ChatDRex: Multi-Agent Bioinformatics System

• ChatDRex is a multi-agent conversational system that enables natural language querying for network-based disease module identification and drug repurposing using the NeDRex knowledge graph.
• It orchestrates specialized agents for literature mining, functional coherence analysis, and network analytics, employing algorithms like DIAMOND and TrustRank for precise predictions.
• The system integrates NLP-driven query decomposition with robust hallucination detection, increasing accessibility for clinical researchers without coding expertise.

ChatDRex is a multi-agent conversational system that provides natural language access to network-based disease module identification and drug repurposing prediction over the integrated biomedical knowledge graph (KG) NeDRex. By deploying specialized bioinformatics and literature-mining agents within a coordinated workflow, ChatDRex enables researchers and clinicians—regardless of coding expertise—to conduct complex translational analyses for drug repurposing, hypothesis generation, and personalized medicine advancement (Süwer et al., 26 Nov 2025).

1. Agent-Oriented System Architecture

ChatDRex follows a multi-agent design paradigm, where a central Planning Agent orchestrates a modular pipeline involving several specialized agents. User queries are first handled by a Summary Agent (dialogue tracking, input guardrails), then routed via the Planning Agent to domain-specific sub-agents—Network (NeDRex), functional coherence (DIGEST), knowledge-graph (Cypher/GraphRAG), Research (literature mining), and visualization modules. Each agent performs a narrowly scoped task: for example, the Network Agent executes algorithms such as DIAMOND, TrustRank, and Closeness Centrality on the KG; the DIGEST Agent conducts in silico functional coherence analysis; the Research Agent decomposes broad literature requests for extraction via Semantic Scholar; the Finalize Agent aggregates and formats results, including verification for hallucination detection. Intermediate states and results are maintained in agent-specific memory, enabling reasoned chaining of tool calls and analysis steps. This distributed reasoning ("agent-oriented chain-of-thought") diverges from standard CoT by explicit modularization, with each agent maintaining self-contained control logic and verification mechanisms (Süwer et al., 26 Nov 2025).

2. Biomedical Knowledge Graph (NeDRex): Schema and Representation

NeDRex is the underlying KG supporting ChatDRex, designed for systems medicine and pharmacological inference. Node types encompass Gene/Protein, Disorder (Disease), Drug, Pathway, Tissue, Genomic Variant, and Side Effect; relation types include standardized biological and medical associations:

• (Gene)–[:GeneAssociatedWithDisorder]→(Disorder)
• (Drug)–[:HasIndication]→(Disorder)
• (Drug)–[:HasContraindication]→(Disorder)
• (Protein)–[:InteractsWith]–(Protein)
• (Gene)–[:ExpressedIn]→(Tissue)

Formally, the KG is represented as $G = (V, E)$ with $|V| = n$ nodes, adjacency matrix $A \in \{0,1\}^{n \times n}$, and node feature matrix $X \in \mathbb{R}^{n \times d}$ encoding attributes and type embeddings. This comprehensive schema supports multi-hop traversals required for disease module discovery and drug-gene association inference (Süwer et al., 26 Nov 2025).

3. Network-Based Disease Module Identification

ChatDRex leverages established network analytic algorithms for disease module construction:

• DIAMOND: Starting from a seed set $S$ of disease-associated genes, DIAMOND iteratively expands the module by adding protein nodes $v \notin S$ that maximize a connectivity significance score (typically a $z$-score of edges to $S$), halting when candidates do not meet a pre-defined threshold.
• Random Walk with Restart (Editor’s term): Often employed in network medicine, this method propagates relevance vectors via $p_{t+1} = (1-r) W p_t + r p_0$ where $W$ is column-normalized adjacency, $p_0$ localizes the walk at seeds, and $r$ is a restart probability.

Graph convolution (using $$H^{(l+1)} = \sigma\bigl(\widehat{D}^{-1/2}\widehat{A}\widehat{D}^{-1/2} H^{(l)} W^{(l)}\bigr)$$ with normalization) is cited as a potential extension for gene neighborhood feature enrichment. Parameters such as significance thresholds in DIAMOND and restart probabilities in random walks are critical for module composition (Süwer et al., 26 Nov 2025).

4. Drug Repurposing: Ranking and Network Proximity

Drug repurposing within ChatDRex involves prioritizing drugs based on their network proximity to disease modules:

• TrustRank: A personalized PageRank variant, with the stationary distribution $\pi$ computed via $\pi = \alpha P \pi + (1-\alpha) e_s$, where $e_s$ seeds the trusted nodes.
• Closeness Centrality: Scores each drug node $d$ by $$C_{\text{closeness}}(d) = \frac{n-1}{\sum_{v \in V} d(d, v)}$$ using shortest path distance.
• Network-based Proximity: The mean shortest-path distance $\Delta(D, M) = |\langle d(D, v)\rangle_{v \in M}|$ quantifies drug-module proximity; lower $\Delta$ signals stronger network relevance.

Drugs are ranked and thresholded by descending TrustRank or ascending proximity, with statistical significance assessed by comparing against random module permutations. These rankings are empirically validated against curated gold standards, with metrics such as precision, recall, and F1 reported per tool and agent (Süwer et al., 26 Nov 2025).

5. Functional Coherence and In Silico Validation

Functional coherence of gene modules is established with the DIGEST Agent. Scores are based on pairwise GO annotation Jaccard similarity:

$$J(g_i, g_j) = \frac{|A(g_i) \cap A(g_j)|}{|A(g_i) \cup A(g_j)|}$$

The overall coherence $C(M)$ is computed as the average of pairwise similarities:

$C(M) = \frac{2}{m(m-1)} \sum_{i<j} J(g_i, g_j)$

Statistical significance is determined as an empirical p-value, $p_{\text{emp}} = \frac{k+1}{R+1}$, where $k$ is the count of random gene sets with coherence scores at least as extreme as $M$ over $R$ samples.

Alternatively, term enrichment can be evaluated with a hypergeometric test, with Benjamini–Hochberg correction applied to adjust for multiple hypotheses. Validation across ChatDRex use cases yields per-tool precision, recall, and answer-accuracy scores (e.g., Closeness Centrality: 0.95 for tool, call, and answer accuracy) (Süwer et al., 26 Nov 2025).

6. NLP Modules: Literature Mining and Hallucination Detection

The Research Agent decomposes complex literature queries via LLM-driven prompt engineering, breaking them into targeted sub-queries, retrieving publication metadata from Semantic Scholar, and synthesizing concise summaries. For KG-access tasks, repeated failures in Cypher execution trigger a Retrieval-Augmented Generation (GraphRAG) fallback.

Robustness is maintained by guardrails against prompt injection, output hallucinations, and citation format errors. Hallucination detection is finalized by a secondary LLM pass that checks factual consistency against dialogue state and KG context, helping ensure reliability in biomedical inference (Süwer et al., 26 Nov 2025).

7. Implementation, Example Use Cases, and Evaluation

ChatDRex is implemented as a backend Quarkus/Java system, using LangChain4j, bespoke NeDRexAPI client, a self-hosted LLM (gpt-oss-20b via Ollama), and a front end developed in Angular. Visualization is enabled by Drugst.One. The no-code web interface and public API endpoints expedite adoption by researchers (https://github.com/SimonSuewerUHH/ChatDRexAPI4J; https://github.com/SimonSuewerUHH/ChatDrexUI).

A representative workflow for Huntington’s Disease proceeds through identification of seed genes (e.g., HTT, GRIK2, TCERG1, PPARGC1A), module expansion via DIAMOND, functional coherence validation (e.g., $C(M)=0.23$, $p=0.002$), drug ranking (Tominersen, score $=0.85$), and literature extraction of supporting findings.

Evaluation metrics include tool-accuracy ($0.86$), call-accuracy ($0.852$), and answer-accuracy ($0.61$) averaged across agents and use cases (Süwer et al., 26 Nov 2025). Functional and network modules achieve high precision in correct tool selection and execution, while answer-accuracy indicates residual areas for LLM output improvement.

---

ChatDRex advances conversational, multi-agent bioinformatics by integrating natural language KG querying, network-based module analysis, and functional/literature validation in a user-controllable framework, thereby increasing accessibility and analytical depth for non-programmer clinical researchers.