Recent advancements in generative AI, including models such as GPT-4 and LLAMA2, have been rapidly integrated into cancer research. It is not clear to which extent the AI assisted outputs can be trusted and used to the advancement of biomedical knowledge. We discuss and demonstrate several ways the genAI could be used in hematology and related fields.

We have compared a number of commonly available LLM based tools as well as custom trained AIs to determine their ability to enhance the accuracy of functional and genomic discovery screens by reducing the incidence of false positives and refining the interpretation process.

We demonstrate and discuss how custom training of the off-the-shelf genAI can be used to reduce or eliminate the AI-generated inaccuracies, often referred to as 'hallucinations“ . We further show the capacity of semantic scoring applied to candidates based on three orthogonally-tuned criteria: effectiveness, confidence, and novelty.

We observe that in benchmarking tests, the multi-agent scoring systems can outperform GPT-4 and prove particularly adept at identifying candidate targets that may have been overlooked in previous research.

The versatility of the custom trained genAI pipelines is demonstrated through their application in three distinct areas of cancer research. First - we discuss their applicability for optimization of chimeric antigen receptor (CAR) T-cell therapies. Second, we use them to elucidate the network of chemotactic molecules that orchestrate immune cell infiltration and tissue remodeling in the lungs of patients with COVID-19. Third, we use genAI to spatially define signaling pathways in humanized xenograft murine tumor experiments, for which existing cross-species database knowledge is limited.

These applications underscore the genAI potential to contribute to the understanding and treatment of complex diseases.

Disclosures

No relevant conflicts of interest to declare.

This content is only available as a PDF.
Sign in via your Institution