MARBLE: Multi-Agent Reasoning for Bioinformatics Learning and Evolution

TL;DR

MARBLE tackles the slow, manual cycle of bioinformatics model refinement by introducing an autonomous, multi-agent framework grounded in scientific literature and empirical performance. It orchestrates paper selection, debate-driven ideation, autonomous execution, and memory-based feedback to iteratively evolve target architectures. Across spatial transcriptomics segmentation, drug-target interaction prediction, and drug response prediction, MARBLE achieves sustained performance gains and high execution robustness, outperforming strong baselines in framework-level metrics such as $NPG$, $NAUI$, $SIC$, and $ESR$. The work demonstrates that structured debate, balanced evidence selection, and performance-grounded memory are critical for stable, repeatable model evolution, suggesting a general path toward automated, scalable bioinformatics refinement.

Abstract

Motivation: Developing high-performing bioinformatics models typically requires repeated cycles of hypothesis formulation, architectural redesign, and empirical validation, making progress slow, labor-intensive, and difficult to reproduce. Although recent LLM-based assistants can automate isolated steps, they lack performance-grounded reasoning and stability-aware mechanisms required for reliable, iterative model improvement in bioinformatics workflows. Results: We introduce MARBLE, an execution-stable autonomous model refinement framework for bioinformatics models. MARBLE couples literature-aware reference selection with structured, debate-driven architectural reasoning among role-specialized agents, followed by autonomous execution, evaluation, and memory updates explicitly grounded in empirical performance. Across spatial transcriptomics domain segmentation, drug-target interaction prediction, and drug response prediction, MARBLE consistently achieves sustained performance improvements over strong baselines across multiple refinement cycles, while maintaining high execution robustness and low regression rates. Framework-level analyses demonstrate that structured debate, balanced evidence selection, and performance-grounded memory are critical for stable, repeatable model evolution, rather than single-run or brittle gains. Availability: Source code, data and Supplementary Information are available at https://github.com/PRISM-DGU/MARBLE.

Figures (6)

• Figure 1: Overview of the MARBLE framework. MARBLE is an end-to-end multi-agent system for iterative bioinformatics model refinement. The framework operates in a closed loop comprising (a) literature-aware paper selection, (b) debate-driven ideation among role-specialized agents, (c) autonomous execution and evaluation, and (d) an evolving memory that records outcomes to guide subsequent refinement iterations.
• Figure 2: Hybrid paper selection with performance-grounded reward update. Candidate papers are scored using embedding-based domain similarity (paper abstract vs. target domain keywords) and architecture similarity (paper method vs. target model source code) via a unified semantic embedder. The two similarity components are combined using domain and architecture weights $(w_d, w_a)$, which vary across high-, middle-, and low-domain relevance groups, to produce a metric score. From an initial pool of 200 papers, top candidates are filtered per group and further ranked through agent-based validation. Execution outcomes from subsequent refinement iterations generate batch-aggregated reward signals, which update paper selection weights and guide performance-grounded reference prioritization across iterations.
• Figure 3: Example of debate-driven ideation in MARBLE. Illustrative ideation trace for refining the STAGATE model on a spatial transcriptomics domain segmentation task. Research Group generates initial architecture modification proposals based on the selected reference set, which are evaluated and critiqued by a Critic. The proposals are then refined and ranked by a Model Principal, who incorporates feedback to balance expected performance impact and feasibility, and makes the final selection. Based on this decision, the Implement Architect formalizes the chosen approach into a concrete specification for execution.
• Figure 4: Iterative performance trajectories of target models refined by MARBLE across bioinformatics domains. Performance trends of MARBLE over refinement iterations for (a) spatial transcriptomics domain segmentation (STAGATE, DeepST; ARI), (b) drug--target interaction prediction (HyperAttentionDTI, DLM-DTI; AUPRC), and (c) drug response prediction (DeepTTA, DeepDR; RMSE). For each model, the left panel shows iteration-wise performance trajectories across independent runs, the middle panel reports cumulative peak performance, and the right panel summarizes mean performance of the top-performing iterations. Across all tasks, MARBLE exhibits repeated performance improvements followed by stable convergence relative to the initial baselines.
• Figure 5: Iterative architectural refinement of STAGATE by MARBLE. (a) Performance trajectory (ARI) across refinement iterations and corresponding architectural modifications. Starting from the original STAGATE model, MARBLE successively introduces probabilistic denoising via attention--VAE encoding (i@5), enhances spatial smoothness through decoder redesign (i@17), and performs final architectural optimization (i@28), yielding consistent ARI improvements. (b) Spatial domain segmentation results for each iteration, showing progressively improved spatial continuity and alignment with known anatomical regions.