CoChat

The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery

Open access

The DOI matches its Crossref record; 48 of the 48 references it cites were found in a registry or an index; 60 of the 92 sources its claims cite were checked for support.

Abstract

Bioinformatics is entering a new phase characterized by the integration of universal biological models and multi-agent systems to enable end-to-end scientific discoveries. This review argues that the next paradigm shift will go beyond traditional predictive models and generative artificial intelligence (AI) toward agentic AI: systems capable of planning, acting through tools, reflecting on results, and iterating until a goal is achieved. We first examine recent foundational models that produce transferable representations across omic modalities, such as scGPT, Nicheformer, and EpiAgent, and discuss their architectural choices, training regimes, and interpretability constraints. We then analyze biomedical agent frameworks through their main components (planning, action, reflection, and memory), highlighting representative systems such as ClinicalAgent and Biomni that operationalize these ideas in controlled environments. Next, we focus on hypothesis validation mechanisms, including retrieval-augmented generation for evidence grounding, sequential statistical testing, and benchmarking methodologies designed to quantify robustness and reproducibility. Finally, we summarize emerging applications in drug discovery and personalized medicine, from molecular literature analysis and protocol automation to drug repurposing for rare diseases and closed-loop synthesis. We conclude by outlining the main challenges ahead, namely hallucinations, interpretability, systemic biases, integration with clinical infrastructures, and regulatory and ethical requirements, and propose a roadmap for the development of scientific agents that are not only high-performing but also reliable, verifiable, and implementable in real biomedical contexts.

Overview

CoChat Summary

TL;DR

This review argues that bioinformatics is entering a third wave—agentic AI—defined by integrating universal biological foundation models, multi-agent architectures, and automated validation systems for end-to-end scientific discovery. It maps the required architecture, representative frameworks, and validation mechanisms, while stating that no published system has yet achieved complete three-pillar integration in a prospectively validated biomedical context.

Key findings

  • The proposed paradigm rests on three pillars: universal biological foundation models (e.g., scGPT, Nicheformer, EpiAgent), multi-agent architectures with planning, action, reflection, and memory (PARM), and automated validation systems incorporating sequential statistical testing, pipeline auditing, and robustness/reproducibility benchmarks.
  • Each pillar alone or in pairwise combination is argued to be insufficient for multistage, multimodal, regulated workflows: without Pillar I the system cannot reliably represent complex biological signals; without Pillar II it cannot orchestrate discovery; without Pillar III it cannot guarantee statistical reliability and regulatory acceptability.
  • Biomedical agent frameworks such as ClinicalAgent and Biomni are described as operationalizing PARM components; validation is structured on three levels—input/protocol validation, evidence/inference validation, and output/implementation validation—and applications span drug discovery, personalized medicine, protocol automation, rare-disease drug repurposing, and closed-loop synthesis.
  • The paper identifies five main challenges—hallucinations, interpretability, systemic biases, clinical integration, and regulatory/ethical requirements—and states that no currently published system has achieved complete three-pillar integration within a prospectively validated biomedical context.

Methods

  • Design: a review/perspective proposing a three-pillar paradigm for agentic AI in bioinformatics, not a primary empirical study.
  • Data or sample: no primary dataset or sample is analyzed; the paper uses an illustrative spinal muscular atrophy (SMA) drug-repurposing workflow and tables mapping representative foundation models, agent frameworks, applications, and validation strategies.
  • Main analysis: conceptual synthesis and taxonomy, including the Planning–Action–Reflection–Memory (PARM) framework, classification of agent frameworks by orchestration strategy, tool governance, and evidence traceability, and a layered automated-validation scheme.

Limitations

  • The paper states that no currently published system has yet achieved the complete three-pillar integration within a prospectively validated biomedical context, and that the SMA workflow is an operational blueprint rather than a completed implementation.
  • The paper states that most current biomedical studies fall within conceptual/architectural proposals or controlled evaluations, with only a limited subset providing robust external validation or generalizability across institutions.
  • Evident from the text: as a review/perspective, it reports no primary quantitative outcomes or prospective validation; cited productivity or time-reduction claims come from other work, and the main challenges—hallucinations, interpretability, bias, clinical integration, and regulatory gaps—remain unresolved.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1093/bib/bbag245

Cites

48 of 48 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Agent-based workflow in drug discovery. A scientific goal (e.g. identifying a drug for a rare disease) is decomposed by an orchestrator agent into subtasks assigned to specialized agents. The literature analyst retrieves relevant publications via RAG; the pipeline executor accesses and analyzes public omic datasets; the statistical validator assesses result robustness through hypothesis testing and uncertainty quantification. A shared memory layer tracks all intermediate outputs, i.e. queries, parameters, results, enabling iterative refinement through reflection loops until verifiable, evidence-supported hypotheses are generated. This architecture operationalizes PARM framework in a distributed multi-agent setting.
Figure 1 Agent-based workflow in drug discovery. A scientific goal (e.g. identifying a drug for a rare disease) is decomposed by an orchestrator agent into subtasks assigned to specialized agents. The literature analyst retrieves relevant publications via RAG; the pipeline executor accesses and analyzes public omic datasets; the statistical validator assesses result robustness through hypothesis testing and uncertainty quantification. A shared memory layer tracks all intermediate outputs, i.e. queries, parameters, results, enabling iterative refinement through reflection loops until verifiable, evidence-supported hypotheses are generated. This architecture operationalizes PARM framework in a distributed multi-agent setting.CC BY · PubMed Central
Figure 2 Multi-agent workflow for rare-disease drug repurposing. A supervisory planning/orchestrator agent decomposes the clinical objective (identifying a candidate drug for SMA) into a structured task graph. The literature analysis agent retrieves publications (e.g. via PubMed Entrez Programming Utilities, E-utilities), processes them through RAG pipelines, and stores structured outputs in a shared memory layer, while reflection monitors coverage and conflicts. The omics pipeline agent retrieves transcriptomic datasets from the GEO or ArrayExpress, performs preprocessing, differential expression, and pathway enrichment, with reflection evaluating stability across preprocessing variants. The validation/reflection agent integrates outputs, performs statistical consistency checks, uncertainty-aware aggregation, and sequential verification to reconcile conflicts across graph-based, molecular, or experimental predictions. All intermediate results are persisted with provenance, enabling reproducible, evidence-backed candidate prioritization (e.g. valproic acid) via cross-referencing with drug-target databases (e.g., DrugBank).
Figure 2 Multi-agent workflow for rare-disease drug repurposing. A supervisory planning/orchestrator agent decomposes the clinical objective (identifying a candidate drug for SMA) into a structured task graph. The literature analysis agent retrieves publications (e.g. via PubMed Entrez Programming Utilities, E-utilities), processes them through RAG pipelines, and stores structured outputs in a shared memory layer, while reflection monitors coverage and conflicts. The omics pipeline agent retrieves transcriptomic datasets from the GEO or ArrayExpress, performs preprocessing, differential expression, and pathway enrichment, with reflection evaluating stability across preprocessing variants. The validation/reflection agent integrates outputs, performs statistical consistency checks, uncertainty-aware aggregation, and sequential verification to reconcile conflicts across graph-based, molecular, or experimental predictions. All intermediate results are persisted with provenance, enabling reproducible, evidence-backed candidate prioritization (e.g. valproic acid) via cross-referencing with drug-target databases (e.g., DrugBank).CC BY · PubMed Central
Figure 3 Challenges versus mitigation strategies heatmap. Rows represent the five main challenges discussed in the text (hallucinations, interpretability, bias, clinical integration, regulatory compliance); columns represent five mitigation families (RAG, explainable AI, dataset auditing, secure APIs, continuous governance). Shading intensity indicates the current maturity level of each solution in addressing each challenge, based on literature and current implementations reviewed in this paper. The scale ranges from low maturity (critical gap, indicated by the lightest/least intense shading) to high maturity (resolved, indicated by the darkest/most intense shading). The heatmap provides an at-a-glance assessment of the field’s status and highlights priority areas for future research. For instance, while RAG shows high maturity for addressing hallucinations, its maturity for interpretability remains low; conversely, continuous governance frameworks show low maturity across all dimensions, underscoring the need for parallel technical and regulatory innovation.
Figure 3 Challenges versus mitigation strategies heatmap. Rows represent the five main challenges discussed in the text (hallucinations, interpretability, bias, clinical integration, regulatory compliance); columns represent five mitigation families (RAG, explainable AI, dataset auditing, secure APIs, continuous governance). Shading intensity indicates the current maturity level of each solution in addressing each challenge, based on literature and current implementations reviewed in this paper. The scale ranges from low maturity (critical gap, indicated by the lightest/least intense shading) to high maturity (resolved, indicated by the darkest/most intense shading). The heatmap provides an at-a-glance assessment of the field’s status and highlights priority areas for future research. For instance, while RAG shows high maturity for addressing hallucinations, its maturity for interpretability remains low; conversely, continuous governance frameworks show low maturity across all dimensions, underscoring the need for parallel technical and regulatory innovation.CC BY · PubMed Central

Verification

Claims the paper makes

28 Valid · 15 Not valid · 8 Not checked · 94 Own results and interpretations

57 more claims · 11 with their cited sources checked

Create a free account to see all 145 claims and to verify the 9 references still open.

See all claims — free account
  • Probably not a substantial claim. 62
  • Borderline: one of the tests for a substantial claim is unclear. 2
  • It may state two separate claims as one. 39

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Research works printed but not linked: 4

  1. Communications of the ACM, 60
  2. Yao, 2022, The Eleventh International Conference on Learning Representations
  3. Wu, 2024, First Conference on Language Modeling
  4. Hong, 2024, The Twelfth International Conference on Learning Representations

Cited by

Metadata

Authors
Branda, F., Ahmed, M. M., Ciccozzi, M., Guzzi, P. H., & Scarpa, F.
Affiliations
  • Università Campus Bio-Medico (IT)
  • SIMAD University (SO)
  • Magna Graecia University (IT)
  • University of Sassari (IT)
Corresponding author
Ahmed, Mohamed M
Venue
Briefings in Bioinformatics
Published
2026-05
Volume
27
Issue
3
ISSN
  • 1467-5463 (ISSN-L)
  • 1477-4054
OpenAlex
W7161752743
Cited by
2
References
35
Citations by year
2026: 2
FWCI
4.45
Type
journal article
Language
en
Open access
Locations
  • Briefings in Bioinformatics (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
Topics
  • Social Sciences › Decision Sciences › Information Systems and Management › Scientific Computing and Data Management
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
Keywords
agentic AI, multi-agent systems, scientific discovery, scGPT, hypothesis verification, retrieval-augmented generation, drug discovery, personalized medicine, drug repurposing, rare diseases, hallucinations, interpretability, systemic bias, clinical integration, AI ethics and regulation
MeSH terms
Generative Artificial Intelligence, Artificial Intelligence, Humans, Computational Biology, Drug Discovery, Precision Medicine

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.