YaoyunHug/realbio_benchmark

RealBio — an open, objectively-scored benchmark for AI agents in drug development & genomics

The one-sentence version

Modern drug discovery and genomics run on multi-step computational pipelines — pick the right analysis, configure it, quality-control the data, model the drug, choose a safe dose. Teams increasingly want an AI agent to drive those pipelines. RealBio asks the blunt question: can an AI agent actually run one correctly end-to-end, or does it only talk about biology fluently?

Why this is hard — and why we can't just trust the vendors' own scores

A large language model can write a confident paragraph about RNA-seq or pharmacokinetics. That is not the same as correctly routing a real request to the right pipeline, filling its parameters without inventing wrong ones, catching a bad sequencing sample, predicting a drug's blood concentration, or picking a safe first-in-human dose — where a mistake wastes real lab money or, in the clinical tasks, is a patient-safety error. Yet almost every agentic-bio benchmark is scored by each system's own AI judge on its own private task set, so the numbers are non-comparable and invite (often unintentional) self-grading.

RealBio fixes that: 505 fixed public items across 5 real drug-development / genomics tasks, each with objective, published ground truth, graded by one shared open scorer (src/score.py — a deterministic checker, no AI judge). Same items, same metric, for every system — so any team can run its own agent and land a directly comparable number.

The headline finding: today's frontier LLMs are strong on biological knowledge but fail at execution — and they fail worst on the clinical-pharmacology tasks (predicting drug exposure, choosing a trial dose), exactly where errors matter most. A purpose-built platform (BioMate) leads every task.

🌐 Project: biomate.ai · 🏆 Leaderboard: on GitHub · 📄 Datasheet: DATASHEET.md

Growing benchmark. Three further tasks (execution auto-repair, workflow generation, drug-discovery ligand ranking) are validated and re-added as every system — including BioMate — is measured on the same items with the same metric. We report a task only when the whole board is measured on it.

Why RealBio exists — and what "running a pipeline" actually means

Taking a therapy from a molecule or a dataset toward the clinic is a computational relay. RealBio's five tasks are five real hand-offs in that relay, in order:

1. Route — turn a plain-English goal ("call germline variants from whole-genome sequencing") into the correct analysis pipeline, out of hundreds.
2. Configure — set that pipeline's parameters from what the user actually said, without inventing settings they didn't specify.
3. Quality-control — decide whether the incoming sequencing data is usable, using the rule that fits its specific assay type (a single universal cutoff false-alarms).
4. Model the drug — predict how much drug reaches the bloodstream (its pharmacokinetics) from the molecule's chemistry.
5. Dose the trial — from emerging toxicity data, recommend the dose to give the next patients in a Phase I clinical trial.

An AI agent that is fluent about biology but gets any of these wrong isn't just unhelpful — it burns real sample and compute budget, and in steps 4–5 it is making drug-safety calls. So the question isn't "does it sound knowledgeable," it's "did it get the objectively-checkable answer right." That is what RealBio measures, identically, for every system: fixed public items, published ground truth, one open deterministic scorer — no AI judge, no private test set, no self-grading.

Repository layout

├── <task>/data.jsonl        # 5 task datasets (public ground truth) + per-task README
├── results/<system>/*.jsonl # committed predictions per system (re-scorable)
├── src/                     # score.py (the scorer) + run/aggregation harness
├── figures/                 # leaderboard + difficulty figures
├── leaderboard.csv / .md    # the committed leaderboard (regenerates from results/)
└── DATASHEET.md             # dataset documentation (motivation, construction, limitations)

The 5 tasks (in plain language)

Two lifecycle bands — Orchestration (set the analysis up) and Execution (run the science, including the clinical-pharmacology computations). Quality control is not a separate band; it is embedded in every task — route to avoid the wrong pipeline, don't over-fill parameters, gate bad data, monitor trial toxicity, sanity-check drug exposure. BioMate leads both bands.

Band	Task (code name)	In plain English	Why it matters — and why an error is costly
Orchestration	Pipeline routing (cross_domain_routing, n=200)	Given a plain-English research request, pick the single correct analysis pipeline from a large catalog (e.g. bulk RNA-seq vs. variant-calling vs. single-cell).	The first decision in any genomics/omics project; the wrong pipeline wastes days of compute and real sample money.
Orchestration	Parameter pre-fill (param_prefill, n=170)	Fill the chosen pipeline's settings from the request — and only the settings actually specified, never invented ones. Scored by F1 (a precision/recall balance).	Over-filling silently corrupts a run; under-filling stalls it. Both can look like success until the results are quietly wrong.
Execution	QC gating (protocol_thresholds, n=100)	Decide PASS / FAIL on a real sequencing sample. One universal quality cutoff false-alarms on specialized assays (single-cell, small-RNA, cryo-EM); the agent must apply the assay-specific rule. Each item carries a public data accession (GEO / PRIDE / EMDB — the standard genomics / proteomics / cryo-EM repositories), expert-labeled.	Discarding good data — or keeping bad data — corrupts every downstream conclusion.
Execution	Trial dose-finding (boin_benchmark, n=20)	BOIN = Bayesian Optimal INterval design, a standard Phase I clinical-trial method. From emerging toxicity data, recommend the MTD (Maximum Tolerated Dose) — the highest dose that isn't too toxic.	A first-in-human patient-safety decision: too high harms patients, too low fails the drug.
Execution	Drug-exposure prediction (pbpk_benchmark, n=15)	PBPK = Physiologically-Based PharmacoKinetic modeling. From a molecule's chemistry, dose and route, predict how much reaches the blood — peak concentration (Cmax), total exposure (AUC), half-life. Scored as within 2-fold of the published human value.	Sets the first-in-human dose and underpins regulatory (FDA) submissions; a wrong exposure estimate mis-doses a trial.

Every abbreviation above (BOIN, PBPK, MTD, Cmax, AUC, GEO/PRIDE/EMDB, F1) is a standard drug-development or bioinformatics term, spelled out so the benchmark is readable without prior background. Items are original tasks authored from public sources (the nf-core pipeline catalog, public data accessions, published human pharmacokinetics, and the published BOIN design tables) — not drawn from any system's training set. Full per-task construction methodology is in the DATASHEET.md.

Coverage at a glance (how wide each task reaches)

The items are not five variations of one thing — each task deliberately spans many distinct pipelines, parameters, assays, trial designs, and drugs:

Task	Items	Distinct coverage inside the task
Pipeline routing	200	63 distinct target pipelines spanning every major omics domain (transcriptomics, genomics/variant-calling, epigenomics, single-cell, proteomics, cryo-EM, metagenomics, immunogenomics, therapeutic design)
Parameter pre-fill	170	11 inference categories (genome, strandedness, tool-flags, numeric, reference-file, multi-param, single-cell, virtual-cell, drug-discovery, name-disambiguation, negative cases) over 143 distinct parameter names
QC gating	100	99 distinct assay / library protocols; real datasets from 3 major repositories — GEO (75), PRIDE (3), EMDB (3) + 19 other public sources
Trial dose-finding	20	20 distinct Phase I trial scenarios (oncology, antibody-drug conjugates, pediatric, steep/shallow toxicity curves, exposure–response models)
Drug-exposure prediction	15	15 distinct marketed drugs (14 oral, 1 IV) across therapeutic classes (analgesic, antidiabetic, statin, antifungal, stimulant, …)

Examples & the domains each task covers

Real items from the benchmark (not toy prompts), to show the breadth:

Pipeline routing — spans essentially every omics domain:

• "Run standard RNA-seq analysis on my paired-end FASTQ files from a mouse study" → nf-core/rnaseq (transcriptomics)
• "Whole-genome bisulfite sequencing to profile CpG methylation" → nf-core/methylseq (epigenomics)
• "AIRR-compliant V(D)J gene-usage analysis for an immunology clinical study" → nf-core/airrflow (immune-repertoire / clinical immunology)
• "Recommend 2′-OMe chemical modifications to reduce siRNA immunogenicity" → sirna_offtarget_analysis (therapeutic-oligonucleotide design)
• Domains covered: transcriptomics · epigenomics · variant calling · single-cell · proteomics · cryo-EM · metagenomics · immunogenomics · therapeutic design.

Parameter pre-fill — extract the right settings, and only those:

• "RNA-seq differential expression on human lung adenocarcinoma" → {genome: GRCh38} (nothing else — don't invent a read length or aligner)
• "WES human GATK best-practices germline variant calling" → {genome: GRCh38, wes: true, tools: haplotypecaller}
• "Synergy analysis: trametinib (MEK inhibitor) + palbociclib (CDK4/6 inhibitor)" → structured drug/target params
• Domains covered: oncology genomics · gene regulation (ChIP-seq) · clinical genetics (WES/WGS) · combination pharmacology.

QC gating — real public datasets on assays where one universal cutoff misfires:

• 10x_genomics_flex_low_input (GEO GSE132044) · cite_seq_low_adt_counts (GSE164378) · smart_seq2_plate_based (GSE118184) · ambient-RNA correction (GSE163530)
• Domains covered: single-cell RNA-seq variants · CITE-seq (joint protein+RNA) · plate-based scRNA · plus proteomics (PRIDE) and cryo-EM (EMDB) library types.

Trial dose-finding — Phase I scenarios across modalities:

• "Standard oncology — myelosuppression-limited" · "Steep dose-limiting-toxicity curve — MTD at the starting dose" · "Linear exposure–response — antibody-drug conjugate" · "Pediatric oncology — weight-normalized"
• Domains covered: oncology Phase I trials · antibody-drug conjugates (ADCs) · pediatric dosing.

Drug-exposure prediction — marketed drugs across therapeutic classes, each with published human PK:

• Aspirin (analgesic, 500 mg oral) · Metformin (antidiabetic, 500 mg) · Atorvastatin (statin, 80 mg) · Ketoconazole (antifungal, 200 mg) · Caffeine (stimulant, 200 mg)
• Domains covered: clinical pharmacology across analgesics, antidiabetics, statins, antifungals — a deliberate spread of absorption/metabolism behaviors.

Figure 1. Deliberate difficulty design — routing spans direct/indirect/adversarial; others grade easy/medium/hard by inference depth.

Results

Every cell re-scores from committed predictions.jsonl via src/score.py. LLMs are evaluated the way they are actually used — model + task, no candidate list injected.

Figure 2. BioMate (product) vs frontier LLMs used directly.

Columns are grouped by capability band: Orchestration (Routing, Param) · Quality control (Protocol) · Execution (PBPK, BOIN).

System	Routing	Param (F1)	Protocol	PBPK (2×)	BOIN
BioMate (product)	0.965	0.848	0.925	1.00	0.80
Claude Opus 5	0.615	0.757	0.330	0.133	0.400
Gemini 3.1 Pro	0.580	0.693	0.630	0.000	0.278
GPT-5.6	0.530	0.671	0.460	0.200	0.500
Kimi K3	0.450	0.628	0.360	0.000	0.167
DeepSeek V4	0.415	0.621	0.420	0.067	0.167
GLM-5.2	0.400	0.641	0.440	0.000	0.111
GPT-5.6-luna	0.355	0.624	0.340	0.200	0.389
Qwen3.8-Max	0.345	0.692	0.460	0.000	0.111
Biomni (Stanford A1)	0.950	0.708	0.500	0.000	0.500

Every cell is a same-item, same-scorer result that re-scores from the committed prediction files under results/.

Efficiency & engine (measured)

Latency below is the routing task (same 200 items) — the one metric measured identically for every system, and so the only clean head-to-head. Cost is reported separately (next block), because no single task logged tokens for all systems.

System	Routing latency (median)	Engine
BioMate (product)	2.3 s	Mixture of LLMs — primary Claude Sonnet 4.5; secondary Claude Haiku 4.5, Gemini 3.5 Flash, Gemini 3.1 Pro, GPT-5.6-luna
Claude Opus 5	2.7 s	Anthropic Claude Opus 5
Gemini 3.1 Pro	3.6 s	Google Gemini 3.1 Pro
GPT-5.6	— (subset only)	OpenAI GPT-5.6
GPT-5.6-luna	1.6 s	OpenAI GPT-5.6-luna
Kimi K3	3.2 s	Moonshot Kimi K3
DeepSeek V4	3.8 s	DeepSeek V4
GLM-5.2	0.8 s	Z.ai GLM-5.2
Qwen3.8-Max	4.0 s	Alibaba Qwen3.8-Max
Biomni (A1)	3.9 s	Agent scaffold — Claude Opus 5 (routing/param/protocol), Claude Sonnet 4.5 (PBPK/BOIN; Opus-5's API rejects Biomni's assistant-prefill loop, removed across Claude ≥ 4.6)

Per-call cost — being revised

Note. The per-call token/cost analysis is temporarily withheld while we make it a clean, same-task comparison across all systems. It will be added back once corrected. Latency above is the one same-task efficiency metric we report in the interim.

Participating systems

Every system was run on the same fixed items and scored by the same src/score.py. Frontier and open-weight LLMs were run the way they are actually deployed (model + task); BioMate was run as the product; Biomni is run as its published agent scaffold.

System	Organization	Link	Reference
BioMate (product)	BioMate AI	biomate.ai · leaderboard	This benchmark — see Citation
Claude Opus 5	Anthropic	anthropic.com/claude	—
Gemini 3.1 Pro	Google DeepMind	deepmind.google/models/gemini	—
GPT-5.6 · GPT-5.6-luna	OpenAI	openai.com	—
Kimi K3	Moonshot AI	moonshot.ai	—
DeepSeek V4	DeepSeek	deepseek.com	—
GLM-5.2	Z.ai (Zhipu AI)	z.ai	—
Qwen3.8-Max	Alibaba Qwen	qwen.ai	—
Biomni (A1)	Stanford (Zou Lab)	biomni.stanford.edu	Huang et al., Science 2025, doi:10.1126/science.adz4351

Key takeaways for the community

1. BioMate leads every task on the board. Routing 0.965, parameter-F1 0.848, protocol-QC 0.925, PBPK 1.00, BOIN 0.80 — first on all five. The nearest competitor differs by task (Biomni ties routing at 0.950; Claude Opus 5 is second on param-F1 at 0.757; Gemini 3.1 Pro is second on protocol at 0.630), so no single LLM is BioMate's runner-up — the lead is broad, not a one-task artifact.
2. LLMs used directly fail execution, not knowledge. They are near-useless at PBPK simulation (mean 0.08 within-2-fold) and mediocre at deterministic BOIN dose-finding (mean 0.24) — the computation/execution tasks — while being competent at knowledge-driven extraction.
3. Output discipline is a real deployment gap. Several open-weight models emit long reasoning with no parseable answer on PBPK/BOIN — which fails deployment even when the reasoning is sound.
4. The lead is verifiable, not self-reported. Open fixed items + one shared deterministic scorer + no LLM judge means you cannot self-grade — any team runs its own system and lands directly comparable to the numbers above.
5. On curated routing, a good catalog beats raw model scale. The largest cross-system spread is routing: BioMate's product (0.965) vs the same frontier LLMs used directly (0.345–0.615).

Run your own system

# produce predictions.jsonl per task — {"id": ..., "prediction": <answer>} — then:
python3 src/score.py cross_domain_routing my_system/cross_domain_routing.jsonl
# → objective metric + 95% CI, no LLM judge. Directly comparable to the leaderboard.

Citation

@misc{realbio2026,
  title  = {RealBio: an open, objectively-scored benchmark for bioinformatics AI agents},
  author = {Zhang, Yaoyun and Dike, Andrew},
  year   = {2026},
  note   = {BioMate AI},
  url    = {https://github.com/bioMate-AI/realbio-benchmark}
}

Ground-truth sources (external references)

• BOIN — Liu & Yuan, JRSS-C 2015, doi:10.1111/rssc.12089; Yuan et al., Clin. Cancer Res. 2016.
• PBPK — FDA drug labels; DrugBank (Wishart et al., NAR 2006, doi:10.1093/nar/gkj067); Rodgers & Rowland, J. Pharm. Sci. 2007.
• Routing — nf-core (Ewels et al., Nat. Biotechnol. 2020, doi:10.1038/s41587-020-0439-x).
• Baseline agent — Biomni (Huang et al., Science 2025, doi:10.1126/science.adz4351).

License

CC-BY-4.0 — data, ground truth, scorer, and harness. See LICENSE.

---

Built by BioMate AI · leaderboard: on GitHub