Introduction
The exponential growth of biomedical literature has created a scalability crisis, in which the volume of information available surpasses the human capacity for synthesis [1]. Lactobacillus rhamnosus GG (LGG) is one of the most studied probiotic strains [2] and its immunological fingerprint evolved across three decades of research [3]. Artificial Intelligence (AI) emerges as a possible solution for redefining research methodologies [4].
Aims & Methods
This study aimed to validate a zero-cost, high-speed, hybrid (human-Artificial Intelligence) workflow using free Large Language Models (LLMs), accessible by any researcher regardless of institutional funding, using LGG literature as a benchmark case study.
An operative pipeline was developed based on four specialised prompts to interrogate three different LLMs architectures: Gemini 3 Flash (Mixture-of-Experts), Perplexity Academic (Retrieval-Augmented Generation), and Claude 4.6 Sonnet (Constitutional AI). LLMs screened over 35 years of PubMed literature (1989-2026). Performance was benchmarked against a human-validated Gold Standard using Sensitivity, Specificity, and Spearman Correlation. A "Semantic Floor" analysis was conducted to investigate the causes of False Negatives (FN), such as the "Title Trap" phenomenon.
Results
The pipeline demonstrated significant efficiency, reducing evidence synthesis time from weeks to days. Gemini 3 Flash showed superior sensitivity (0.93-0.97), while Perplexity Academic achieved near-perfect specificity (0.99) due to its Retrieval-Augmented Generation (RAG) architecture which mitigated hallucinations. The Spearman Heatmap revealed a correlation near zero between Gemini and Perplexity sessions, statistically validating their architectural complementarity.
The analysis identified a paradigm shift in LGG research: historical studies (pre-2019) focused primarily on mucosal barrier integrity and IgA production, whereas recent research (2019-2026) reveals a focus on complex molecular signaling, including IL-22, Type-I Interferons, and Inflammasome modulation. Despite this shift, a consistent core fingerprint was identified, characterised by TNF-α and IL-6 suppression alongside IL-10 and IFN-γ upregulation.
| Model | Architecture | Role in Pipeline | Key performance metric |
| Gemini 3 Flash | Mixture-of-Experts | Rapid Screening | Sensitivity: 0.97 |
| Perplexity Academic | Multistage Retrieval-Augmented Generation | Quality validation | Specificity: 0.99 |
| Claude 4.6 Sonnet | Constitutional AI | Refinement of complex cases (errors from Gemini and/or Perplexity) | Specificity: 0.88 |
Conclusion
The synergy between different LLMs’ architectures allows the researcher to act as an AI-orchestrator, focusing on critical interpretation rather than on mechanical extraction. This methodology democratises high-level research, providing a replicable solution to the scalability crisis. Additionally, we identified a shift in LGG’s research, which moved from mucosal barrier protection to complex molecular signaling.
References
[1] Scientific literature: Information overload | Nature. (n.d.). Retrieved 17 April 2026, from https://www.nature.com/articles/nj7612-457a
[2] Segers, M. E., & Lebeer, S. (2014). Towards a better understanding of Lactobacillus rhamnosus GG--host interactions. Microbial Cell Factories, 13 Suppl 1(Suppl 1), S7. https://doi.org/10.1186/1475-2859-13-S1-S7
[3] Capurso, L. (2019). Thirty Years of Lactobacillus rhamnosus GG: A Review. Journal of Clinical Gastroenterology, 53, S1. https://doi.org/10.1097/MCG.0000000000001170
[4] Ge, L., Agrawal, R., Singer, M., Kannapiran, P., De Castro Molina, J. A., Teow, K. L., Yap, C. W., & Abisheganaden, J. A. (2024). Leveraging artificial intelligence to enhance systematic reviews in health research: Advanced tools and challenges. Systematic Reviews, 13, 269. https://doi.org/10.1186/s13643-024-02682-2
