Open cross-system benchmark · drug development & genomics

RealBio: can an AI agent actually run a biomedical pipeline — not just discuss it?

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 trial dose. Teams increasingly want an AI agent to drive them. RealBio asks the blunt question: can an agent get the objectively-checkable answer right? 505 fixed public items across 5 real drug-development / genomics tasks, each with published ground truth, graded by one shared open scorer (score.py — deterministic, no AI judge). The finding: frontier LLMs know biology but fail at execution — worst on the clinical-pharmacology tasks (predicting drug exposure, choosing a trial dose) — while a purpose-built platform (BioMate) leads every task.

Data + scorer on GitHub → Dataset on HuggingFace See the leaderboard

What makes it different

Anti-self-grading, by construction

505
fixed items
5 objective-ground-truth tasks
1
shared open scorer
score.py · no LLM judge
10
systems on the board
8 frontier LLMs + a domain agent + BioMate
CC-BY-4.0
fully open
data, ground truth, scorer, harness

Tasks are real bioinformatics automation — not multiple-choice trivia. Every number traces to committed score.py output on committed predictions; every leaderboard cell re-scores from the released files.

The five tasks, in plain language

The real drug-development relay

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 it — grouped into Orchestration (set the analysis up) and Execution (run the science, including the clinical-pharmacology computations). Quality control isn't a separate stage — it's embedded in every task.

BandTaskIn plain EnglishWhy an error is costlyCoverage
OrchestrationPipeline routingPick the single correct analysis pipeline for a plain-English request (RNA-seq vs. variant-calling vs. single-cell …).The wrong pipeline wastes days of compute and real sample money.200 items · 63 pipelines, all omics domains
OrchestrationParameter pre-fillFill the pipeline's settings from the request — and only the ones actually specified, never invented.Over-filling silently corrupts a run; both failures look like success.170 items · 11 categories, 143 parameters
ExecutionQC gatingDecide PASS/FAIL on a real sequencing sample; a universal cutoff false-alarms on specialized assays (single-cell, small-RNA, cryo-EM).Discarding good data — or keeping bad data — corrupts every downstream result.100 items · 99 assays · GEO/PRIDE/EMDB
ExecutionTrial dose-finding (BOIN)BOIN = Bayesian Optimal INterval, a Phase I clinical-trial design. From toxicity data, recommend the Maximum Tolerated Dose (MTD).A first-in-human patient-safety call: too high harms patients, too low fails the drug.20 items · 20 trial scenarios (oncology, ADCs, pediatric)
ExecutionDrug-exposure prediction (PBPK)PBPK = Physiologically-Based PharmacoKinetics: predict blood concentration (Cmax, AUC, half-life) from a molecule's chemistry, dose and route.Sets the first-in-human dose and underpins FDA submissions.15 items · 15 marketed drugs across classes

Every abbreviation (BOIN, PBPK, MTD, Cmax, AUC, GEO/PRIDE/EMDB) is a standard drug-development or bioinformatics term. Items are original tasks authored from public sources — not drawn from any system's training set. The pattern: LLMs used directly do fine on knowledge-heavy orchestration but collapse on the execution tasks — PBPK simulation (mean 0.08 within-2-fold) and BOIN dose-finding (mean 0.24 exact-MTD) — exactly where a wrong answer is a safety error.

Cross-system leaderboard

BioMate leads every task

On the same fixed items and the same open scorer, BioMate is first on all five tasks — routing 0.965, parameter-F1 0.848, protocol-QC 0.925, PBPK 1.00, BOIN 0.80. The runner-up differs by task (Biomni ties routing; Claude Opus 5 is second on parameters; Gemini 3.1 Pro second on protocol), so the lead is broad, not a one-task artifact. The clearest pattern among the LLMs: used the way they are actually deployed, they fail execution tasks — PBPK simulation (mean 0.08 within-2-fold) and deterministic BOIN dose-finding (mean 0.24 exact-MTD) — while doing well on knowledge-driven extraction.

SystemRoutingParam (F1)Protocol QCPBPK (2×)BOIN (MTD)
BioMate (product)0.9650.8480.9251.000.80
Claude Opus 50.6150.7570.3300.1330.444
Gemini 3.1 Pro0.5800.6930.6300.0000.278
GPT-5.60.5300.6710.4600.2000.500
Kimi K30.4500.6280.3600.0000.167
DeepSeek V40.4150.6210.4200.0670.167
GLM-5.20.4000.6410.4400.0000.111
Qwen3.8-Max0.3450.6920.4600.0000.111
Biomni (Stanford A1)0.9500.7080.5000.0000.500

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, so the only clean head-to-head. Cost is reported separately, because no single task logged tokens for all systems.

SystemRouting latency (median)Engine
BioMate (product)2.3 sMixture 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 52.7 sAnthropic Claude Opus 5
Gemini 3.1 Pro3.6 sGoogle Gemini 3.1 Pro
GPT-5.6— (subset only)OpenAI GPT-5.6
GPT-5.6-luna1.6 sOpenAI GPT-5.6-luna
Kimi K33.2 sMoonshot Kimi K3
DeepSeek V43.8 sDeepSeek V4
GLM-5.20.8 sZ.ai GLM-5.2
Qwen3.8-Max4.0 sAlibaba Qwen3.8-Max
Biomni (A1)3.9 sAgent scaffold — Claude Opus 5 (routing/param/protocol), Claude Sonnet 4.5 (PBPK/BOIN)

Per-call cost — being revised. The 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 interim same-task efficiency metric.

Who is on the board

Participating systems

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

SystemOrganizationLink
BioMate (product)BioMate AIbiomate.ai
Claude Opus 5Anthropicanthropic.com/claude
Gemini 3.1 ProGoogle DeepMinddeepmind.google/models/gemini
GPT-5.6 · GPT-5.6-lunaOpenAIopenai.com
Kimi K3Moonshot AImoonshot.ai
DeepSeek V4DeepSeekdeepseek.com
GLM-5.2Z.ai (Zhipu AI)z.ai
Qwen3.8-MaxAlibaba Qwenqwen.ai
Biomni (A1)Stanford (Zou Lab)biomni.stanford.edu · Science 2025

Run your own system

The items, ground truth, and scorer are public. Run your system on the fixed items, score with the shared score.py, and your result is directly comparable to the leaderboard.

GitHub: bioMate-AI/realbio-benchmark → HuggingFace dataset