Data Extraction Method

1. Publication Node: Canvas of Nodes representing all publications that have passed the inclusion criteria from the previous stage of the Method.

2. Loop Node: An instructional Node that directs the AI to loop through all Nodes in the previous Canvas, ensuring comprehensive coverage of each Research Paper from #1 Publication

3. Analysis Context Node: Relevant context/instructions for the AI to review the publication and extract data. In the Dual Targeting section, the scientists created an outline split into three parts: 1) introduction and initial focus on hematological malignancies 2) complexity in targeting solid tumors 3) combining tumor associated antigens. Each part had sub-bullets with their perspectives and details. We then provided a reference for how this data will be used and instructions for how to extract data according to the outline.

4. Prompt Node: Provides the AI with detailed instructions to follow the predefined Analysis Template for extracting relevant information from each publication. In the Dual Targeting section, we instructed the AI to follow the instructions in the Analysis Context Node and label each node with the publication title so that we can retain citations.

5. Data Extraction Generation Node: Creates a Canvas populated with Nodes containing the extracted data from each publication, structured according to the Analysis Template.

6. Synthesis Instruction Node: Directs the AI to synthesize the extracted data across all Nodes into a single consolidated Node, maintaining the title of each publication for citations

7. Synthesis Node: Generates a final Node page that consolidates all synthesized findings, ready for further analysis or reporting.

The Synthesis Node was then selected as context where we wrote various prompts to create a writeup from the data extractions. A few other prompt themes we used in the Dual Targeting and other sections are here:

· Devil’s Advocate: AI stating opposing perspectives and challenging the thinking

· Open questions: What questions could a peer or other expert still ask. Where is more evidence required

· Insight Development: Generating ideas on what could be added, altered, or removed from the writeup or fact list. Here, additional context not directly extracted from our Method was incorporated, often embedding output from other sections to surface new insights.

· Similarities and Differences: Reviewing the publications to identify any conflicting data, noting what information was repeated versus what was contradictory.

This Method is designed to be expandable, allowing for the addition of these and other analytical branches as research progresses. As we continue to refine our approach, the AI Business Engineer will play a key role in identifying best practices and integrating them into the primary Method.

The Dual Targeting section underwent several iterations of the Data Extraction Method. Portions of the Method were rerun, both individually and in their entirety, with additional manual edits made to refine the narrative and structure of the literature review as it evolved.