Artificial intelligence engine for directed hypothesis generation and ranking

Inventors

Stumpe, Martin • HARLEY, Alena

Assignees

Tempus AI Inc

Abstract

An artificial intelligence engine for directed hypothesis generation and ranking uses multiple heterogeneous knowledge graphs integrating disease-specific multi-omic data specific to a patient or cohort of patients. The engine also uses a knowledge graph representation of ‘what the world knows’ in the relevant bio-medical subspace. The engine applies a hypothesis generation module, a semantic search analysis component to allow fast acquiring and construction of cohorts, as well as aggregating, summarizing, visualizing and returning ranked multi-omic alterations in terms of clinical actionability and degree of surprise for individual samples and cohorts. The engine also applies a moderator module that ranks and filters hypotheses, where the most promising hypothesis can be presented to domain experts (e.g., physicians, oncologists, pathologists, radiologists and researchers) for feedback. The engine also uses a continuous integration module that iteratively refines and updates entities and relationships and their representations to yield higher quality of hypothesis generation over time.

Core Innovation

The invention provides a person-centric data storage that includes biomarker data entities specific to a person, where the biomarker data comprises whole transcriptome-derived RNA data. The person-centric data storage also includes one or more of clinical findings, disease state diagnostic data, or imaging-derived data, and further includes additional biomarker data and annotations associated with the additional biomarker data.

The method determines relationships between biomarker entities and outcomes events within the person-centric data storage. It then references relationships between biomarker entities and outcomes events within a background knowledge data storage generated from a plurality of structured annotation data sources, such as scientific literature, guidelines, domain-specific knowledge bases, clinico-genomic data sets, and/or large-scale EHR data sets, where the background knowledge data storage stores the biomarker entities, the outcomes events, and respective relationships between them.

The method generates ranked candidate relationships of observational significance between biomarker entities and outcomes events within the person-centric data storage. For each generated ranked candidate relationship, it attributes a confidence score by translating the generated ranked candidate relationships into queries against the referenced relationships within the background knowledge data storage and inputting the queries into an artificial intelligence engine configured to determine the confidence scores.

It ranks the generated ranked candidate relationships based on the attributed confidence scores to classify predictive accuracy and clinical actionability, including decreasing a rank for candidate relationships that match referenced relationships in the background knowledge data storage and increasing a rank for candidate relationships reflecting previously unknown relationships, and returns one or more of the ranked candidate relationships to a user.

Claims Coverage

The partial content includes three independent claims. Across them, each claim uses the same core pipeline: person-centric biomarker data storage including whole transcriptome-derived RNA, determination of biomarker-to-outcome relationships, referencing a background knowledge data storage built from structured annotation sources, generating ranked candidate relationships of observational significance, attributing AI-derived confidence scores via query translation, ranking based on predictive accuracy and clinical actionability with explicit up/down rank behavior for known-versus-previously unknown relationships, and returning ranked relationships to a user.

The independent claims cover a person-centric biomarker-to-outcome relationship discovery and ranking workflow that leverages a background knowledge data storage built from structured annotation sources and an artificial intelligence engine to compute confidence scores. The ranking is explicitly structured to increase priority for previously unknown relationships and decrease priority for relationships that match referenced background knowledge, with classification according to predictive accuracy and clinical actionability, and with the outputs returned to a user.

Stated Advantages

Documented Applications