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.
The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery
Open accessAbstract
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
Cites
Statistics
Figures



Verification
Claims the paper makes
28 Valid · 15 Not valid · 8 Not checked · 94 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“Foundation models for single-cell and multi-omic biology, such as scGPT [20], Nicheformer [21], and EpiAgent [22, 23], produce high-quality continuous representations of biological entities that support cross-tissue transfer, perturbation inference, and regulatory feature learning.”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction > Limitations of single-pillar approaches and the necessity of integration
What the sources say
- Supportedcites [20] Cui H, Wang C, Maan H et al. scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nat Methods 2024; 21: 1470– 80. 10.1038/s41592-024-02201-0 38409223 source
The Results (section 2.1, p. 3) identify scGPT as a single-cell foundation model that learns cell and gene embeddings and supports perturbation prediction and gene regulatory network inference. The multi-batch integration results (section 2.4) explicitly describe transferability of features learned from whole-human data to a specific tissue. This verdict concerns scGPT [20], not the separately cited Nicheformer or EpiAgent. (LLM, conf=0.90)
- Partly supportedcites [21] Tejada-Lapuerta A, Schaar AC, Gutgesell R et al. Nicheformer: a foundation model for single-cell and spatial omics. Nat Methods 2025; 22: 2525– 38. 10.1038/s41592-025-02814-z 41168487 PMC12695652 sourceExperimental — accuracy not yet measured
The Main and Discussion (page unavailable) describe Nicheformer as learning cell representations that transfer spatial information between single-cell and spatial data; the Results also report attention patterns that generalize across analyzed tissues. The paper discusses perturbation prediction and gene-regulatory inference as tasks for single-cell foundation models generally, but does not establish that Nicheformer's representations support those tasks. Its perturbation test concerns embedding stability under missing or reordered genes, not perturbation inference. (LLM, conf=0.86)
- Disputedcites [22] Kedzierska KZ, Crawford L, Amini AP et al. Zero-shot evaluation reveals limitations of single-cell foundation models. Genome Biol 2025; 26: 101. 10.1186/s13059-025-03574-x 40251685 PMC12007350 sourceExperimental — accuracy not yet measured
The Background (page unavailable) says this paper tested claims that scGPT produces robust embeddings and generalizes to unseen datasets, but found that scGPT and Geneformer underperformed simpler methods. The Results and discussion (page unavailable) likewise report inconsistent zero-shot embedding performance and failure to correct batch effects between techniques. The paper does not evaluate Nicheformer or EpiAgent, so its contrary finding is limited to the models it tested. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
- Not checked — no full text of the sourcecites [23] Ma Q, Jiang Y, Cheng H et al. Harnessing the deep learning power of foundation models in single-cell omics. Nat Rev Mol Cell Biol 2024; 25: 593– 4. 10.1038/s41580-024-00756-6 38926531 source
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The first wave saw the emergence of deep learning as the dominant paradigm for analyzing high-dimensional biological data.”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction
What the sources say
- Partly supportedcites [1] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Communications of the ACM. 2017; 60:84–90. sourceExperimental — accuracy not yet measured
The Prologue says SuperVision triggered a paradigm shift in computer vision, and the Epilogue describes the computer vision community switching to deep neural networks (page numbers unavailable). Neither passage establishes that deep learning became the dominant paradigm for analyzing high-dimensional biological data. (LLM, conf=0.86)
- Not checked — no full text of the sourcecites [2] Poplin R, Chang PC, Alexander D et al. A universal SNP and small-indel variant caller using deep neural networks. Nat Biotechnol 2018; 36: 983– 7. 10.1038/nbt.4235 30247488 source
- Not checked — no full text of the sourcecites [3] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015; 521: 436– 44. 10.1038/nature14539 26017442 source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“while systems such as AlphaFold have demonstrated that the integration of deep learning and physical-structural constraints can achieve accuracy comparable to that of experiments in predicting three-dimensional structure [9].”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction
What the sources say
- Supportedcites [9] Jumper J, Evans R, Pritzel A et al. Highly accurate protein structure prediction with AlphaFold. nature. 2021; 596: 583– 9. 10.1038/s41586-021-03819-2 34265844 PMC8371605 source
The AlphaFold network section says the model combines deep learning with physical and geometric constraints to predict 3D protein coordinates. The Main section reports near-experimental accuracy in a majority of cases in the CASP14 assessment, supporting the claim that it can achieve accuracy comparable to experiments. (LLM, conf=0.93)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In parallel, foundation models for single-cell omics have begun to produce universal embeddings of cells and genes that are robust to batch effects and technical variability and transferable across tissues, conditions, and platforms [10, 11].”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction
What the sources say
- Partly supportedcites [10] Theodoris CV, Xiao L, Chopra A et al. Transfer learning enables predictions in network biology. Nature. 2023; 618: 616– 24. 10.1038/s41586-023-06139-9 37258680 PMC10949956 sourceExperimental — accuracy not yet measured
The ‘Context-awareness and batch integration’ section reports both gene and cell embeddings that resist tested batch-related artifacts, including patient variability, and cell embeddings that transfer between two sequencing platforms. ‘Geneformer architecture and pretraining’ describes training on cells from a broad range of tissues, but the cited passages do not establish that the embeddings are universal or demonstrate transfer across tissues and conditions. (LLM, conf=0.85)
- Partly supportedcites [11] Zhang F, Chen H, Zhu Z et al. A survey on foundation language models for single-cell biology. In: Che W, Nabende J, Shutova E, Pilehvar MT. (eds), Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, Vienna, Austria: Association for Computational Linguistics, Vol. 1, pp. 528– 49. Long Papers, 2025. sourceExperimental — accuracy not yet measured
The Introduction says single-cell foundation language models produce robust, generalizable cell representations across datasets and tasks, and the Continuous Embeddings section describes gene embeddings. But the Positional Information/Batch Effect section says sequencing technologies, platforms, and laboratory conditions cause severe batch effects that remain a challenge for unified representations; it does not establish the claimed robustness to that variability or transfer across all the specified settings. No page number is available in our copy for these passages. (LLM, conf=0.88)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Language models for proteins have shown that structural and functional principles emerge spontaneously from training on millions of unannotated sequences [8]”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction
What the sources say
- Supportedcites [8] Rives A, Meier J, Sercu T et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc Natl Acad Sci 2021; 118: e2016239118. 10.1073/pnas.2016239118 33876751 PMC8053943 source
The Discussion reports training protein language models on 250 million sequences and says that knowledge of intrinsic biological properties emerges without supervision, with no learning signal beyond sequences during pretraining. The Abstract identifies both structural properties and biological activity in the learned representations. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
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.
- 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
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Universal representations and operational considerations in agentic systems
“Complementary methodological criticalities and validation strategies are summarized in Table 5.”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Use cases of agentic AI in bioinformatics
“The representative applications summarized in Table 5 highlight that the most immediate impact of agentic AI lies in compressing the information bottleneck that characterizes the early phase of drug discovery.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Frameworks for biomedical AI agents
“we classify representative agent frameworks according to orchestration strategy, tool-governance design, and evidence-traceability mechanisms, as summarized in Table 3.”
§ The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery > Introduction > Limitations of single-pillar approaches and the necessity of integration
“The historical analogy with prior paradigm-defining reviews is summarized in Table 3, situating the present three-pillar framework within the broader context of computational biology and AI methodology.”
Citations
Cites
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
LeCun, Y., Bengio, Y., & Hinton, G. (2015) · Nature · doi:10.1038/nature14539 · Cited by 84,126
Highly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764 · cited 2×
“In structural biology, agents may orchestrate structure prediction, molecular docking, and binding affinity estimation, iteratively validating results against experimental databases and triggering re-runs when confidence thresholds are not met [8, 9].”
The FAIR Guiding Principles for scientific data management and stewardship
Wilkinson, M. D., Dumontier, M., Aalbersberg, I. J., Appleton, G., Axton, M., Baak, A., Blomberg, N., Boiten, J.-W., da Silva Santos, L. B., Bourne, P. E., et al. (2016) · Scientific Data · doi:10.1038/sdata.2016.18 · Cited by 18,423 · cited 3×
“Reflection provides continuous iterative error control, uncertainty monitoring, and dynamic replanning capabilities, while memory ensures context persistence and complete provenance tr ackingthroughacombinationofshort-termworkingstatesandlong-term retrieval mechanisms [ 17, 18, 32].”
Neural Machine Translation by Jointly Learning to Align and Translate
Bahdanau, D., Cho, K., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1409.0473 · Cited by 16,213
Fast, sensitive and accurate integration of single-cell data with Harmony
Korsunsky, I., Millard, N., Fan, J., Slowikowski, K., Zhang, F., Wei, K., Baglaenko, Y., Brenner, M., Loh, P.-R., & Raychaudhuri, S. (2019) · Nature Methods · doi:10.1038/s41592-019-0619-0 · Cited by 10,488 · cited 1×
“…another natural application area, where specialized agents analyze transcriptomic, epigenomic, proteomic, and spatial datasets, reconciling heterogeneous signals through shared representation spaces, a task particularly suited to foundation models capable of cross-modal alignment [12, 43, 44].”
1,500 scientists lift the lid on reproducibility
Baker, M. (2016) · Nature · doi:10.1038/533452a · Cited by 4,209 · cited 3×
“Each stage introduces degrees of analytical freedom and potential sources of bias, contributing to the reproducibility crisis documented in the scientific literature [19].”
Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C. L., Ma, J., et al. (2021) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.2016239118 · Cited by 3,286
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101
On the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
Research works printed but not linked: 4
- Communications of the ACM, 60
- Yao, 2022, The Eleventh International Conference on Learning Representations
- Wu, 2024, First Conference on Language Modeling
- Hong, 2024, The Twelfth International Conference on Learning Representations
Cited by
Ang, M. Y., Lipovich, L., Choo, S. W., Chen, L., & Song, L. (2026) · Biology · doi:10.3390/biology15171537 · Cited by 2
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
- Publisher
- Oxford University Press (OUP)
- Published
- 2026-05
- Volume
- 27
- Issue
- 3
- ISSN
- 1467-5463 (ISSN-L)
- 1477-4054
- PMC
- PMC13189163
- OpenAlex
- W7161752743
- Cited by
- 2
- References
- 35
- Citations by year
- 2026: 2
- FWCI
- 4.45
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- 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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery. https://app.cochat.ai/papers/10.1093/bib/bbag245
- The paper
- Branda, F., Ahmed, M. M., Ciccozzi, M., Guzzi, P. H., & Scarpa, F. (2026). The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery. Briefings in Bioinformatics. https://doi.org/10.1093/bib/bbag245
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.