To meet the requirements of pre-clinical research and discovery for biopharma companies, EC created Cora for Life Sciences. By combining the wide coverage for natural language understanding provided by the BERT and T5 word embedding models, the generative AI capabilities of PaLM 2, and EC’s proprietary semantic analysis and reasoning capabilities, Cora automatically analyzes and ingests content in the life sciences domain.
The figure above shows the high-level architecture and workflow. Data flows into the Cora content analysis and ingestion process, where Cora automatically extracts rich knowledge structures using word embeddings, semantic parsing, and deep automatic analysis of concepts, relations, and qualifiers. Cora automatically identifies key concepts in the domain, such as genes, proteins, biomarkers, symptoms, etc., and how these concepts relate to each other. After additional processing to cluster, type, and link concepts, Cora loads the analysis results into the knowledge index.
At runtime, the Cora semantic query engine and logical reasoning engine leverage the knowledge index and any domain models to process the query, analysis, dialog, and evidence summarization requests from the front-end API. Cora uses PaLM 2 to both interpret natural language questions and generate provably correct search result summaries grounded in the specific evidence found by Cora. Any GUI can be connected to the Cora APIs, though Cora also includes a default UI/UX with the SaaS product.
The base system has already analyzed and ingested all of the content in PubMed (the portion freely available for commercial use), and is immediately available for SaaS use in Google Cloud. Cora can also easily ingest proprietary customer content and satisfy all of the customer’s security and privacy requirements, leveraging Google Cloud’s comprehensive data management and security support.
The overall system provides a powerful research and discovery platform for researchers conducting preclinical drug discovery literature research. In one evaluation, Cora reduced the time required to perform a drug repurposing research task from two weeks to less than two hours. Cora’s usefulness extends well beyond pre-clinical literature research, however, and can be applied to content analysis across the entire drug discovery life cycle.
Considerations and Tradeoffs
The first consideration with any application of generative AI is understanding the requirements for veracity. For applications requiring creativity, or where secondary validation of results is expected or required, applying generative AI is relatively straightforward. EC’s primary focus, however, is on applications where veracity, transparency, and provable correctness are essential. To provide trustworthy and accurate results from LLMs, EC applies them with constraints and guardrails that ensure the results are grounded in reliable evidence and logically correct explanations. EC solutions cannot produce hallucinations because they never ask the LLM to generate answers without grounding them in a human-approved domain model or reliable evidence.
Another consideration is the cost and speed of invoking an LLM. This is where the Vertex AI solutions excel and provide better response time at lower cost than competitors. There are still situations, however, where invoking the full LLM with hundreds of billions of parameters is too costly or inefficient. In those cases, EC uses the full LLM to generate training data for the task, then trains a much smaller, fine-tuned LLM from the foundation model and the training data to produce a highly-optimized and efficient model that meets use case accuracy requirements as well as cost and latency requirements.
Better Together
Google Cloud is a leader in the industry with massive natural language data sets and large-scale solutions. By partnering with Google Cloud, EC can leverage their expertise with LLMs and cloud-based solutions while focusing on the unique and differentiating AI technology they have developed to support deep content understanding and sophisticated logical reasoning.
Google Cloud has the reach to build large, accurate foundation models and host them reliably at scale and deliver them with speed.
Finally, Google has a strong research tradition that aligns very well with EC’s core values in research and innovation. This results in a unique partnership where EC and Google are willing to push the boundaries of the technology and work together to experiment, learn, and drive meaningful solutions for their customers.
Learn more about Google Cloud’s open and innovative generative AI partner ecosystem. Read more about the partnership between Elemental Cognition and Google Cloud here and request a demo here.
