bioreason-go-rl

A single-turn RL environment for protein Gene Ontology (GO) function prediction. The model reads protein evidence, treats GO-GPT candidate terms as noisy hypotheses, and emits one strict JSON object containing a functional summary plus GO IDs grouped by ontology aspect.

Task

Given the protein evidence, predict GO annotations as JSON:

{
  "functional_summary": "2-3 sentence description of function, role, and location",
  "molecular_function": ["GO:0005524", "GO:0004672"],
  "biological_process": ["GO:0006468"],
  "cellular_component": ["GO:0005737"]
}

Data

Train: wanglab/bioreason-pro-rl-reasoning-data
Eval: wanglab/bioreason-pro-test-data

Examples without any gold GO annotation are dropped. go_mf, go_bp, and go_cc are the gold per-aspect labels. go_pred is shown to the model as noisy candidate evidence and is never used as a target.

Reward Profiles

The environment supports only the BioReason reward profiles used by the RL config pack:

profile	weighted signals
`ted_f1_presence_v1`	`go_f1_mean` 0.35 + `go_ted_score` 0.30 + `aspect_presence_accuracy` 0.25 + `json_format_score` 0.10
`json_allaspect_f1_v1`	`go_f1_all_aspects` 0.90 + `json_format_score` 0.10
`json_go_f1_mean_v1`	`go_f1_mean` 0.90 + `json_format_score` 0.10

go_ted_score computes Zhang-Shasha tree-edit-distance similarity over the GO JSON, excluding functional_summary. GO lists are sorted for order invariance but not deduplicated, so repeated GO IDs are penalized as extra leaves.

Diagnostic metrics go_mf_f1, go_bp_f1, go_cc_f1, go_pair_f1, and empty_aspect_accuracy are logged at weight 0 for analysis.

Contract

from bioreason_go_rl import load_environment

env = load_environment(
    dataset_name="wanglab/bioreason-pro-rl-reasoning-data",
    split="train",
    num_examples=-1,
    max_sequence_chars=2000,
    max_candidate_chars=1800,
    include_protein_function=False,
    reward_profile="json_go_f1_mean_v1",
)

Unsupported reward_profile values raise ValueError with the supported profile list.

Dependencies

datasets, verifiers, and zss. No external API keys or network tools are needed beyond Hugging Face dataset download.

Published environment handle: lokahq/bioreason-go-rl@1.