BioPathfinder: Evidence-guided multi-agent platform enables hypothesis discovery for CAR-T engineering

Technical Summary: BioPathfinder

Problem Statement
Clinical studies of chimeric antigen receptor (CAR)-T therapy generate vast amounts of molecular and clinical evidence that remains fragmented across publications, public repositories, and patient-derived datasets. This fragmentation limits systematic therapeutic discovery because relevant information is distributed across primary articles, supplementary tables, and study-specific annotations without a coherent structure. A critical bottleneck exists in connecting publications to dataset accessions, sample contexts, CAR targets, and clinical settings. Consequently, unsupported claims or overinterpretations of computational patterns can misdirect experimental validation. While recent AI efforts have utilized Large Language Models (LLMs) for literature synthesis and antigen nomination, there is a lack of agentic, provenance-tracked systems capable of generating mechanistic hypotheses directly from CAR-T-treated patient evidence.

Methodology
The authors present BioPathfinder, an evidence-guided multi-agent workflow designed for closed-loop biomedical discovery. The system operates through four distinct, linked agentic components:

1. Curator (Evidence Construction): This agent retrieves publications and repository records, extracting dataset accessions and evidence spans. It resolves paper–dataset links and normalizes patient–sample contexts and CAR targets. Starting from 16,597 initial records, the Curator employs pre-filtering, SQL-based dataset identification, LLM-assisted promotion, and expert adjudication to construct a "frozen core corpus" of 57 dataset-linked patient resources (plus 10 supporting papers). This creates a provenance-tracked evidence base separating evidence construction from hypothesis generation.
2. Planner (Hypothesis Generation): Using the frozen evidence artifacts and a specific exploration question, the Planner generates structured hypothesis cards. These cards contain bounded claims, proposed analyses, confounders, stop rules, and source lineage. The planner is designed to explore diverse mechanistic and operational hypotheses rather than returning narrow claims.
3. Reviewer (Prioritization): A reviewer workflow evaluates draft hypotheses using two routes: an LLM baseline (scoring based on a rubric alone) and an agent-based route (incorporating curated context and network-check information). This process assesses novelty, biological relevance, methodological correctness, and dataset suitability. The agent-based reviewer demonstrated higher repeatability (measured by Spearman rank correlation and ICC) than the baseline.
4. Bioinfo Analyst (Computational Validation): This component translates reviewer-selected hypotheses into executable bioinformatics analysis plans. It utilizes a controlled local tool registry (including scVI, Seurat, Monocle3, and scTenifoldKnk) to perform quality control, clustering, differential expression, trajectory analysis, and virtual knockout simulations on patient-derived single-cell RNA sequencing (scRNA-seq) data.

Key Results
Applied to a curated corpus of CAR-T-treated patient studies, BioPathfinder successfully identified candidate mechanisms underlying CAR-T persistence, dysfunction, and resistance.

• Hypothesis Generation: The system prioritized the hypothesis that genes associated with an NK-like transition programme could be targeted to reduce CAR-T exhaustion and improve persistence.
• Computational Validation: Analysis of the prioritized dataset (GSE125881) revealed that the NK-like transition programme and its receptor-focused subset were significantly enriched in exhausted post-infusion CAR-T cells. Pseudotime analysis showed a progressive acquisition of this programme along a trajectory from central memory-like (Tcm) to terminal effector memory-like (Temra) states.
• Target Selection: Virtual perturbation analysis prioritized receptor genes KLRC1, KLRD1, and KLRG1. Expert review selected KLRC1 (encoding the inhibitory receptor NKG2A) for experimental validation.
• Experimental Validation:
  • In vitro: Chronic antigen stimulation of CD19 CAR-T cells led to a progressive increase in NKG2A expression, particularly in the CD8+ compartment, correlating with activation and exhaustion markers (PD-1, LAG-3, CTLA-4).
  • In vivo: In NSG mice bearing Raji tumors, NKG2A expression was enriched in CAR-T cells across multiple tissues. Crucially, NKG2A blockade (using anti-NKG2A antibody) significantly enhanced antitumor activity, reduced tumor burden, extended survival, and increased systemic effector cytokine production (IFN-γ, TNF-α, IL-2) compared to CAR-T cells alone.

Significance and Claims
The paper claims that BioPathfinder establishes a generalizable framework that transforms fragmented clinical single-cell evidence into experimentally validated therapeutic hypotheses. Its primary significance lies in:

• Provenance-Aware Discovery: Unlike systems relying on unstructured document retrieval, BioPathfinder builds a structured, provenance-tracked evidence base linking publications to specific dataset accessions and sample contexts.
• Closed-Loop Workflow: It demonstrates a reproducible route from dispersed clinical observations to testable CAR-T engineering strategies through a cycle of evidence curation, hypothesis generation, expert-style review, and experimental validation.
• Actionable Therapeutic Insight: The study provides a concrete example where AI-guided discovery identified NKG2A as a targetable inhibitory axis. The authors state that this finding connects a computationally prioritized NK-like receptor programme to an experimentally testable intervention that improves CD8+ CAR-T cell antitumor activity and persistence.

The authors emphasize that BioPathfinder does not replace expert judgment or autonomously establish causal mechanisms; rather, it expands the hypothesis search space, organizes evidence boundaries, and enables structured prioritization to guide human-led experimental design. The work is presented as a scalable strategy for AI-guided biomedical discovery in translational medicine.