Bioinformatics Workflow Decay: 80% of Taverna Pipelines Are Dead

Before LLMs entered the picture, the reproducibility crisis was already severe at the infrastructure level. An analysis of the myExperiment repository - once the primary hub for sharing bioinformatics pipelines - found that nearly 80% of tested Taverna workflows fail to execute or reproduce their original results (DOI: 10.48550/arXiv.2511.19510). Workflows published between 2007 and 2009 exhibit failure rates exceeding 80%, and the official retirement of the Taverna system by Apache in 2020 left thousands of validated scientific methods stranded.

The root cause is rarely incorrect science. Approximately 50% of workflow decay is attributed to volatile third-party resources: unavailable web services, inaccessible databases, and unannounced API changes. Many Taverna-era pipelines relied on SOAP-based web services that have since been replaced by RESTful APIs, rendering original service calls inoperable. Missing example data, insufficient environment documentation, and incomplete parameter records compound the problem.

A parallel study found that roughly 75% of code released alongside published research papers could not run without errors (DOI: 10.48550/arXiv.2511.19510). Even when researchers shared their code and data on request, external teams successfully reproduced scientific results in only about 60% of cases - and 93% of requests for data sharing went unanswered entirely (DOI: 10.48550/arXiv.2210.02593).

Citation Fabrication: 19.9% of GPT-4o References in Biomedical Research Are Invented

The most quantified failure mode of general-purpose LLMs in biomedicine is citation fabrication. In a systematic within-domain test of GPT-4o generating literature reviews on mental health topics, 19.9% (35 of 176) of citations were entirely fabricated - no identifiable source existed in any database (PMID: 41223407). Among the remaining real citations, only 54.6% were fully accurate; specific error types included incorrect DOIs (37.8%), wrong volume numbers (30%), and erroneous issue numbers (27.9%).

The fabrication rate was not uniform. Less-studied disorders like body dysmorphic disorder and binge eating disorder showed fabrication rates of 29% and 28% respectively, compared to 6% for the well-documented major depressive disorder. Specialised prompts amplified the problem: for binge eating disorder, specialised reviews had a 46% fabrication rate versus 17% for general overviews (PMID: 41223407).

Probabilistic Inconsistency: Why LLMs Fail Deterministic Biomedical Tasks

Reproducibility requires determinism: the same input must produce the same output. LLMs are probabilistic by design - identical prompts can yield different responses across runs unless parameters like temperature are strictly controlled (PMID: 38722813). In regulated environments, this is disqualifying.

A direct evaluation of GPT-3.5 and GPT-4 for medical named entity recognition (NER) found that even with fixed seeds and controlled parameters, both models failed to demonstrate reproducible results. The study concluded that the lack of reproducibility, combined with the opacity of externally hosted systems, undermines the use of proprietary models in GxP-validated workflows (PMID: 39661234).

Detecting LLM Confabulations in Biomedical Reasoning

LLMs do not merely make errors - they produce "confabulations," defined as claims that are wrong, arbitrary, and often scientifically plausible (PMID: 38898292). Semantic entropy methods, developed to detect these confabulations, compute uncertainty at the level of meaning rather than specific word sequences. The approach works across datasets and tasks without task-specific training data, making it applicable to novel biomedical questions.

However, the reasoning gap extends beyond text generation. The SylloBio-NLI benchmark tested LLMs on formal syllogistic reasoning with biomedical content and found that the Gemma-7b model achieved a faithfulness score below 0.1 - meaning it almost never adjusted its predictions appropriately when the truth value of a premise was altered (DOI: 10.48550/arXiv.2410.14399). Zero-shot accuracy on complex tasks like molecular cloning scenarios remained near or below random guessing (DOI: 10.48550/arXiv.2407.10362).

Emerging Safeguards: From RAG Pipelines to Standardised Reporting

The research community is converging on a set of complementary safeguards. Retrieval-Augmented Generation (RAG) architectures force models to link claims to verifiable primary sources - PMIDs and DOIs - rather than relying on parametric memory. This approach underlies tools that automate the generation of BioCompute Objects (BCOs) from published papers and code repositories, reducing the overhead of retroactive compliance with documentation standards (DOI: 10.48550/arXiv.2409.15076).

A seven-step "Safe and Transparent" workflow for LLM-assisted clinical trials mandates expert sign-off at checkpoints for literature selection and statistical verification (PMID: 41111869). Proposed reporting extensions - COREQ+LLM for qualitative research (PMID: 40991937) and PRISMA-AI for systematic reviews - aim to require disclosure of model versions, temperature settings, and prompting strategies so that AI-assisted syntheses become auditable.

At the infrastructure level, the Model Context Protocol (MCP) provides a standardised semantic layer that allows LLMs to query fragmented bioinformatics web services reliably. Implementations across GEO, STRING, and UCSC Cell Browser demonstrate that MCP can operationalise FAIR principles for autonomous agents (PMID: 41729821). Meanwhile, scholars advocate for open-source LLM infrastructure to prevent proprietary API costs from creating academic "caste systems" (PMID: 38722813).

The Research Landscape: From BioBERT to Autonomous Lab-Pilots

The scientific landscape has evolved through three distinct phases. The early phase (2019–2022) was dominated by domain-specific pre-training: BioBERT demonstrated that masked language modelling on PubMed abstracts was essential for effective named entity recognition (PMID: 31501885), and BioGPT established records in relation extraction by generating natural-language triplets rather than structured text (PMID: 36156661).

The stable phase (2022–2024) shifted to zero-shot clinical reasoning and standardised benchmarking. Med-PaLM, built on a 540-billion-parameter architecture, achieved state-of-the-art accuracy on the USMLE at 67.6% on MedQA - surpassing prior models by more than 17 percentage points (PMID: 37438534). This period also consolidated the hallucination problem: systematic reviews catalogued ethical, copyright, transparency, and legal concerns alongside the risk of bias and fabricated content (PMID: 36981544).

The emerging phase (2025–present) focuses on "agentic" systems and the scAInce paradigm. Research is entering a "co-pilot to lab-pilot" transition in which AI no longer merely interprets knowledge but increasingly generates and executes it (PMID: 40951330). The MCPmed initiative provides a standardised, machine-actionable layer for bioinformatics web services, preparing the infrastructure for next-generation research agents (PMID: 41729821).

Who Benefits from Citation-Grounded Biomedical AI

Frequently asked questions