The field of medical AI presently produces papers with the vigor of a steam press and, alas, many contain more smoke than locomotive - but AI-CURA is the uncommon specimen that made me put down my tea.

In Science Translational Medicine, Ma and colleagues describe AI-CURA, a workflow that nearly automates the classification of genetic variants using the ACMG/AMP guidelines and ClinGen recommendations (DOI: 10.1126/scitranslmed.adz4172). This is not merely another chatbot wearing a white coat and hoping nobody asks for its license. It is a structured system that divides the labor: ordinary bioinformatics tools handle evidence that can be checked mechanically, while a large language model tackles the literature-heavy parts, where papers must be read, weighed, and summarized without wandering into the shrubbery.

The Beast Called a Variant

A genetic variant is a small difference in DNA. Most are harmless. Some cause disease. Many sit in the maddening middle drawer labeled variant of uncertain significance, or VUS, which is science-speak for "we found something, but please do not ask us to explain it before lunch."

Clinicians classify variants into categories such as benign, likely benign, uncertain significance, likely pathogenic, and pathogenic, following the 2015 ACMG/AMP framework (Richards et al., 2015). ClinGen then refines how those rules should be applied across genes and diseases through expert guidance (ClinGen Variant Classification Guidance). The work is careful, slow, and evidence-hungry. One must consult population databases, computational predictions, family studies, functional assays, clinical reports, and the literature. It is less "press button, receive diagnosis" and more "assemble a legal case against one suspicious nucleotide."

Enter the Mechanical Naturalist

AI-CURA’s sensible trick is that it does not ask the LLM to do everything. That would be like appointing a parrot to run the Royal Society because it speaks with confidence. Instead, the workflow separates non-literature evidence, which conventional bioinformatics tools can gather, from literature-based evidence, where reading comprehension matters.

The researchers tested two reasoning models, DeepSeek-R1 and o3-mini-high, on ACMG rules requiring interpretation of scientific papers. With carefully engineered prompts and rule-specific knowledge bases, DeepSeek-R1 performed better in their experiments, reaching high sensitivity and 100% specificity for literature-based rule interpretation. Specificity matters here: in clinical genetics, a confidently wrong answer is not a charming eccentricity. It is the laboratory equivalent of labeling a jar "probably soup" and serving it to a patient.

They then tested the system on 150 variants curated by ClinGen experts and found high concordance with human curators in final classification. They also tried AI-CURA on 150 ClinVar variants with conflicting interpretations, where the system could support reanalysis. That last bit matters because ClinVar contains a great many variants whose interpretations disagree across submitters, and disagreement is where curators spend much of their candlelight.

Why This Is More Than a Fancy Filing Cabinet

Recent work has been circling the same quarry. AutoPM3 used open-source LLMs to extract PM3 evidence from scientific literature and reported strong performance on a ClinGen-derived benchmark (Li et al., 2025). Another study used RAG and fine-tuning to connect GPT models with 190 million variant annotations, finding retrieval more useful than simply cramming facts into model weights like socks into a valise (Lu and Cosgun, 2025). VariantBench, meanwhile, argues that we should test not only whether models choose the right label, but whether their justifications make sense (Basharat et al., 2025).

AI-CURA fits into this movement but pushes closer to a full clinical workflow. It treats the LLM less as an oracle and more as a junior curator with a checklist, a stack of papers, and strict instructions not to improvise opera.

And yes, the PDF problem is real. Anyone who has wrestled genetic evidence out of supplementary tables knows the pain. Private browser-based tools like pdfb2.io live in the same practical universe: before AI can reason over documents, someone must first tame the documents, preferably without sacrificing them to a cloud cauldron.

The Proper Amount of Astonishment

Let us not uncork the champagne with a sword. AI-CURA still needs broader validation across more genes, diseases, variant types, ancestry groups, and real laboratory settings. Models drift. Literature is messy. Guidelines change. A workflow that performs beautifully in a curated test can still trip over the botanical garden of actual clinical practice.

Yet the study points to a useful future: not AI replacing geneticists, but AI doing the exhausting first pass so experts can spend more time on judgment. If reproducible at scale, that could speed rare disease diagnosis, clean up conflicting ClinVar records, and reduce the backlog of VUS reanalysis. The human curator remains the natural philosopher. The model becomes the tireless apprentice, squinting through the literature by lamplight and trying not to hallucinate a citation from 1893.

References

Ma W, Fong G, Lai J, et al. AI-CURA, an automated LLM workflow for high-accuracy genetic variant classification. Science Translational Medicine. 2026. https://doi.org/10.1126/scitranslmed.adz4172

Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants. Genetics in Medicine. 2015. https://doi.org/10.1038/gim.2015.30

Li S, Wang Y, Liu C-M, et al. AutoPM3: enhancing variant interpretation via LLM-driven PM3 evidence extraction from scientific literature. Bioinformatics. 2025. https://doi.org/10.1093/bioinformatics/btaf382

Lu S, Cosgun E. Boosting GPT models for genomics analysis. Bioinformatics Advances. 2025. https://doi.org/10.1093/bioadv/vbaf019

Basharat H, Plotkin S, Le C, Zhu K, Pink M, Alfaro I. VariantBench: A Framework for Evaluating LLMs on Justifications for Genetic Variant Interpretation. IJCNLP-AACL 2025. https://doi.org/10.18653/v1/2025.ijcnlp-srw.26

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.