OpenAI 发布 Genebench-Pro 基准测试：10个案例研究详解
阅读原文· openai.comOpenAI 推出 GeneBench-Pro 生物医学基准测试，包含 10 个案例研究。每个案例提供原始提示词、数据集和支持材料，覆盖体细胞肿瘤学（结构变异指导的肿瘤治疗获益-风险决策）、功能基因组学（CRISPR 靶点验证：lncRNA 转录本或基因组位点）和统计遗传学（连锁遗传位点中蛋白质药物靶点优先排序）等方向，要求模型输出 JSON 格式分析结果。
Inside Genebench-Pro | OpenAI
June 30, 2026
Inside Genebench-Pro
A closer look at the benchmark, its questions, and supporting materials.
Case studies
These 10 case studies showcase representative questions from GeneBench-Pro. Each case study includes the original prompt, datasets, and supporting materials. For an overview of the benchmark and key findings, see the announcement blog.
Note: File previews show excerpts from the full datasets.
Case study 1
Somatic oncology: Structural variant-guided tumor therapy benefit-risk decision
Estimate whether a synthetic TXR1-directed inhibitor has positive clinical utility in tumors whose target activation is driven by a structural variant. TXR1, TXR1i, DLR1, and star-allele labels are synthetic benchmark labels.
The target subgroup has to be recovered from long-read, expression, tumor-quality, and pharmacogenomic evidence before benefit and toxicity can be interpreted as a treatment decision.
Released prompt shown to the model
A molecular tumor board registry contains trial-eligible advanced solid-tumor cases considered for a TXR1-directed inhibitor. Estimate, for tumors with SV-driven TXR1 target-mediated activation at time zero, the marginal effect of TXR1i versus non-TXR1 systemic therapy on week-16 clinical benefit as if all patients had an assessable week-16 visit. Also estimate the 8-week treatment-limiting toxicity/discontinuation risk under TXR1i in the same target population. Report net clinical utility = benefit risk difference (percentage points) - 0.35 * toxicity risk (percentage points), and choose therapy_class_code 1 if TXR1i has positive net utility and 0 otherwise.
Use percentage-point units for all non-code quantities. Positive benefit means TXR1i improves week-16 clinical benefit relative to non-TXR1 systemic therapy.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "therapy_class_code": <int>,4 "benefit_rd_pp": <float>,5 "toxicity_dropout_risk_pp": <float>,6 "net_clinical_utility_pp": <float>7 },8 "reasoning": "<description of method and QC>"9}
Files provided to the model
patient_idanalysis_setagesexsitecalendar_periodecogtumor_burdenprior_linesprior_resistancelineage_classtherapy_classassessed16benefit16tox_stop_8wk****time_zero_day MTB0001 1 73.8 M S1 P2 2 0.787 3 1 A TXR1i 0 1 0 MTB0002 1 55.2 M S3 P1 1 2.637 0 1 A TXR1i 1 0 0 0 MTB0003 1 68.8 F S4 P2 0 0.891 2 1 A TXR1i 1 1 1 0 MTB0004 1 82.8 F S2 P2 2 4.101 0 0 B TXR1i 1 0 0 0 MTB0005 1 65.5 F S1 P3 1 7.0 1 1 A TXR1i 1 0 0 0
Registry covariates, therapy, week-16 assessment, benefit, and early toxicity.
Case study 2
Functional genomics: CRISPR target validation: lncRNA transcript or genomic locus?
Decide whether an apparent lncRNA dependency is transcript-specific or driven by nearby-locus and neighbor-gene effects.
Transcript-directed evidence has to survive controls for local DNA-locus perturbation, neighbor-gene repression, guide swaps, GC toxicity, and plate effects.
Released prompt shown to the model
You are given pooled CRISPRi screening data, guide-level local expression measurements, transcript-targeting CasRx follow-up data, and single-guide follow-up growth measurements for a nominated lncRNA program (LINC473) and a nearby coding gene (KIN1). The identifiers LINC473, KIN1, and ANKRD42 are synthetic benchmark labels; any resemblance to real human genes is coincidental.
Estimate the requested quantities.
Definitions:
- lncrna_specific_lfc: the pooled-screen matched-control day-10 log2 growth effect expected at 100% effective knockdown of the dominant LINC473 transcript, not local DNA-locus effects.
- neighbor_mediated_lfc: the pooled-screen matched-control day-10 log2 growth effect expected at 100% KIN1 repression in the local LINC473-locus model after accounting for concomitant LINC473 transcript knockdown.
- advance_target: 1 if the evidence supports advancing LINC473 as a transcript-directed target, else 0.
Conventions:
- all growth effects are log2(day10/day0) competitive-growth effects relative to matched controls;
- more negative numbers indicate stronger loss of fitness;
- set advance_target to 1 only if lncrna_specific_lfc <= -0.08 and neighbor_mediated_lfc > -0.25; otherwise 0.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "advance_target": <int>,4 "lncrna_specific_lfc": <float>,5 "neighbor_mediated_lfc": <float>6 },7 "reasoning": "<description of method and QC>"8}
Files provided to the model
guide_idnominal_targetchrcoordstranddist_lnc_tss_bpdist_neighbor_tss_bp****guide_gc_frac g001 LINC473 chr7 100014+14 30 0.624 g002 LINC473 chr7 100035-43 67 0.584 g003 LINC473 chr7 100051+116 56 0.622 g004 LINC473 chr7 100066-59 66 0.617 g005 LINC473 chr7 100088+74 77 0.715
Guide coordinates, targets, distances, and GC features.
Case study 3
Statistical genetics: Prioritizing protein drug targets in a linked genetic locus
Estimate direct disease effects for two nearby proteins using cis multivariable Mendelian randomization (cis-MVMR) while handling assay scale, allele orientation, winner's curse, LD, and residual local pleiotropy.
The two proteins share a correlated locus. The analysis has to move from marginal associations to conditional, LD-aware disease effects on a common protein scale.
Released prompt shown to the model
You are given association summary statistics and metadata for two nearby proteins (PROTA and PROTB), a binary disease outcome, a locus correlation reference, and protein measurement records.
Goal: estimate the direct log-odds effect of each protein on the disease outcome per +1 SD increase in log10 concentration, conditional on the other protein.
Interpretation: theta_PROTA and theta_PROTB use the same log-odds per-SD scale defined in the goal.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "theta_PROTA": <float>,4 "theta_PROTB": <float>5 },6 "reasoning": "<description of method and QC>"7}
Files provided to the model
snppos_bpeffect_alleleother_allelemafbetase****pval rs200000 50000000 A C 0.42215 0.006438668310706808 0.003267330091203412 0.04876727714241972 rs200001 50010126 A C 0.05709 0.011008993337581301 0.006955239208750407 0.11345916603941006 rs200002 50020253 G T 0.09021 0.009922014757116319 0.005633023027015518 0.07817048492026045 rs200003 50030379 G T 0.48399 0.010569215614164573 0.0032291419740237445 0.0010638520681901973 rs200004 50040506 A G 0.37703 0.007036551378238654 0.0033297592321269802 0.034580976884336506
Screening-stage protein association summaries for PROTA.
Case study 4
Clinical genomics / carrier screening: DRX1 carrier-screening residual risk under CNV and pseudogene calibration
Estimate ancestry-specific carrier frequencies, residual risk after a negative screen, partner carrier frequency, and affected-conceptus risk from carrier-screening assay data.
The residual-risk estimate depends on pseudogene-aware carrier calls, founder-haplotype collapse, ancestry-specific assay calibration, and standardization from tested partners back to the full partner roster.
Released prompt shown to the model
Using cohort_roster.tsv.gz, partner_roster.tsv.gz, calibration_controls.tsv.gz, target_metadata.tsv.gz, and assay_observations.tsv.gz, estimate residual reproductive risk for an autosomal recessive DRX1 condition. Report all quantities on the probability scale, not as percentages: carrier_frequency_afr and carrier_frequency_eur among screening-roster adults; residual_carrier_risk_afr_negative for an AFR screening-roster adult with a negative DRX1 screen; partner_carrier_frequency_full_roster for a uniformly sampled partner_roster.tsv.gz row; and couple_reproductive_risk for an affected conceptus when the index person is AFR and screen-negative and the partner is drawn from partner_roster.tsv.gz. Assume autosomal recessive inheritance with a 1/4 affected-conceptus risk conditional on both biological parents being carriers.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "carrier_frequency_afr": <float>,4 "carrier_frequency_eur": <float>,5 "residual_carrier_risk_afr_negative": <float>,6 "partner_carrier_frequency_full_roster": <float>,7 "couple_reproductive_risk": <float>8 },9 "reasoning": "<description of method and QC>"10}
Files provided to the model
sample_idcollectionancestry****family_history_tier S_EUR_0001 screening EUR 0 S_EUR_0002 screening EUR 0 S_EUR_0003 screening EUR 0 S_EUR_0004 screening EUR 0 S_EUR_0005 screening EUR 1
Screening-roster adults with ancestry and screening context.
Case study 5
Single-cell genomics: Activated-monocyte eQTL after ambient RNA correction
Estimate a genotype effect on activated-monocyte expression after removing ambient RNA and technical contamination from single-cell RNA-seq data.
Ambient RNA affects both target expression and the marker panel used to call activation state, so correction has to occur before the eQTL model.
Released prompt shown to the model
Estimate the per-allele log rate ratio for CXCL10 expression in the activated monocyte subpopulation from the provided single-cell RNA-seq data.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "beta_activated": <float>4 },5 "reasoning": "<description of method and QC>"6}
Files provided to the model
cell_iddonortotal_umiHBBIFI6ISG15LST1****CXCL10 D01_C001 D01 1113 7 3 4 83 5 D01_C002 D01 1103 6 3 3 112 10 D01_C003 D01 1141 9 8 12 63 9 D01_C004 D01 1250 7 60 43 2 17 D01_C005 D01 1045 9 1 2 51 15
Per-cell UMI counts for marker genes, contamination markers, and the target gene.
Case study 6
Structural genetics: Nested structural variant: expression support and clinical association
Estimate whether a nested structural subhaplotype inside an anonymous inversion-like locus has a calibrated clinical association and credible expression support.
A nested copy-dosage signal can be confounded by the broader inversion orientation, so dosage calibration, expression support, and clinical modeling have to remain distinct.
Released prompt shown to the model
Analyze the released files for anonymous Locus Q. Estimate the full-cohort source-population clinical association and molecular expression support for the calibrated nested segment-B structural copy dosage, separating the nested segment-B dosage from the broader outer-orientation dosage. Report subhap_log_or as the natural-log source-population total-effect odds ratio for case status per additional calibrated segment-B copy. Report expression_log_fc as the natural-log expression fold-change per calibrated segment-B copy for the expression-supported gene. Report target_support_code as 1 if the supported gene has a positive expression_log_fc and the clinical association is protective (subhap_log_or < 0), otherwise 0. Report n_calibrated_carriers as the number of reliable breakpoint-panel samples carrying at least one segment-B copy.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "n_calibrated_carriers": <int>,4 "target_support_code": <int>,5 "expression_log_fc": <float>,6 "subhap_log_or": <float>7 },8 "reasoning": "<description of method and QC>"9}
Files provided to the model
sample_idcaseageage_bandsexpc1pc2pc3ancestry_groupclinic_stratumrecruitment_stream Q00012 1 50.45 50_64 0-1.01514-0.21032-0.08849 EUR tertiary clinic Q00028 0 57.39 50_64 0-1.25987-0.12498 0.2344 EUR regional registry Q00029 1 68.4 65_plus 0 0.91598 0.62177 0.01891 AFR tertiary clinic Q00030 1 74.07 65_plus 1 0.21125-0.59634-0.08197 EAS community registry Q00032 1 82.82 65_plus 0-1.12034-0.24372 0.14665 EUR community clinic
Clinical and covariate data for the full cohort.
Case study 7
Regulatory genomics: Measuring chromatin loop strength after structural-variant and mapping artifact masking
Quantify a focal case-control Hi-C loop-strength difference after removing low-mappability and structural-variant artifacts from the expected-contact background.
The target loop is defined at 20 kb resolution, but the expected-contact model is distorted unless low-mappability contacts and a case-only SV stripe are masked first.
Released prompt shown to the model
You are given Hi-C contact matrices at 20 kb and 40 kb resolution plus bin annotations. Estimate the loop enrichment at the 20 kb interaction between bin_id = 8 and bin_id = 17 in bins_20kb.tsv.gz. Report three quantities: case_loop_strength (mean log2(observed/expected) across case replicates), control_loop_strength (mean log2(observed/expected) across control replicates), and delta_loop_strength (case minus control).
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "case_loop_strength": <float>,4 "control_loop_strength": <float>,5 "delta_loop_strength": <float>6 },7 "reasoning": "<description of method and QC>"8}
Files provided to the model
bin_idchromstartendgc_contentmappabilityre_sites 0 chr8 400000 420000 0.46199033821572594 0.9787574214704273 5 1 chr8 420000 440000 0.5044124208534677 0.8901084943498397 5 2 chr8 440000 460000 0.43218451584938194 0.9056879289326712 3 3 chr8 460000 480000 0.4733197282681218 0.9376529840664789 3 4 chr8 480000 500000 0.4444956062150748 0.8682565517981877 4
Target-resolution bin annotations.
Case study 8
Statistical genetics: Multi-parent QTL mapping with founder reconstruction
Map a chromosome-1 quantitative-trait locus in an eight-founder recombinant population by reconstructing founder ancestry before testing the phenotype association.
The visible marker data are biallelic, but the biological signal is founder ancestry. A defensible analysis therefore has to reconstruct founder state, check marker orientation, and separate the QTL from a batch-aligned nuisance peak.
Released prompt shown to the model
Map the chromosome 1 QTL in an 8-founder multi-parent population. Report the position (cM) and which founder carries the high-effect allele.
Report high_founder as "F1".."F8".
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2"answer": {3"high_founder": "<string>",4"qtl_pos_cM": <float>5},6"reasoning": "<description of method and QC>"7}
Data files:
- markers.tsv.gz: marker metadata
- founders.tsv.gz: founder alleles at each marker
- ril_genotypes.npz: observed RIL genotypes (biallelic)
- phenotypes.tsv.gz: phenotype and covariates
Files provided to the model
marker_idchrpos_cM m2_065 2 59.762431265596575 m2_103 2 94.52656615104739 m2_107 2 98.18761427503033 m2_079 2 72.20130244108847 m1_054 1 49.907510212292195
Marker identifiers, chromosomes, and genetic-map positions.
Case study 9
Population genetics: Parent-specific ancestry and recent admixture timing
Infer parent-specific ancestry proportions and recent admixture timing from phased local-ancestry tracts after repairing reciprocal artifacts and a chromosome-specific label inversion.
Ancestry fractions and pulse times both change if reciprocal tract artifacts, chromosome-local label inversion, or map denominators are handled incorrectly.
Released prompt shown to the model
You are given phased local-ancestry tracts for one admixed individual. Estimate, for each transmitted parental haplotype, the fraction of ancestry A across the called tract span and the number of generations since a single recent admixture pulse. Label parent1 as the haplotype with the smaller ancestry-A fraction and parent2 as the haplotype with the larger ancestry-A fraction.
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2 "answer": {3 "parent1_A_fraction": <float>,4 "parent1_t": <float>,5 "parent2_A_fraction": <float>,6 "parent2_t": <float>7 },8 "reasoning": "<description of method and QC>"9}
Files provided to the model
chromhapstart_morganend_morganancposteriorlow_complexity_frac chr1 h1 0.03 0.505 A 0.985 0.08 chr1 h1 0.505 0.535 B 0.62 0.92 chr1 h1 0.535 1.478849 A 0.985 0.08 chr1 h1 1.503727 1.852681 B 0.985 0.08 chr1 h1 1.852681 2.422373 A 0.985 0.08
Phased local-ancestry tracts with coordinates, ancestry labels, posterior values, and QC annotations.
Case study 10
Population genetics: Estimating selection from noisy ancient-DNA time series
Infer which of two haploid loci is under stronger positive selection from ancient allele-frequency time series while accounting for allele orientation, directional error, drift, and changing population size.
Noisy ancient trajectories are not directly comparable until both loci are placed on the same derived-allele scale and the provided sample-level sequencing-error values are modeled directly.
You are given allele-frequency time series data from two haploid loci sampled over multiple generations.
One locus is under stronger positive selection than the other. Estimate the selection coefficient s for the more strongly selected locus, where s > 0 means the derived allele is favored.
Assume instrument-driven sequencing error is ~1%. The seq_error column is the average of the two directional allele-miscall rates for that locus and sample.
The selected_locus value must be "A" or "B".
These data came from a real experiment; you will be graded not just on numerical correctness but the quality of analytical reasoning you exhibit; do not attempt to take any shortcuts.
Return your final answer as exactly one JSON object.
Do not wrap the JSON in markdown.
Do not add prose before or after the JSON.
Do not omit any keys shown in the example.
Return the JSON object in your final answer:
JSON
1{2"answer": {3"selected_locus": "<string>",4"s": <float>5},6"reasoning": "<description of method and QC>"7}
Files provided to the model
generationalt_readstotal_readsseq_errorsample_year 6 36 40 0.16-4500 12 34 45 0.16-4278 18 41 55 0.16-4056 24 38 70 0.16-3833 30 36 90 0.16-3611
Read-count time series for locus A.