Deterministic retrieval recovers biomedical associations lost by language models

Halder, A.; Singh, M.; Kesarwani, R.; Mathew, B.; Bhattacharya, N.; Chikhaliya, O.; Motwani, D.; Peela, S. C. M.; Samanta, S.; Muddemmanavar, P.; Farooq, M.; Ahuja, G.; Sengupta, D.

doi:10.64898/2026.04.25.720782

bioRxiv · 10.64898/2026.04.25.720782

Deterministic retrieval recovers biomedical associations lost by language models

Halder, A.·Singh, M.·Kesarwani, R.·Mathew, B.·Bhattacharya, N.·Chikhaliya, O.·Motwani, D.·Peela, S. C. M.·Samanta, S.·Muddemmanavar, P.·Farooq, M.·Ahuja, G.·Sengupta, D.

View original source ↗Export metadata ↓

Abstract

Biomedical research increasingly relies on expert-curated databases to connect diseases, genes, variants, phenotypes, pathways and therapeutics. Incomplete or irreproducible retrieval can distort interpretation, mechanistic inference, variant assessment or therapeutic prioritization despite an unchanged evidence base. Agentic natural-language-to-SQL (NL2SQL) systems and Model Context Protocol-enabled agents can autonomously retrieve evidence from biomedical databases, including Open Targets and the Highly Confident Drug-Target Database, in response to natural-language questions. However, they may truncate large result sets, miss records when terminology differs from database vocabulary and return different outputs across runs. We systematically quantify these failures and introduce BioChirp, a goal-directed autonomous system for reliable retrieval from structured biomedical databases, built on interpretation-execution separation. A coordinated language-model layer interprets intent and resolves database-field assignments, after which entity resolution maps user terms to database terminology. A deterministic Steiner-tree planner and executor then retrieve matching records without further language-model involvement. Across curated biomedical databases spanning more than one million associations, BioChirp retrieved thousands of records for exhaustive queries, whereas baseline agents returned at most a few hundred or failed. Across 70 queries, five runs and three BioChirp database backends, BioChirp achieved a median cross-run Jaccard similarity of 1.0. Across 910 database-question evaluations derived from expert-written BioASQ questions and spanning ten databases, BioChirp retrieved at least one database record in 85.4% of cases and produced a fully correct answer in 51.1% of cases, compared with 44.6% and 17.4% for agentic NL2SQL. BioChirp is publicly available at https://biochirp.iiitd.edu.in.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Halder, A., Singh, M., Kesarwani, R., Mathew, B., Bhattacharya, N., Chikhaliya, O., Motwani, D., Peela, S. C. M., Samanta, S., Muddemmanavar, P., Farooq, M., Ahuja, G., Sengupta, D.. 2026-04-29. Deterministic retrieval recovers biomedical associations lost by language models. https://doi.org/10.64898/2026.04.25.720782

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Targeted finetuning enables co-folding models to learn ligand-induced protein conformational states

Advances in protein structure prediction have enabled all-atom protein-ligand co-folding models that predict bound conformations directly from sequence and small-molecule structure. However, these models often fail to generalize to novel binding sites or alternative protein conformational states, limiting their utility for chemical biology and drug discovery. Here we show this limitation reflects training data bias rather than architectural constraints and can be overcome through targeted finetuning. Using ten previously unseen X-ray structures of Werner (WRN) helicase from a drug discovery program, we finetune Boltz-1 to learn both an allosteric binding site and a large conformational change locking the enzyme in an inactive state, while preserving accuracy on the ATP-bound state. The finetuned model generalizes to different chemical series and transfers the conformational logic across RecQ-family helicases in a binding-site sequence-dependent manner. This approach provides a blueprint for adapting foundation models as new structural and mechanistic data emerge, enabling co-folding networks to capture ligand-induced conformational switches and binding poses absent from their training data but central to biological regulation and therapeutic intervention.

bioinformatics↗

bioRxiv2026

Benchmarking single-cell foundation models for aging biology

Single cell foundation models (scFMs) provide representations of cellular states, but their utility across biological questions in aging research remains unclear. We established a benchmark of cellular representations for aging research, evaluating ten general-purpose scFMs, three aging-specific models and conventional methods across five biological questions using more than 2.5 million single cell transcriptomes. Using frozen pretrained representations, Geneformer performed best among scFMs for chronological age prediction and age pseudotime concordance, although 2,000 highly variable genes achieved higher mean performance. Several scFMs captured positive molecular age shifts across three disease contexts, consistent with reported aging-associated changes. SCimilarity performed well for rare cellular state identification across out-of-distribution datasets, exceeding aging specific models and conventional baselines. At the gene level, scGPT showed the highest recovery of reference TF target interactions, including aging-related regulatory hubs. Overall, scFMs supported diverse aging analyses, but performance depended on the biological question, highlighting their utility for rare cellular state identification and regulatory analysis.

bioinformatics↗

bioRxiv2026

CryoMV: Structure-Prior-Guided Modeling and Real-Particle Validation of Continuous Conformational Transitions in Cryo-EM

Continuous protein conformations are essential for understanding fundamental biological processes and supporting drug discovery. Although cryo-EM can resolve individual states at high resolution, recovering continuous heterogeneity from 2D particle images remains challenging. High noise, motion blur, and limited structural priors make it difficult to accurately generate and validate high-resolution continuous conformations using raw particle data. Here, we introduce cryoMV, a framework that integrates structure-prior-guided modeling with real-particle validation for continuous conformational transitions. CryoMV uses reference density maps to establish structural anchors and motion priors, models candidate transition paths between selected conformations, and transfers the learned representation to raw 2D cryo-EM particle images. Each candidate conformation is subsequently evaluated using the estimated particle poses and contrast transfer functions. Supported conformations are reconstructed through raw particle back-projection and assessed using canonical half-maps and Fourier shell correlation. On EMPIAR-10516 and EMPIAR-10345, cryoMV achieves excellent performance in terms of robustness, verifiability, and reconstruction resolution. By incorporating structure-prior modeling and evidence from the raw particles, cryoMV offers an explicit mechanism for assessing whether generated conformations are supported by experimental data and provides a practical approach to reducing model-induced artifacts in continuous cryo-EM heterogeneity analysis.

bioinformatics↗