0

Frontierbioml

Fresh

FrontierBioML: machine learning tasks in computational biology, graded against a sealed held-out split by a hosted grader.

Type
RL Env
License
unknown
Size
v0.1.7
Published
Nov 2026
Updated
Nov 2026

Cite

Notes

Only stored in your browser.

FrontierBioML by Valen: evaluating AI agents on machine learning tasks in computational biology

Leaderboard and traces · Blog post · Valen Labs

Machine learning tasks in computational biology, for coding agents. An agent reads the task statement, works in /workspace, and writes a prediction script to /app/submission. Grading re-runs that script twice on a sealed held-out split in a fresh, network-disabled sandbox and scores its output. The hidden split and the scorer never enter the agent's sandbox.

Scores are rebased so that the task's reference method scores 0 and a perfect submission scores 100. A submission that fails any validity check scores 0.

Scope of this leaderboard

This environment ships three of the 25 FrontierBioML tasks, so its leaderboard ranks models on those three only, and the order differs from the full benchmark. The full leaderboard, 11 models across all 25 tasks with a trace for every run, is at frontierbioml.com, and the blog post describes the benchmark and the results.

Grader access

Scoring sends the submission to a hosted grader and needs a grader token in FRONTIERBIOML_GRADER_TOKEN. Tokens are available on request from Valen Labs. The held-out split is never distributed.

Tasks

TaskTitleWhat the agent predicts
t01Nanobody binder triageBinding probability for every VHH-IL-6 pair, and a family-diverse candidate panel
t09Docking pose scoring for antibody complexesA score for every docking pose, so the top-scored pose in each panel is near-native
t22Partner-selective KRAS variantsKRAS folding and partner binding, and variants that keep one partner and lose another

Data sources and licences

Each task redistributes third-party data. The full notice, including the changes made, ships with the task at environment/input/licenses/DATA-SOURCE.md.

TaskDatasetSourceLicence
t01AVIDa-hIL6COGNANO — datasetCC BY-NC 4.0
t09DeepRank-AbZenodo 17911452CC BY 4.0
t22KRAS deep mutational scanWeng, Faure, Escobedo and Lehner, Nature 626, 643-652 (2024) — doiCC BY 4.0

t01 is licensed CC BY-NC 4.0: noncommercial use only. Running this task, and any use of its data or results, must be noncommercial. t22 additionally redistributes six PDB entries from the RCSB Protein Data Bank, unchanged; please cite those entries individually.

Changelog

  • 0.1.7 — Describe the benchmark as machine learning tasks in computational biology; updated header image.
  • 0.1.6 — README: FrontierBioML header image instead of the title and leaderboard graphic.
  • 0.1.5 — README: updated leaderboard graphic.
  • 0.1.4 — README: leaderboard graphic, links to the website and the blog post.
  • 0.1.3 — the hosted grader moves to its new address (the frontierbioml-grader app).
  • 0.1.2 — README: scope of the leaderboard, and how to get a grader token.
  • 0.1.1 — fix the taskset export: frontierbioml/__init__.py imported from the old package name.
  • 0.1.0 — first release: t01, t09, t22, each with its data-source notice.