AssayLLM-Handoff-GRPO

Website | Checkpoints | Paper |

Qwen3.6-27B supervised-fine-tuned and jointly GRPO-trained to open a CRISPR-screen campaign that AssayFormer finishes. Given a screen description and everything assayed so far, it returns a ranked list of 100 HGNC gene symbols for each of the first two rounds; the amortized ranker takes over from round three using those observations as context.

This is the LLM half of the Joint AssayLLM-AssayFormer GRPO result in Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens.

Usage

vllm serve Genentech/assayllm-handoff-grpo --served-model-name assayllm-handoff-grpo \
    --reasoning-parser qwen3 --port 8061 --api-key token-abc123

Within the AssayLoop harness:

Copy configs/lm/collect-qwen3.6-27b.yaml, point lm.model and lm.api_base at that server, collect the warm start, then hand off:

uv run assayloop run --model null --acq llm_single --screen-set public --full-genome \
    --lm-config configs/lm/collect-assayllm-handoff-grpo.yaml

uv run assayloop eval-ranker-handoff \
    --checkpoint <assayformer-rl-checkpoint-dir> --ckpt-file model_last.pt \
    --warm-dir output/runs --warm-prefix sweep-<id>- --n 2

--n 2 is the handoff point this model was trained for. AssayFormer is a separate checkpoint; see the AssayLoop repo.

Thinking mode is on, with the Qwen3.6 preset the paper used: temperature=1.0, top_p=0.95, top_k=20. Those are also the defaults in generation_config.json.

Without the harness:

The checkpoint accepts the same chat-formatted screen prompt and comma-separated output format as Genentech/assayllm-sft. Load it directly with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Genentech/assayllm-handoff-grpo"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, dtype="bfloat16", device_map="auto")

Scoring an output

pip install assaybench scores the pipeline's picks against a screen's ground truth. It carries the screens and the candidate pool as well as the metrics, so nothing else is needed, and no GPU:

from assaybench import enrichment_factor, gene_universe, load_screens

screens = load_screens("assayloop-test")      # the paper's 20-screen test set
universe = gene_universe(screens)             # the f2 pool, 21,147 genes

screen = screens[0]
hits = [g for g, h in zip(screen.genes, screen.hits) if h]
picked = [...]                                # all ten rounds, in order: this model's
                                              # first two, then AssayFormer's eight

enrichment_factor(picked, screen.genes, hits, universe=universe, budget=1000)

Pass universe=. It says which names outside this screen's library are still real genes: those picks are forgiven and leave the effective budget, everything else is charged as a miss. Omit it and every out-of-library name is forgiven, hallucinations included, which can only shrink the denominator, so the EF you get back sits at or above the number in the table above.

Loading

# text-only -- what you want. Builds 851 tensors, skips the vision tower.
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(REPO, dtype="bfloat16", device_map="auto")

Results

Results are evaluated on AssayBench-Loop.

Method	EF	nAUC	FH	SF
AssayLLM alone (SFT + GRPO)	3.70	0.156	0.193	0.173
AssayFormer alone (BPMF + GRPO)	4.83	0.172	0.232	0.100
AssayLLM → AssayFormer, handoff	5.05	0.186	0.247	0.077
Handoff-trained AssayLLM →AssayFormer	5.23	0.187	0.251	0.091

Citation

@article{edwards2026biologyloop,
  title={Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens},
  author={Edwards, Carl and De Brouwer, Edward and Li, Xiner and Lee, Namkyeong and
          Hajiramezanali, Ehsan and Biton, Anne and Mostafavi, Sara and Scalia, Gabriele},
  journal={arXiv preprint arXiv:2609.11877},
  url={https://arxiv.org/abs/2609.11877},
  year={2026}
}