PathwaySeeker: Evidence-Grounded AI Reasoning over Organism-Specific Metabolic Networks

PathwaySeeker is an evidence-grounded AI system that integrates proteomic and metabolomic data to reconstruct organism-specific metabolic networks, utilizing an "Oracle-in-the-Loop" inference method to generate and verify condition-specific biological hypotheses with explicit experimental provenance.

The Gist

Imagine you are trying to understand how a specific, unique factory (an organism) works. You have two main sources of information:

1. The Master Blueprint: A giant, generic instruction manual for all factories in the world. It tells you what could happen, but it doesn't know which machines are actually turned on in your specific factory today.
2. The Factory Floor Cameras: Real-time video feeds showing which machines are running and what materials are moving around right now. But these cameras are blurry, and you can't see every single corner of the factory.

The Problem:
Current AI tools are like a very smart intern who has memorized the Master Blueprint perfectly. If you ask, "How does this factory make product X?" the intern will give you a perfect answer based on the blueprint. But they might be wrong about your factory because they don't know which machines are actually running today.

On the other hand, if you just look at the blurry camera feeds (the experimental data), you see movement, but you don't have a map to connect the dots into a complete story.

The Solution: PathwaySeeker
The paper introduces PathwaySeeker, a new AI system that acts like a brilliant detective who combines the Master Blueprint with the Factory Floor Cameras.

Here is how it works, using a simple analogy:

1. Building the "Live Map" (The Knowledge Graph)

First, PathwaySeeker takes the blurry camera feeds (proteomics and metabolomics data) and overlays them onto the Master Blueprint.

• If the camera sees a machine (enzyme) running, it draws a solid line on the map.
• If the camera sees materials (metabolites) moving, it draws a line even if it didn't catch the machine running.
• The Result: A "Live Map" of the factory that only shows what is actually happening in this specific organism under these specific conditions. It ignores the parts of the blueprint that aren't being used right now.

2. The "Oracle-in-the-Loop" (The Truth-Checker)

This is the most clever part. Usually, when an AI guesses something, it just guesses. PathwaySeeker has a built-in "Truth-Checker" (called an Oracle).

• The Process: The AI tries to solve a puzzle (e.g., "How does the factory turn sugar into energy?"). It generates a few possible routes.
• The Check: Before it tells you the answer, it pauses and asks the Oracle: "Hey, does our Live Map show that this specific step is happening?"
• The Feedback:
  • If the map shows it, the AI says: "Confirmed Fact." (Green light).
  • If the map is silent but the step makes chemical sense, the AI says: "Plausible Hypothesis." (Yellow light). It clearly tells you, "I think this happens, but we haven't seen it on camera yet."
  • If the map shows it's impossible, the AI stops and tries a different route.

3. The "Evidence Tagging" System

When PathwaySeeker gives you an answer, it doesn't just give you a story. It tags every single sentence with a label:

• GRAPH_FACT: "We saw this happen on the cameras."
• GRAPH_PATH: "We saw the whole chain of events happen."
• HYPOTHESIS: "We didn't see this, but based on chemistry, it's very likely. You should check this out!"

Why This Matters (The Trametes versicolor Example)

The researchers tested this on a fungus called Trametes versicolor (a white-rot fungus that eats wood).

• Old Way: Scientists would guess how this fungus eats wood by looking at how other fungi (like yeast) do it. They would assume the fungus works the same way.
• PathwaySeeker Way: It looked at the actual data from the wood-eating fungus. It confirmed the main path the fungus uses to break down wood. But it also found a new shortcut the fungus uses that no one knew about because it wasn't in the standard manuals.
• The Bonus: It also told the scientists, "We are 90% sure this other path exists, but we need to run one more experiment to be 100% sure."

The Big Takeaway

PathwaySeeker changes AI from a "know-it-all" that might be confidently wrong, into a "humble expert" that knows exactly what it knows and what it doesn't.

• Old AI: "Here is the answer." (Sometimes a hallucination).
• PathwaySeeker: "Here is the answer. This part is proven by your data. This part is a smart guess based on chemistry. This part is a wild guess you should test."

It bridges the gap between what we think is possible (the blueprint) and what is actually happening (the experiment), giving scientists a clear roadmap of what to study next.

Technical Summary

1. Problem Statement

Metabolic activity is not an intrinsic property of an organism but an emergent state shaped by specific environmental and experimental contexts. Current computational frameworks face three fundamental limitations in addressing this:

• Static Databases: Curated pathway databases (e.g., KEGG) provide organism-agnostic, static maps that often lack condition-specific regulatory constraints and are incomplete for non-model organisms.
• Disconnected Multi-Omics: While multi-omics datasets (proteomics, metabolomics) capture dynamic, condition-specific molecular evidence, they lack a structured reasoning layer to transform this data into mechanistic pathway hypotheses.
• LLM Limitations: Large Language Models (LLMs) possess advanced biochemical reasoning capabilities but rely on pre-trained, static knowledge. They cannot inherently distinguish between experimentally supported relationships and plausible inferences drawn from their training corpus, leading to "hallucinations" or ungrounded claims in scientific contexts.

The core challenge is to create an AI system that can reason over metabolic networks using organism-specific experimental evidence while transparently distinguishing between confirmed facts and testable hypotheses.

2. Methodology: PathwaySeeker Framework

PathwaySeeker is an AI system that couples LLM-based inference with explicit verification against experimental data. It operates in three main stages:

A. Condition-Specific Metabolic Knowledge Graph Construction

• Data Integration: The system integrates proteomic and metabolomic data from specific experimental conditions (e.g., Trametes versicolor under varying cultivation regimes).
• Inclusive Logic: It uses an "OR-based" integration strategy. A reaction is included in the graph if supported by either proteomic evidence (detected enzyme) or metabolomic evidence (detected substrate/product), rather than requiring both simultaneously. This prevents network fragmentation due to incomplete omics coverage.
• Evidence Annotation: Every node (compound, enzyme) and edge (reaction) is annotated with its evidence source (Proteomics-only, Metabolomics-only, or Both).
• Cofactor Constraints: To prevent biologically implausible shortcuts through highly connected hubs (e.g., ATP, NADH), the system enforces a three-tier constraint strategy:
  1. Cofactors cannot be start/end nodes.
  2. Hypothesized edges cannot be mediated solely by cofactors.
  3. Cofactor-mediated edges incur a high traversal penalty during search.

B. Schema-Aware Training Data Generation & Fine-Tuning

• Training Data: The system generates supervised training examples programmatically from the constructed graph, adapting the Graph-aligned Language Model (GLaM) approach.
• Evidence Typing: Training examples are categorized into distinct epistemic regimes:
  • GRAPH_FACT: Single verified reactions.
  • GRAPH_PATH: Multi-step chains where all edges are experimentally confirmed.
  • HYPOTHESIS: Biochemically plausible connections absent from the graph but supported by neighborhood evidence.
  • NO_PATH: Pairs in disconnected components (teaching the model that absence of evidence $\neq$ biological impossibility).
  • INVALID: Queries violating metabolic constraints (e.g., cofactor endpoints).
• Fine-Tuning: A base LLM (GPT-4.1) is fine-tuned on this schema-aware dataset to learn the distinction between verified and inferred relationships.

C. Oracle-in-the-Loop (OitL) Inference

This is the core reasoning engine, designed to prevent the model from reverting to pre-training priors when experimental data is silent.

• Iterative Search: The system uses a beam search approach (beam width $k=3$) with a maximum of 3 iterations.
• The Oracle: A "positive-evidence-only" graph oracle acts as a constraint. It verifies each step of a generated hypothesis against the experimental graph.
  • Crucial Logic: If evidence exists, it is returned. If evidence is absent, it is not treated as a rejection (acknowledging data incompleteness) but as a signal to generate a hypothesis.
• Scoring: Hypotheses are scored based on:
  1. Evidence Support (40%): Direct, partial, or alternative support from the oracle.
  2. Biochemical Coherence (40%): Logical consistency.
  3. Parsimony (20%): Simplicity of the pathway.
• Output: The final response explicitly labels every pathway edge as GRAPH_FACT, GRAPH_PATH, or HYPOTHESIS (with confidence scores and reasoning).

3. Key Contributions

1. Evidence-Grounded Reasoning: Introduces a framework where AI reasoning is constrained and verified by organism-specific multi-omics graphs, moving beyond static database lookups.
2. Oracle-in-the-Loop Architecture: Proposes a novel inference pattern where an external "oracle" iteratively verifies LLM hypotheses, allowing the system to separate confirmed knowledge from structured hypotheses in real-time.
3. Epistemic Transparency: Every output claim carries explicit provenance. The system does not just predict a pathway; it categorizes the certainty of each step (Verified vs. Hypothesized).
4. Handling Non-Model Organisms: Demonstrates the ability to reconstruct metabolic networks for organisms with poor genome annotation (like T. versicolor) by leveraging an inclusive, evidence-driven graph construction strategy.

4. Results

The system was validated using multi-omics data from the white-rot fungus Trametes versicolor.

• Graph Construction:

  • Reconstructed a graph with 3,402 unique reactions connecting 1,192 compounds.
  • Achieved 90.1% compound coverage for detected metabolites.
  • 26% of edges were supported by both omics layers, 54% by proteomics only, and 20% by metabolomics only.
  • Compared to the global KEGG universe, recall was 27.6% (intentionally restricted to experimental evidence); compared to S. cerevisiae, recall was 72.6%.

• Case Study: Phenylpropanoid Pathways:

  • Recovery of Canonical Paths: Successfully reconstructed the L-phenylalanine to ferulate backbone and the L-tyrosine entry route, matching manually curated results from a 2025 study (Monteiro et al.) without access to that literature (due to LLM training cutoffs).
  • Novel Discovery: Identified a branch from 4-coumarate to 4-hydroxybenzoate directly from the experimental graph, a route that manual curation had missed because it relied on cross-species homology rather than direct organism-specific evidence.
  • Differentiation: Clearly distinguished verified steps (GRAPH_PATH) from extensions (HYPOTHESIS).

• Validation:

  • Thermodynamic Feasibility: Post-hoc analysis using eQuilibrator confirmed that the AI-inferred pathways were thermodynamically feasible (positive Max-min Driving Force) and identified specific rate-limiting bottlenecks.
  • Biological Coherence: Overlaying metabolite abundance profiles showed that "GRAPH_PATH" segments exhibited coordinated condition-dependent dynamics, while "HYPOTHESIS" segments did not, validating the system's ability to distinguish active flux from mere structural connectivity.
  • Systematic Evaluation: Across 64 queries, the system achieved an Experimental Evidence Ratio (EER) of 26.4% (meaning ~26% of edges were experimentally verified, while the rest were labeled hypotheses). An LLM judge rated the system highly for Evidence Transparency (4.98/5) and Scientific Reasoning (4.93/5).

5. Significance

PathwaySeeker represents a paradigm shift in computational biology by operationalizing uncertainty. Instead of treating AI as a black box that generates plausible-sounding but potentially hallucinated pathways, it creates a structured, auditable reasoning process.

• Bridging AI and Experiment: It bridges the gap between frontier AI capabilities and organism-specific experimental evidence, allowing researchers to systematically differentiate between what is known, what is plausible, and what requires experimental validation.
• Non-Model Organism Discovery: It provides a scalable method to uncover metabolic capabilities in non-model organisms where curated models do not exist, reducing reliance on biased, model-organism-centric assumptions.
• Actionable Hypotheses: By explicitly labeling unverified connections as "HYPOTHESES" with confidence scores, the system directly guides experimental design (e.g., which enzymes to assay or which metabolites to trace), accelerating the discovery cycle in metabolic engineering and bioprocess optimization.