Literature workflows
Separate potent non-oncology activity from broad high-dose killing
Read the task instructions and deliverables
PRISM measured cancer-cell viability after drug treatment. The supplied tables contain the original secondary screen's full dose series for 1,448 compounds. Which non-oncology compounds have the clearest evidence of low-concentration activity supported by their dose profiles, and how does the group's average response change with dose? Report compound names once in qualifying_compounds.
For the two numerical fields, use the source category noncancer and every matrix row. Relative viability is 2 raised to the supplied log2 fold change, relative to DMSO-treated controls. Count a row as sensitive when relative viability is strictly below 0.3. For each treatment condition (one matrix column), calculate 100 times the sensitive-row count divided by the number of rows with a measurement. Exclude NA entries and omit entirely unmeasured conditions. Keep screens and batches separate.
For mean_sensitive_percent_0625 and mean_sensitive_percent_10, report the unweighted arithmetic mean of those condition-level percentages at 0.625 and 10 micromolar (uM), respectively, across all noncancer compounds, not just your selected set. Treat doses within 0.000001 uM of the named dose as equal. Report both means on a 0–100 scale to four decimal places.
Also submit reasoning.md with 4–8 bullet points giving the key observations, the inference each supports, and how controls and possible sources of bias affect interpretation.
Return EXACTLY, as the contents of answer.json:
{"qualifying_compounds": ["<compound name>"], "mean_sensitive_percent_0625": <float, 4 decimal places>, "mean_sensitive_percent_10": <float, 4 decimal places>}How was it graded?
Deterministic checks compare the submitted answer fields with the reference values and tolerances. An LLM judge assesses the remaining reasoning criteria and cannot override the deterministic results.
One point per passed criterion, divided by the total number of criteria. Unresolved judgments produce a score range. Critical failures are reported separately and do not add weight. No overall pass threshold is specified.
LLM judge: Not recorded in this packet
Task files
Choose an evidence file.
Grading criteria
Identify non-oncology compounds that meet the killing-response criteria
Earns credit
Identify the qualifying non-oncology compounds from the supplied dose-response measurements. Submit qualifying_compounds as a JSON list of strings. After alias normalization and removal of optional members, set overlap must be at least 85%. Overlap means shared members divided by all distinct members in either set (Jaccard similarity); both omissions and extra members reduce it. Reference set (19 members): BMS-387032, brilliant-green, CHIR-124, colchicine, digitoxin, digoxin, EVP4593, halofuginone, mefexamide, nanchangmycin, oligomycin-a, ouabain, parbendazole, peruvoside, podophyllotoxin, PP-121, proscillaridin-a, sirolimus, thiram.
Does not earn credit
No credit for a missing or malformed answer, a field that is not a list of strings, or set overlap below the required threshold. Optional members do not count as errors.
Accepted alternatives
Order does not affect the set comparison. Repeated members are counted once by the scorer. Letter case is ignored. Leading and trailing whitespace is ignored. Accepted aliases: BRD-K02407574-001-09-7 → parbendazole; BRD-A93236127-001-04-5 → digitoxin; BRD-A72716251-001-01-5 → peruvoside; BRD-K60348325-065-09-9 → brilliant-green; BRD-K47869605-001-32-0 → podophyllotoxin; BRD-A81541225-001-06-8 → oligomycin-a; BRD-A73871920-236-03-6 → nanchangmycin; BRD-A11815689-001-01-1 → proscillaridin-a; BRD-K33277808-331-01-3 → ouabain; BRD-K02241333-001-02-7 → CHIR-124; BRD-K68395654-001-03-1 → EVP4593; BRD-K81801188-001-02-8 → PP-121; BRD-K20655524-003-12-0 → mefexamide; BRD-K29254801-001-14-7 → thiram; BRD-K00259736-001-16-4 → colchicine; BRD-K43389698-001-09-9 → BMS-387032; BRD-A94756469-001-04-7 → digoxin; BRD-K54247840-004-01-3 → halofuginone; BRD-K84937637-001-09-9 → sirolimus. Optional members, ignored in the comparison: albendazole, anisomycin, aphidicolin, azalomycin-b, benzethonium, brivaracetam, CGS-15943, chloropyramine, clonazepam, crystal-violet, DCEBIO, deoxycorticosterone-acetate, dichlorvos, difluprednate, digitoxigenin, disulfiram, fenbendazole, flubendazole, fluvastatin, harringtonine, ingenol-mebutate, JIB04, KD-023, lanatoside-c, LY2784544, mevastatin, mexiletine, monensin, narasin, ONX-0914, oxyphencyclimine, pardoprunox, penfluridol, pentagastrin, phenylmercuric-acetate, pirenperone, pitavastatin, puromycin, pyrithione-zinc, RN-1734, salinomycin, spiradoline, SR-33805, STA-5326, VU0238429.
Input evidence
Reference evidence
S03 · S4 paragraph 2, S15 sensitivity definition, S16 dose-response analysis, F2b caption; reference/paper-result-passages.txt and answer_key.json, _scoring.
The authors identify potent non-oncology activity and compare dose-wise killing. The packet reproduces the two group means. The set is the stable core across seven conservative analyses of sustained low-dose activity, documented in reference/key_derivation.py and reference/baseline_audit.json; it is not presented as the original 91-compound operational list.
Recorded result / GPT 6 astra
The supplied scorer reports fail. The list matches 13 of 19 core compounds, below the authored similarity requirement. It omits CHIR-124, colchicine, EVP4593, mefexamide, oligomycin-a, and sirolimus. Harringtonine, ONX-0914, and pyrithione-zinc are accepted optional members and were not penalized.
Measure mean sensitivity at 0.625 µM
Earns credit
At 0.625 µM, calculate the unweighted mean percentage of measured rows with relative viability below 0.3 across measured noncancer conditions. Use the 0–100 percentage scale. In the submitted JSON, mean_sensitive_percent_0625 must be approximately 2.9145, with an absolute tolerance of ±0.00005 (inclusive). The displayed reference is rounded for readability; the original contract retains the exact value.
Does not earn credit
No credit for a missing field, a value that is not a finite JSON number, or a result outside the allowed tolerance.
Input evidence
Reference evidence
S03 · S4 paragraph 2, S15 sensitivity definition, S16 dose-response analysis, F2b caption; reference/paper-result-passages.txt and answer_key.json, _scoring.
The authors identify potent non-oncology activity and compare dose-wise killing. The packet reproduces the two group means. The set is the stable core across seven conservative analyses of sustained low-dose activity, documented in reference/key_derivation.py and reference/baseline_audit.json; it is not presented as the original 91-compound operational list.
Recorded result / GPT 6 astra
The scorer passes 2.9145. Independent calculation from all 489 matrix rows gives 2.9145302934% across 728 measured noncancer conditions at 0.625 uM, which rounds to the submitted value.
Measure mean sensitivity at 10 µM
Earns credit
At 10 µM, calculate the unweighted mean percentage of measured rows with relative viability below 0.3 across measured noncancer conditions. Use the 0–100 percentage scale. In the submitted JSON, mean_sensitive_percent_10 must be approximately 37.2194, with an absolute tolerance of ±0.00005 (inclusive). The displayed reference is rounded for readability; the original contract retains the exact value.
Does not earn credit
No credit for a missing field, a value that is not a finite JSON number, or a result outside the allowed tolerance.
Input evidence
Reference evidence
S03 · S4 paragraph 2, S15 sensitivity definition, S16 dose-response analysis, F2b caption; reference/paper-result-passages.txt and answer_key.json, _scoring.
The authors identify potent non-oncology activity and compare dose-wise killing. The packet reproduces the two group means. The set is the stable core across seven conservative analyses of sustained low-dose activity, documented in reference/key_derivation.py and reference/baseline_audit.json; it is not presented as the original 91-compound operational list.
Recorded result / GPT 6 astra
The scorer passes 37.2194. Independent calculation from all 489 matrix rows gives 37.2193638790% across 728 measured noncancer conditions at 10 uM, which rounds to the submitted value.
Explain how low average killing can coexist with potent compounds
Earns credit
C-001 passes and the response supports a selected compound with observed, sustained low-dose killing in the correct screen, including evidence that the affected cells overlap across doses. It contrasts that example with the roughly 2.9145% group mean at 0.625 uM and explains why weak average activity does not rule out potent drugs. For example, sirolimus HTS002 has 83 of 480 eligible rows sensitive at two or more doses at or below 0.625 uM and at one or both of 2.5 and 10 uM; requiring both higher doses still leaves 57 of 480. A measured overlap or fitted dose-response result, not just separate marginal percentages, must support the inference.
Does not earn credit
C-001 fails or is missing; the example uses the wrong sample, screen, dose, group or response scale; the response gives only a single-condition hit or unrelated marginal percentages; or it omits the group-versus-compound inference.
Accepted alternatives
Any correctly selected compound with quantified same-cell dose support, or a fitted dose-response analysis showing the same low-dose potency, contrasted with the group average. This includes a supported uncertain member from an accepted variant; it need not be a stable-core example. Small rounding differences are allowed.
Prerequisites
C-001
Input evidence
Reference evidence
S03 · S4 paragraph 2, S15 final sensitivity paragraph, S16 dose-response fitting; private answer_key.json and reference/variant-support.json, V1 and V6 sirolimus records.
The paper separates group-average activity from potent individual compounds and interprets dose responses. Counts are recomputed from the release; they are not asserted to be printed in the paper.
Recorded result / GPT 6 astra
This criterion explicitly requires C-001 to pass, so the incorrect compound list blocks it. The reasoning also reports separate dose percentages without measuring overlap in sensitive cells or fitting dose-response curves. Proposed future same-cell validation does not supply that evidence.
Distinguish low-dose activity from broad high-dose killing
Earns credit
C-001 passes and the response compares a selected compound with a nonselected compound using low- and high-dose measurements from the same screen for each drug. It explains both that high-dose killing alone cannot justify selection and that sustained low-dose activity does not guarantee narrow killing at high dose. One valid comparison is sirolimus HTS002 (116/480 sensitive at 0.15625 uM, 103/466 at 0.625 uM, 468/480 at 10 uM) versus noretynodrel HTS002 (0/426 at 0.625 uM and 375/426 at 10 uM, with no sustained low-dose response).
Does not earn credit
C-001 fails or is missing; only a high-dose result is given; different screens are silently treated as a dose pair; a nonselected compound is presented as a stable-core selection; or the response fails to distinguish low-dose potency from the breadth of high-dose killing.
Accepted alternatives
Other selected and excluded compounds with verified same-screen dose profiles that establish both inferences. No claim that all selected drugs remain selective is required.
Prerequisites
C-001
Input evidence
Reference evidence
S03 · S4 paragraph 2; F2b.
The authors emphasize that breadth of killing is dose-dependent and that activity at low concentration is a separate observation.
Recorded result / GPT 6 astra
C-001 is an explicit prerequisite and fails. The reasoning also lacks a selected-versus-excluded comparison using low- and high-dose measurements that establishes both required conclusions: high-dose killing alone does not justify selection, and low-dose potency does not guarantee narrow killing at high dose.
Keep a threshold hit separate from strong biological confirmation
Earns credit
C-001 passes and the response uses a nonselected compound with uneven measurement coverage to show why an isolated threshold hit is weaker evidence than the sustained activity of the selected set. For example, aspirin HTS002 has 1/49 sensitive rows at 0.0024414 uM, but 0/416 at 0.0097656 and 0.0390625 uM, 0/426 at 0.15625 uM, and 0/182 at 0.625 uM. The response explains why that sparse hit does not establish convincing low-dose potency, and why missing values cannot be counted as resistant cells. It does not claim that failure to select aspirin proves absence of all biological activity.
Does not earn credit
C-001 fails or is missing; missing values are counted as measured cells; a sparse single-dose result is treated as sufficient proof of sustained activity; or the response gives generic missing-data advice without a compound, coverage counts, a dose comparison and the resulting strength-of-evidence distinction.
Accepted alternatives
Another excluded single-condition hit with explicit counts demonstrating uneven coverage or a reversed low-dose profile and the same cautious inference. A curve-based uncertainty analysis is acceptable if it demonstrates this distinction from the actual measurements.
Prerequisites
C-001
Input evidence
Reference evidence
S03 · S15 quality filters and sensitivity definition; S16 dose-response analysis; private reference/condition-counts.csv aspirin rows and answer_key.json.
The authors filter poor measurements and interpret drug-dose responses. The aspirin coverage example is directly measured in the release. The task distinguishes a threshold hit from sustained evidence; it does not overturn the authors’ operational hit rule or claim aspirin has no activity.
Recorded result / GPT 6 astra
C-001 is an explicit prerequisite and fails. The submission correctly warns about reversed profiles and missing measurements, but its excluded-compound examples lack coverage counts. Coverage counts elsewhere concern selected compounds, so they do not complete the required comparison between an excluded threshold hit and sustained evidence.
Put the group dose trend in context with a reference comparison
Earns credit
Puts the non-oncology group's dose trend in context using an appropriate reference comparison computed from the packet. Identifies the compared groups or distributions, doses, and response measure; reports concrete, correct observations; and connects them to what the comparison shows about the group trend. Explains both what the reference supports and a relevant limit on that inference, including its control status when used as a control. The comparison must respect compound, screen, and batch identities, measured-row denominators, and dose matching where required. No particular reference group or numerical example is required.
Does not earn credit
Gives only the non-oncology means without a reference comparison, names a reference without measured evidence and an inference, uses incorrect calculations or mismatched conditions, or does not explain what the comparison can and cannot show. Fails if it treats a drug category as an untreated negative control, treats unmatched screen differences as a dose effect, or substitutes generic control advice for a packet-based comparison.
Accepted alternatives
One accepted example compares the source drug categories: at 0.625 uM, mean sensitive percentages are about 40.07 for chemo, 15.56 for targeted cancer, and 2.91 for noncancer; at 10 uM they are about 65.61, 62.10, and 37.22. These observations place the non-oncology increase alongside broader killing in the other groups and the narrowing chemo-versus-targeted gap. Explaining that these are DMSO-normalized, batch-adjusted drug responses, not untreated negative-control groups, supplies a relevant limit. Equivalent correct distribution summaries are accepted; these categories and values are examples, not mandatory findings. An equally valid reference is the distribution of compound-level responses at the compared doses, with screens and batches handled explicitly. For example, across noncancer treatment conditions the median sensitive percentage is 0 at 0.625 uM and about 25.29 at 10 uM, with 575/728 and 94/728 conditions, respectively, having no sensitive rows. Comparing these with the means shows that the group average does not describe every compound and that the increase includes a shift in the response distribution; it does not establish a shared mechanism or activity for every compound. A correctly computed dose-matched comparison across screens is equally acceptable if it identifies the observations and sample coverage, explains whether the non-oncology trend is consistent across those screens, and limits conclusions about screen differences when compounds or measured cell rows differ. Other packet-based references that give the same kind of concrete evidence, inference, and limits are accepted.
Input evidence
Reference evidence
S03 · S4 paragraph 2; F2b; S15 control normalization and ComBat paragraphs.
The authors find chemotherapy has the broadest killing, non-oncology the least, and targeted agents become less selective at high dose.
S20 · Figure 2b: 625 nM and 10 uM panels, x-axis cell lines killed (%), y-axis proportion of drugs.
Red chemotherapy curve exceeds orange targeted curve at 625 nM; the two curves nearly overlap at 10 uM. Blue non-oncology remains below both. This is a distributional result, not an exact mean reported in the figure.
Recorded result / GPT 6 astra
Judged independently of C-001. The reasoning correctly describes the noncancer means and assay controls, but provides no measured reference comparison explaining the group's dose trend. The disulfiram comparison covers only 10 uM, lacks sample coverage, and illustrates screen variability without showing whether the group trend holds across screens. Generic control explanations do not fill this gap.