovative breakthroughs in the field of pharmaceutical research and discovery. As the pharmaceutical industry continues to grapple with the complexities of bringing new drugs to market, the Atlas AI platform offers a new path forward, leveraging AI to provide researchers the tools and insights they need to drive innovation and breakthroughs. With its efficient integration of data, analytics, and AI, Atlas AI indicates that the future of scientific research is not just promising; it’s here.
.
Connect with Deloitte.


Deloitte – AWS Partner Spotlight

Deloitte is an AWS Premier Tier Services Partner and MSP. Through a network of professionals, industry specialists, and an ecosystem of alliances, Deloitte assists clients in turning complex business issues into opportunities for growth, helping organizations transform in the digital era.

Contact Deloitte | Partner Overview | AWS Marketplace