Impacts of AI in the Evolving Oncology Landscape

As we navigate the rapidly evolving landscape of oncology, this month’s blog post focuses on recent advancements in AI and their potential impact in oncology, while also addressing key challenges in research and clinical trials.

AI in Histopathology: A recent study published in Nature Machine Intelligence introduces VirtualMultiplexer, a generative AI toolkit that synthesizes multiplexed immunohistochemistry images for several antibody markers (AR, NKX3.1, CD44, CD146, p53, and ERG) from a single hematoxylin & eosin (H&E) image. This innovation can significantly accelerate histopathology workflows and enhance our ability to predict clinically relevant end points. In the paper, the authors also show how multiplexed imaging can learn from the spatial distribution of proteins to potentially predict patient outcomes. The challenge, as mentioned in last month’s blog post, is the fragmented nature of health systems that makes AI-based pathology difficult to validate and incorporate at scale.

AI Agents in Scientific Research: PaperQA2 is a new AI language model that is pushing the boundaries of scientific literature review. This tool has demonstrated the ability to produce Wikipedia-style summaries of scientific topics with accuracy surpassing existing human-written articles in certain areas. One of the mode’s core features is the ability to identify contradictions within scientific literature, averaging 2.34 contradictions per paper in a subset of 93 biology publications, with 70% validated by human experts. This capability could be invaluable for literature reviews and hypothesis generation moving forward.

Figure: Schematic of PaperQA2’s agentic toolset along with relevant action representations within each tool.

Challenges and Opportunities With Large Language Models (LLMs): Even with AI models being able to review scientific papers, they come with several important limitations (as discussed in a recent Nature Machine Intelligence editorial):

1. Factual inaccuracies and hallucinations remain significant issues, as discussed in previous blog posts reviewing AI chat boxes
2. LLMs struggle with contextual coherence and domain-specific reliability
3. They perform better in deductive reasoning than inductive reasoning

As mentioned above, LLMs do present opportunities for enhancing fact-checking processes, potentially assisting in stance detection and domain-specific verification.

Clinical Trial Accrual Challenges: Switching gears to other aspects of clinical research, a concerning trend in radiotherapy cooperative group trials highlights the need for improved trial design and recruitment strategies. Two-thirds of currently enrolling trials activated at least 6 months ago have achieved less than half of expected accruals, with 40% enrolling no more than 25% of projected participants to date. This recent editorial in JAMA Oncology addresses strategies on how to tackle this problem.