Draft:BioMate AI

BioMate AI (also referred to as BioMate) is a software-as-a-service computational biomedicine platform developed by BioMate AI, Inc. that enables researchers without specialist bioinformatics training to run complex computational analyses through plain-language requests. The platform uses large language model-based agents to route research requests to validated bioinformatics and drug discovery workflows, execute them on cloud infrastructure, apply evidence-graded quality control, and return structured, audit-ready results. Its significance lies in removing the specialist bottleneck that has historically required separate bioinformaticians, computational chemists, and clinical pharmacologists at each stage of a research project. BioMate AI, Inc. was founded in 2025 and is headquartered in Houston, Texas, within the Texas Medical Center.

The platform indexes more than 4,000 validated workflows across 36 biomedical domains, including genomics, transcriptomics, drug discovery, structural biology, proteomics, and cryo-electron microscopy. It integrates established open-source tools such as GATK, STAR, Salmon, Seurat, AlphaFold, CryoSPARC, and AutoDock Vina, covering the full arc from raw sequencing data to regulatory-submission-ready reports.

Background

Drug discovery and biomedical research depend on dozens of specialized computational tools — RNA-seq analysis, variant calling, molecular docking, ADMET profiling, PBPK simulation — each with distinct software, parameters, and domain expertise requirements. Historically, researchers have needed bioinformaticians, computational chemists, and clinical pharmacologists as separate specialists to operate these tools.

Independent research has identified the gap between language model reasoning and deterministic computational execution as a core obstacle in AI-assisted biomedicine. A June 2026 benchmark study from Anthropic (VirBench) evaluated six frontier AI systems on virus sequence retrieval tasks and found that without deterministic tool integration, accuracy ranged from 16.9% to 91.3% across systems with high run-to-run variance; adding a single structured retrieval tool raised all systems above 90% accuracy — demonstrating that reliable biological AI requires validated tool calls, not language model generation alone. A concurrent study published in Nature described Robin, an AI research system developed by FutureHouse, which demonstrated strong performance in literature synthesis and hypothesis generation but required a separate code execution component (Finch) to perform statistical data analysis tasks such as RNA-seq processing and figure generation — work the language model alone could not reliably complete.

BioMate AI addresses this execution gap by coupling language model intent-parsing directly to validated, deterministic workflow execution on cloud infrastructure — connecting a researcher's plain-language request to an indexed, containerized pipeline, running it on AWS Batch, applying QC gates to the outputs, and returning structured results. Both the VirBench and Robin findings independently support the architectural premise that language model reasoning and computational execution must be tightly integrated in biomedical AI rather than treated as interchangeable. Beyond workflow execution, the platform also supports literature survey (retrieval-grounded synthesis across PubMed, bioRxiv, ChEMBL, and ClinicalTrials.gov) and experimental protocol generation (step-by-step wet-lab procedures derived from actual run parameters, formatted as SOPs).

Company

BioMate AI, Inc. was founded in 2025 by Yaoyun Zhang, who serves as CEO. The company is based at the Texas Medical Center in Houston, Texas. The platform is available under a freemium model, with a free tier, a Pro subscription at $49 per month, a Team subscription at $199 per month, and an Enterprise tier with custom pricing.

Platform

Workflow routing

BioMate uses a hybrid search architecture combining vector embeddings, domain-specific scoring, and large language model reasoning to map a natural-language request to the most appropriate workflow in its index. For example, a request to "run differential expression on RNA-seq samples with batch correction" is routed to the nf-core/rnaseq pipeline followed by nf-core/differentialabundance, with parameters inferred from the user's sample metadata.

Execution infrastructure

Workflows execute on managed cloud compute infrastructure (AWS Batch). GPU-accelerated workflows — including cryo-EM reconstruction and AlphaFold structure prediction — use GPU-equipped instances. Results are stored in cloud object storage and linked to the run provenance record.

Quality control and auto-remediation

The platform applies configurable quality control gates to workflow outputs, using a Gold / Silver / Bronze grading system referenced against published community standards. The importance of standardized, tiered QC in high-throughput biomedical data analysis has been established by large-scale consortium projects: the ENCODE Project formalized tiered data quality standards for genome-wide assays, and clinical sequencing guidelines from the ACMG established minimum QC thresholds for next-generation sequencing in diagnostic contexts. When a gate fails — for example, a hERG flag in an ADMET screen, or a low resolution estimate in cryo-EM reconstruction — the agent proposes a parameter change or alternative approach and re-runs automatically without manual intervention.

Drug discovery workflows

BioMate supports multiple phases of the drug discovery process, including:

• Target identification — AlphaFold structure prediction, pocket druggability analysis via fpocket, and CRISPR essentiality scoring using DepMap data
• Virtual screening — large-scale molecular docking with AutoDock Vina and pharmacophore filtering
• ADMET profiling — prediction of absorption, distribution, metabolism, excretion, and toxicity properties, including CYP inhibition and hERG liability
• PBPK simulation — physiologically based pharmacokinetic modeling for first-in-human dose estimation
• IND document assembly — automated drafting of FDA Investigational New Drug application sections §2.6.1–§2.6.5 from underlying study data
• Bispecific antibody format triage — scoring of bispecific antibody formats (CrossMab, IgG-scFv, BiTE, DART, KiH) across six axes including CMC complexity, immunogenicity, and manufacturability
• In vivo CAR-T co-design — simultaneous design of CAR target, lipid nanoparticle tropism, and CAR construct with cross-referenced fratricide safety checking
• GLP-1 receptor modality selection — multi-axis scoring of GLP-1 receptor agonist candidates across pharmacology, receptor profile, liver selectivity, manufacturing complexity, and clinical evidence
• Modality triage — a two-workflow chain querying single-cell expression atlases to recommend therapeutic modality (mAb, bispecific, ADC, CAR-T, PROTAC, ASO, or small molecule) from target expression data

Bioinformatics workflows

The platform integrates established open-source tools for standard bioinformatics analyses:

• RNA-seq — nf-core/rnaseq pipeline using STAR or HISAT2 alignment and Salmon quantification, followed by differential expression analysis with DESeq2, edgeR, or limma
• Whole-genome sequencing — nf-core/sarek pipeline for germline and somatic variant calling using GATK HaplotypeCaller and DeepVariant
• Single-cell RNA-seq — Seurat and Scanpy workflows with Cell Ranger preprocessing, clustering, and cell-type annotation using SingleR
• Cryo-EM — single-particle analysis with CryoSPARC and RELION, including GPU-accelerated 2D classification, 3D refinement, and resolution estimation
• Structural biology — AlphaFold structure prediction, molecular docking with AutoDock Vina, and molecular dynamics simulation with GROMACS and OpenMM
• Epigenomics / DNA methylation — bisulfite sequencing analysis using minfi and DMRcate, with differentially methylated region calling and pathway enrichment
• Metagenomics — QIIME 2 amplicon and shotgun metagenomics
• Proteomics — MaxQuant label-free quantification and metabolic network modeling

Bioconductor integration

BioMate incorporates 1,818 Bioconductor packages, making bioinformatics statistical methods available through plain-language requests without requiring users to write code.

The Bioconductor ecosystem — comprising over 2,300 R packages for bioinformatics — presents a known challenge for large language models: the APIs are complex, heavily object-oriented (S4 class system), and underrepresented in general-purpose training corpora, leading to high rates of hallucinated or non-functional code generation. To address this, BioMate has published BioMate-KB, a knowledge base of 15,641 real-execution-validated workflow steps extracted across the Bioconductor 3.20 ecosystem, annotated with NAMESPACE validation, EDAM ontology terms, and container links. The top 200 packages — collectively accounting for 57% of all Bioconductor downloads — have been released as an open-access GitHub repository providing structured, machine-readable workflow definitions reusable by AI systems, workflow engines, and educators. The repository has received more than 400 community stars.

A companion preprint introduces an execution-grounded validation hierarchy addressing how such knowledge bases should be validated. It finds that structural grounding (NAMESPACE verification, S4 type checking, vignette cross-referencing) raises LLM function-citation accuracy from 71.4% to 88.2%, but that real end-to-end execution on representative data remains necessary as the final ground truth.

Data integration and connectors

The platform supports connections to laboratory instruments and information systems, including Illumina BaseSpace, Oxford Nanopore MinKNOW, Benchling LIMS, and standard laboratory instruments (flow cytometers, LC-MS, qPCR, plate readers). It implements the Model Context Protocol (MCP) for integration with compatible AI assistant environments.

Compliance and security

BioMate AI states that the platform is HIPAA-compliant and offers Business Associate Agreement (BAA) support for healthcare and life science enterprise customers. A private VPC deployment option is available for customers requiring data residency controls. The company states that tenant data is not used to train shared AI models.

Benchmarks

BioMate publishes benchmark results on its website, reporting 94.6% workflow routing accuracy, a 100% pass rate on PBPK validation cases, and an 87.1% score on regulatory language model evaluation.

The platform's modality triage workflow chain correctly predicted the therapeutic modality for seven of seven oncology drugs approved by the FDA in Q4 2024, using only the target gene name and indication as input.

Relationship to existing platforms

BioMate leverages open-source community resources extensively. The platform can accommodate existing Nextflow scripts and Galaxy workflows, allowing researchers to bring established pipelines and run them through BioMate's AI-guided interface. Community tools such as GATK, STAR, DESeq2, AlphaFold, and the full nf-core pipeline catalog are available through plain-language requests. The platform also integrates community reference databases including gnomAD, ClinVar, GTEx, Reactome, KEGG, STRING, and OpenTargets as queryable resources within workflows.

BioMate also covers drug discovery and laboratory informatics workflows that overlap with commercial offerings from Schrödinger, Dotmatics, and Benchling, with a focus on AI-orchestrated execution across the full discovery pipeline rather than standalone desktop tools.

External links

• Official website
• Bioconductor workflow repository (GitHub)
• Wikidata entity Q140074014