BioMate AI
Newsroom

Publications

The methodology and validation behind BioMate, published openly. Organized by topic and month — newest first.

Open knowledge base — free under CC-BY-4.0

BioMate-KB: 15,641 real-execution-validated Bioconductor workflow steps. The public skill bundle is on GitHub for the community to clone, use, and build on.

Cyclic-peptide drug design

November 2026

A knowledge-aware, constraint-guaranteed decoder for cyclic peptides

Zhang, Y., Li, X., Zhang, Y., Dike, A. (2026) · Zenodo

A cyclic-peptide generator that compiles the design rules — macrocycle closure, ring size, N-methyl count, molecular-weight bound and structural alerts — directly into a token-level decoding automaton, so every molecule it completes is valid by construction. Delivers about three times the valid, novel yield of monomer-library enumeration at a matched sampling budget, evaluated under a pre-registration that fixes margins, controls and verdicts in advance.

Zenodo DOI: 10.5281/zenodo.24112462 Read →

Chain to loop: cost-aware value-of-information scheduling for cyclic-peptide discovery

Zhang, Y., Li, X., Zhang, Y., Dike, A. (2026) · Zenodo

Automated discovery must decide not only which candidate to test next but which assay to run on it, across a cascade whose costs differ by orders of magnitude. A cost-aware value-of-information scheduler chooses the candidate-and-assay pair that most reduces advance-or-kill uncertainty per unit cost — beating a fair cheapest-first cascade on real cyclic-peptide permeability and binding cascades, judged by a cluster bootstrap over whole laboratories and targets.

Zenodo DOI: 10.5281/zenodo.24114055 Read →

A benchmark framework to guide automated cyclic peptide discovery

Zhang, Y., Li, X., Zhang, Y., Dike, A. (2026) · Zenodo

Automated cyclic-peptide discovery rests on a benchmark that can tell real learning from an artefact. This framework separates the questions with coordinated controls — frozen-model, one-shot, shuffled-feedback and retrieval comparisons; acquisition versus held-out-source ranking; and planted-valid and planted-shortcut tests of the validity gate itself. It finds that most of the selection advantage precedes feedback, that held-out-source ranking barely transfers, and that a null-calibrated gate holds its false-alarm rate where a registered gate flags clean benchmarks: trust comes from controls, not a headline metric.

Zenodo DOI: 10.5281/zenodo.24112627 Read →

AI for Bioinformatics & Executable Knowledge Bases

June 2026

BioMate-KB: A Real-Execution-Validated Workflow Knowledge Base for Bioconductor

Zhang, Y. (2026) · Zenodo

A curated knowledge base of 15,641 workflow steps extracted across the Bioconductor 3.20 ecosystem — NAMESPACE-validated, EDAM-annotated, container-linked, and, for the validated head, confirmed by real end-to-end execution. Reframes executability as an empirical, reproducible property.

Zenodo DOI: 10.5281/zenodo.20616355 Read →

Structure Grounding Is Not Enough: Real Execution as the Ground Truth for LLM-Generated Bioinformatics Workflows

Zhang, Y. (2026) · Zenodo

Bioconductor's formal structure (NAMESPACE, S4 types, BiocViews, vignettes) suppresses LLM hallucination — raising function-citation accuracy from 71.4% to 88.2% — but does not guarantee execution. Introduces execution-grounded validation as the strongest tier of a lexical ⊂ structural ⊂ executional grounding hierarchy.

Zenodo DOI: 10.5281/zenodo.20616543 Read →

Citations use the Zenodo concept DOIs (all versions). To pin v3 exactly: BioMate-KB 10.5281/zenodo.20616356 · Structure Grounding 10.5281/zenodo.20616544.

BioMate AI

Try BioMate free.

Move faster from scientific question to actionable results.