task_id stringclasses 10
values | title stringclasses 10
values | description stringclasses 10
values | instruction stringclasses 10
values | task_path stringclasses 10
values | scored_outputs int64 2 2 |
|---|---|---|---|---|---|
cellsim-screen-v4-01 | dADP screen | Identify the strongest dADP-lowering knockdown in a supplied CRISPRi screen and estimate its effect. | # Nominate the knockdown that lowers DADP C10H12N5O9P2
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic DADP C10H12N5O9P2 (BiGG `M_dadp_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-targeting guides. After ten days ... | tasks/cellsim-screen-v4-01 | 2 |
cellsim-screen-v4-02 | Acyl-CoA screen | Identify the gene knockdown that lowers a mitochondrial acyl-CoA reporter and estimate its effect. | # Nominate the knockdown that lowers 2E,6Z,9Z,12Z-Octadecatetraenoyl Coenzyme A
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for mitochondrial 2E,6Z,9Z,12Z-Octadecatetraenoyl Coenzyme A (BiGG `M_CE2437_m`). The library targets 400 metabolic genes with 4 gu... | tasks/cellsim-screen-v4-02 | 2 |
cellsim-screen-v4-03 | C14:0 PE screen | Identify the gene knockdown that lowers the cytosolic C14:0 PE reporter and estimate its effect. | # Nominate the knockdown that lowers 1-Myristoylglycerophosphoethanolamine (C14:0 Pe)
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic 1-Myristoylglycerophosphoethanolamine (C14:0 Pe) (BiGG `M_pe14_hs_c`). The library targets 400 metabolic genes ... | tasks/cellsim-screen-v4-03 | 2 |
cellsim-screen-v4-04 | Aminopropionaldehyde screen | Identify the gene knockdown that lowers cytosolic beta-aminopropionaldehyde and estimate its effect. | # Nominate the knockdown that lowers Beta-Aminopropion aldehyde
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic Beta-Aminopropion aldehyde (BiGG `M_bamppald_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-targeting g... | tasks/cellsim-screen-v4-04 | 2 |
cellsim-screen-v4-05 | Heptadecanoate screen | Identify the gene knockdown that lowers cytosolic heptadecanoate and estimate its effect. | # Nominate the knockdown that lowers Heptadecanoate C170 C17H33O2
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic Heptadecanoate C170 C17H33O2 (BiGG `M_hpdca_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-target... | tasks/cellsim-screen-v4-05 | 2 |
cellsim-screen-v4-06 | Leukotriene B4 screen | Identify the strongest leukotriene B4-lowering knockdown in a supplied CRISPRi screen and estimate its effect. | # Nominate the knockdown that lowers Leukotriene B4(1-)
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic Leukotriene B4(1-) (BiGG `M_leuktrB4_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-targeting guides. After ten... | tasks/cellsim-screen-v4-06 | 2 |
cellsim-screen-v4-07 | Protein methyl-lysine screen | Identify the strongest knockdown lowering a cytosolic protein methyl-lysine reporter and estimate its effect. | # Nominate the knockdown that lowers Aggregate protein N6-methyl-lysine residue, cytosolic
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic Aggregate protein N6-methyl-lysine residue, cytosolic (BiGG `M_M00213_c`). The library targets 400 metabol... | tasks/cellsim-screen-v4-07 | 2 |
cellsim-screen-v4-08 | GT1b screen | Identify the strongest GT1b-lowering knockdown in a supplied CRISPRi screen and estimate its effect. | # Nominate the knockdown that lowers GT1b (homo sapiens)
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic GT1b (homo sapiens) (BiGG `M_gt1b_hs_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-targeting guides. After te... | tasks/cellsim-screen-v4-08 | 2 |
cellsim-screen-v4-09 | L-Homoserine screen | Identify the strongest L-homoserine-lowering knockdown in a supplied CRISPRi screen and estimate its effect. | # Nominate the knockdown that lowers L-Homoserine
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic L-Homoserine (BiGG `M_hom__L_c`). The library targets 400 metabolic genes with 4 guides each plus 100 non-targeting guides. After ten days of knock... | tasks/cellsim-screen-v4-09 | 2 |
cellsim-screen-v4-10 | Oleoyl GPE screen | Identify the strongest knockdown lowering the oleoylglycerophosphoethanolamine reporter and estimate its effect. | # Nominate the knockdown that lowers 1-Oleoylglycerophosphoethanolamine (Delta 9)
A simulated pooled CRISPR interference screen was run in a human cell line carrying a fluorescent biosensor for cytosolic 1-Oleoylglycerophosphoethanolamine (Delta 9) (BiGG `M_peole_hs_c`). The library targets 400 metabolic genes with 4 ... | tasks/cellsim-screen-v4-10 | 2 |
CellSim CRISPRi Screens
Ten synthetic functional-genomics tasks in the GeneBench-Pro style. Each task asks an agent to analyze guide-level screen counts, identify the strongest reporter-lowering gene knockdown, and estimate its enrichment.
The task catalog, model comparison, and project overview are available at symbolic.life/tasks.
Contents
tasks/: runnable Harbor tasks, each with an instruction, container environment, verifier, and reference solution.task_catalog.jsonl: browsable task descriptions and prompts.download/: the exact task archive served on symbolic.life.benchmark-results.json: completed six-model benchmark results.
Run
Install Docker and uv, then run from this repository’s directory:
uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent oracle
To evaluate an agent, configure its provider credentials in your own environment and run:
uv run --with harbor==0.23.0 python run_verified.py run --path ./tasks --agent terminus-2 --model PROVIDER/MODEL --ak reasoning_effort=high
The supplied launcher uses the tested Harbor version and mounts each task’s verifier separately and read-only. Keep tests/ and solution/ outside the evaluated agent’s environment.
Scoring
The answer is a gene symbol and gene-level log2 enrichment. The task prompt defines the aggregation statistic. Whole-task success requires the correct gene and an effect estimate within 0.15 log2 units of the reference. Partial credit is the mean of gene and effect correctness.
Results
Each of the six models completed three attempts on each of the ten tasks, with high reasoning effort and a 30-minute task budget. All 180 recorded scores were independently regraded against the released task files.
| Model | Correct attempts | Whole-task success |
|---|---|---|
| Claude Opus 5 | 30/30 | 100.0% |
| GPT-6 Astra | 28/30 | 93.3% |
| Gemini 3.8 Flash | 11/30 | 36.7% |
| Grok 4.6 | 8/30 | 26.7% |
| Grok 4.7 | 6/30 | 20.0% |
| Nemotron 3 Ultra | 0/30 | 0.0% |
Data source
Counts are synthetic and use CellSim’s metabolic reaction structure. These tasks evaluate recovery of a statistic from the supplied data; the counts are not measured laboratory observations or numerically integrated cellular trajectories.
License
The task data, prompts, and included code are released under the Apache License 2.0.
- Downloads last month
- 110