MAEN / A personal public-data project

From cellular change
to the next hypothesis.

How does a substance change a cell? Can that change lead us back to what was added? We use public cancer-cell data to run small computations and compare them with measured results.

Explore the latest results Where this could lead

A personal project by Yuu Honda, run mainly with AI.
No expert supervision or peer review. Not for medical decisions.

78treatment conditions / MCF7

69.9%retrieved first across replicates

03research lines / RNA, protein & repurposing

THE QUESTION

Explore response.
Read the change.

MAEN has three research lines with different aims. All are development-stage records based on cultured cells. None evaluates treatment effects in people.

01 / RNA

Predict drug response from a cell’s features.

Use RNA measured before treatment to predict the concentration at which six existing drugs suppress cell growth. Compare predictions with measured responses in unseen cells.

See prediction error

NCI-60 / 44 development cells / 20 splits

02 / PROTEIN

Retrieve treatment conditions from protein changes.

Use protein abundance after treatment to rank 78 known conditions. Test whether a measurement from one biological replicate retrieves the original condition in the other.

See retrieval results

MCF7 / after 12 hours / 2 biological replicates

03 / REPURPOSING

Find differences in cell sensitivity among existing compounds.

Use public data on 4,532 compounds across 568 cancer cell lines to select groups with reduced viable-cell signal, then test whether relative sensitivity ranks agree in the secondary screen and CTRP. This does not determine cell death.

See the candidates

PRISM Repurposing / CTRP v2 / pre-registered

03 — REPURPOSING / EXISTING COMPOUNDS

Among existing compounds,
look for different sensitivities.

We searched existing and investigational compounds for reduced viable-cell signal relative to vehicle in a subset of cancer cell lines. We then checked whether the primary-hit group remained relatively sensitive in the secondary screen and CTRP. Criteria were fixed before viewing the data, and the evaluation ran once.

IN PICTURES

In pictures: what was tested

Every animated figure explains numbers in this section. “Schematic” figures show the idea; “Real data” figures use aggregated counts as they are.

Schematic · A three-step filter

From 4,532 to 50.

All4,532
①710
②311
③50

Remaining compounds
Remaining after step 3
Filtered out

Each dot is about 10 compounds (arrangement is schematic). ① 710 lowered the viability readout in only some lines. ② Of 638 assessable in the secondary screen, 311 had primary-hit lines relatively more sensitive. The other 72 were unassessed in that screen. ③ 50 also met the CTRP relative-rank criterion and passed step 2. Of the step-1 compounds, 105 were assessable in CTRP; 54 passed step 3 and 50 also passed step 2. Unassessed compounds are distinct from discordant results.

Schematic · Step-1 classes

Three classes by primary-hit count.

Only some lines710 compounds5 or more primary hits, at most half (selective)
More than half202 compoundsBroad activity
Fewer than 5 primary hits3,620 compoundsInactive by this rule; this does not rule out decreases that do not cross the hit threshold.

Primary hit (below 30% of control)
Other lines

Each dot is one cell line (up to 568 per compound in reality; 36 shown here). Dots that change color and shrink are primary hits: the viability readout fell below 30% of control. This readout cannot tell growth arrest from cell death.

Schematic · Separation at steps 2 and 3

Are the same lines on the sensitive side again?

Top: primary result (colored = primary hits)

← More sensitive in the secondary screen (lower AUC)Less sensitive →

Primary hit (below 30% of control)
Other lines

Separation

0.5 no differenceStep-2 threshold 0.75

0.960.47

Auxiliary control: how many pass after shuffling values

② real data311

② shuffled8

③ real data54

③ shuffled1

Shuffling once per compound is a rough control. Real data passed more often than this control, but no per-candidate p-value, true-positive rate or FDR was estimated. The 54 met the step-3 criterion; the 50 that also passed step 2 form the top tier.

We check whether primary-hit lines (colored) also rank toward the sensitive end in the secondary screen. Separation counts random hit/non-hit pairs with a lower hit AUC as 1, ties as 0.5, and a higher hit AUC as 0: 0.5 means no rank difference and 1 means complete separation. It compares relative ranks; it does not mean the same lines fell below 30% again or that cell death was reproduced.

Cultured cancer cells only · no human effect tested

A three-step filter

PRISM Repurposing 19Q4
4,532 compounds × 568 lines

① Reduced viability in some lines

710Nominal 2.5 µM primary screen: at least 5 lines, and at most half, below 30% of vehicle-control viable-cell signal

② Relative sensitivity in the secondary screen

311Out of 638 decidable in the 8-dose secondary screen

③ Relative ranks agree in CTRP

5054 of 105 decidable in CTRP met its rule; 50 also passed step 2

Primary hits have viable-cell signal below 30% of vehicle control. This viability endpoint does not distinguish growth arrest from cell death. Separation is the probability that a primary-hit line has lower AUC than a non-hit line, with ties counted as 0.5: 0.5 means no rank separation and 1 means complete separation. Step 2 required at least 0.75 with a 95% interval lower bound above 0.5; step 3 required at least 0.70 with a lower bound above 0.5. These stages test relative ranks, not whether the same lines again fall below 30% or reproduce an absolute effect or cell death.

In one auxiliary AUC shuffle per compound, 8 passed step 2 and 1 passed step 3. More passed with the observed data than with this single control; individual candidate p-values, true-positive rates and FDR were not estimated.

Positive controls: five drugs with reported selective sensitivity. Before results, they were expected to pass step 2. Thresholds were unchanged when expectations were not met. The tissue annotation “skin” means skin-derived; histology and BRAF mutation status were not matched.

Drug	Result	Detail
dabrafenib	As expected	Passed step 3. 15 of 25 primary-hit lines are skin-derived
vemurafenib	As expected	Passed step 3. 14 of 17 primary-hit lines are skin-derived
lapatinib	As expected	Passed step 3
disulfiram	Not met	Selective in step 1. Step-2 separation 0.70 (lower bound 0.64) missed the 0.75 threshold
tepoxalin	Not met	Only 5 primary-hit lines; the step-2 interval 0.40–0.95 is too wide

How to read it

Known targeted drugs met relative-sensitivity criteria, suggesting the procedure recovers known activity. Disulfiram and tepoxalin, repurposing examples in the original paper, did not meet step 2. Failing a rule does not establish absence of activity; passing does not confirm a mechanism or clinical efficacy.

HUB ANNOTATION / HISTORICAL USE CATEGORIES

Read the 47 compounds with their annotations.

These 47 compounds were annotated Launched in the historical Repurposing Hub and passed at least step 2 after excluding use annotations containing oncology or malignancy. Current country-specific human non-cancer approvals were not checked. The set includes compounds with veterinary product records, valnemulin (EMA) and tilmicosin (FDA), and can include cancer indications, as illustrated by the FDA announcement approving trifluridine/tipiracil with bevacizumab, which also cites the prior Lonsurf approval. The single-compound culture experiment is distinct from a clinical formulation.

Simvastatin and nintedanib were the two in this set that passed step 3. Four statins, three artemisinin-class compounds, corticosteroids and progestins remained together. These groups motivate hypotheses about structure, targets and sensitivity; they do not establish a common mechanism or rule out chance.

Step 3 2Step 2 · no CTRP 43Below CTRP rule 2

Anagrelide has a descriptive median IC50 of 0.01 µM (its use annotation contains malignancy, excluding it from the 47). Possible membrane effects of cetylpyridinium and unexpected activity of oxiracetam or brivaracetam warrant checks of mechanisms, identity and purity; sample problems are not an established cause. Descriptive median IC50 values for the four statins range from 0.65 to 2.67 µM. Human total/free concentrations, administration routes, formulations and exposure duration were not checked.

Real data · Positive control

Vemurafenib’s primary hits leaned toward skin-derived lines.

Skin-derived annotation (43 lines)
Other lines (521)
Primary hits (17: 14 skin-derived, 3 other)

Each dot is one of the 564 lines tested with vemurafenib (arrangement schematic, counts real). 14 of 17 primary hits carry the skin-derived annotation, against 43 of 564 lines overall. “Skin” is a tissue-of-origin annotation; melanoma diagnosis and BRAF mutation were not checked. This checks whether the procedure picks up a known pattern; it is not proof of mechanism.

Do not take a drug or change treatment because of these results. They come from cultured cancer cells. Human efficacy, safety, achievable exposure and effects on normal cells were not tested. Next, separately pre-register checks of current approvals, formulations, target species and exposure, and tests of mechanism hypotheses.

Table: 47 historical Hub compounds selected by use annotation

The 47 compounds annotated Launched in the historical Hub after excluding oncology and malignancy use labels. This is not a list of current human-approved non-cancer drugs. Primary hits/tested means below 30% of vehicle-control signal in the primary screen / lines tested. Order is by step, then step-2 separation; original rows and values are retained.

Compound	Use annotation (historical Hub)	Step	Primary hits/tested	Step 2 separation (lower)	Step 3 CTRP	Median IC50 µM
simvastatin	endocrinology, cardiology, neurology/psychiatry	Step 3	39/541	0.88 (0.82)	0.71	2.29
nintedanib	pulmonary	Step 3	15/554	0.82 (0.63)	0.86	0.78
danazol	obstetrics/gynecology, cardiology	Step 2 · no CTRP	8/564	0.99 (0.98)	—	0.65
atorvastatin	neurology/psychiatry, cardiology	Step 2 · no CTRP	46/562	0.94 (0.91)	—	1.22
artesunate	infectious disease	Step 2 · no CTRP	7/562	0.92 (0.80)	—	2.16
flumethasone-pivalate	—	Step 2 · no CTRP	8/559	0.92 (0.84)	—	3.85
eltrombopag	hematology, infectious disease	Step 2 · no CTRP	13/530	0.89 (0.84)	—	1.59
chlorhexidine	dermatology, infectious disease, dental	Step 2 · no CTRP	20/559	0.89 (0.83)	—	2.07
pyrimethamine	infectious disease	Step 2 · no CTRP	16/561	0.88 (0.84)	—	1.84
norethindrone	obstetrics/gynecology	Step 2 · no CTRP	5/562	0.87 (0.82)	—	5.99
chlorquinaldol	infectious disease	Step 2 · no CTRP	10/548	0.87 (0.78)	—	2.23
pitavastatin	endocrinology	Step 2 · no CTRP	184/562	0.86 (0.82)	—	0.65
forskolin	—	Step 2 · no CTRP	7/541	0.86 (0.75)	—	0.94
methylprednisolone	endocrinology, rheumatology, dermatology, infectious disease, allergy, hematology, neurology/psychiatry, gastroenterology	Step 2 · no CTRP	10/540	0.86 (0.68)	—	0.61
trifluridine	infectious disease	Step 2 · no CTRP	51/548	0.85 (0.79)	—	0.15
hexachlorophene	infectious disease	Step 2 · no CTRP	22/561	0.85 (0.77)	—	1.87
clocortolone-pivalate	dermatology	Step 2 · no CTRP	5/564	0.84 (0.73)	—	0.84
mycophenolic-acid	transplant	Step 2 · no CTRP	41/562	0.84 (0.74)	—	0.62
papaverine	cardiology, hematology, gastroenterology	Step 2 · no CTRP	5/562	0.83 (0.74)	—	2.25
idoxuridine	infectious disease	Step 2 · no CTRP	28/541	0.82 (0.73)	—	1.15
sulfasalazine	gastroenterology	Step 2 · no CTRP	7/541	0.82 (0.65)	—	—
nonoxynol-9	endocrinology	Step 2 · no CTRP	37/560	0.81 (0.73)	—	1.02
hexaminolevulinate	radiology	Step 2 · no CTRP	7/541	0.81 (0.58)	—	1.65
cetylpyridinium	dental	Step 2 · no CTRP	83/559	0.81 (0.77)	—	0.86
dihydroartemisinin	infectious disease	Step 2 · no CTRP	11/556	0.81 (0.66)	—	2.34
pentamidine	infectious disease	Step 2 · no CTRP	62/559	0.81 (0.75)	—	1.08
chlormidazole	infectious disease	Step 2 · no CTRP	20/541	0.80 (0.71)	—	1.83
monobenzone	dermatology	Step 2 · no CTRP	6/553	0.80 (0.69)	—	3.88
tedizolid-phosphate	infectious disease	Step 2 · no CTRP	21/541	0.80 (0.72)	—	0.90
gestrinone	obstetrics/gynecology	Step 2 · no CTRP	11/564	0.80 (0.59)	—	1.03
doxercalciferol	nephrology, endocrinology	Step 2 · no CTRP	16/530	0.79 (0.69)	—	1.72
valnemulin	infectious disease, gastroenterology	Step 2 · no CTRP	6/548	0.79 (0.60)	—	0.71
oxiracetam	—	Step 2 · no CTRP	235/556	0.78 (0.74)	—	2.09
thioproperazine	neurology/psychiatry	Step 2 · no CTRP	14/562	0.78 (0.65)	—	—
cyclovalone	gastroenterology	Step 2 · no CTRP	53/541	0.78 (0.70)	—	2.44
curcumin	—	Step 2 · no CTRP	68/530	0.78 (0.71)	—	1.79
azathioprine	transplant, rheumatology	Step 2 · no CTRP	18/562	0.77 (0.67)	—	1.83
norethindrone-acetate	endocrinology, orthopedics, obstetrics/gynecology	Step 2 · no CTRP	14/540	0.77 (0.59)	—	1.23
brivaracetam	neurology/psychiatry	Step 2 · no CTRP	91/540	0.77 (0.71)	—	0.51
lovastatin	cardiology, endocrinology	Step 2 · no CTRP	38/559	0.77 (0.65)	—	2.67
ingenol-mebutate	dermatology	Step 2 · no CTRP	56/564	0.76 (0.69)	—	0.07
artemisinin	infectious disease	Step 2 · no CTRP	5/556	0.76 (0.54)	—	1.81
tilmicosin	pulmonary	Step 2 · no CTRP	7/548	0.76 (0.59)	—	—
narasin	infectious disease	Step 2 · no CTRP	279/560	0.76 (0.71)	—	0.82
adefovir-dipivoxil	infectious disease	Step 2 · no CTRP	79/554	0.75 (0.68)	—	0.89
masitinib	dermatology	Step 2 · CTRP disagrees	8/562	0.90 (0.71)	0.72	0.46
niclosamide	infectious disease	Step 2 · CTRP disagrees	56/559	0.79 (0.73)	0.60	2.00

Table: 50 passing both steps 2 and 3 (including cancer drugs)

The 50 compounds meeting both secondary and CTRP relative-rank rules. Most are targeted drugs. Tissue counts only describe the primary-hit group's origin annotations; histology and BRAF status were not checked, and multiple comparisons were not corrected.

Compound	Phase (historical Hub)	Primary hits/tested	Step 2 separation (lower)	Step 3 CTRP	Most common tissue among primary hits	Median IC50 µM
nilotinib	Launched	5/539	0.99 (0.98)	0.97	thyroid 1/5	1.35
vemurafenib	Launched	17/564	0.99 (0.97)	0.89	skin 14/17	0.32
sorafenib	Launched	5/556	0.98 (0.97)	0.99	thyroid 1/5	0.61
lenvatinib	Launched	7/548	0.98 (0.97)	0.97	rhabdoid 3/7	0.38
pazopanib	Launched	5/548	0.94 (0.81)	0.95	rhabdoid 2/5	0.16
regorafenib	Launched	9/562	0.93 (0.88)	0.91	rhabdoid 3/9	0.54
ibrutinib	Launched	47/556	0.93 (0.91)	0.84	upper_aerodigestive 8/47	0.40
olaparib	Launched	8/557	0.93 (0.87)	0.94	bone 2/8	0.90
dabrafenib	Launched	25/541	0.92 (0.85)	0.95	skin 15/25	0.01
gefitinib	Launched	32/554	0.91 (0.87)	0.78	esophagus 4/32	0.36
trametinib	Launched	109/560	0.90 (0.87)	0.86	skin 22/109	0.03
sunitinib	Launched	13/556	0.88 (0.73)	0.85	rhabdoid 2/13	0.96
simvastatin	Launched	39/541	0.88 (0.82)	0.71	bone 6/39	2.29
cabozantinib	Launched	10/546	0.87 (0.67)	0.84	gastric 2/10	0.37
vandetanib	Launched	31/564	0.83 (0.75)	0.76	urinary_tract 5/31	1.39
erlotinib	Launched	26/562	0.83 (0.72)	0.79	urinary_tract 7/26	0.45
afatinib	Launched	32/548	0.83 (0.71)	0.83	breast 5/32	0.06
lapatinib	Launched	22/560	0.82 (0.71)	0.74	urinary_tract 5/22	1.46
nintedanib	Launched	15/554	0.82 (0.63)	0.86	rhabdoid 4/15	0.78
quizartinib	Phase 3	14/561	0.92 (0.79)	0.80	rhabdoid 2/14	0.57
PAC-1	Phase 1	8/541	0.91 (0.88)	0.81	colorectal 4/8	2.84
alpelisib	Phase 3	12/564	0.91 (0.85)	0.79	gastric 2/12	1.02
indisulam	Phase 2	57/541	0.90 (0.86)	0.76	colorectal 7/57	0.29
selumetinib	Phase 3	69/564	0.90 (0.85)	0.82	skin 23/69	0.27
R406	Phase 1	21/541	0.90 (0.79)	0.71	rhabdoid 4/21	0.33
PD-153035	Phase 1	23/548	0.89 (0.84)	0.88	esophagus 4/23	0.99
cediranib	Phase 3	15/548	0.89 (0.79)	0.98	rhabdoid 4/15	0.64
tipifarnib	Phase 3	15/559	0.89 (0.80)	0.87	liver 2/15	0.04
I-BET-762	Phase 2	13/541	0.87 (0.78)	0.82	mesothelioma 1/13	0.49
saracatinib	Phase 2/Phase 3	39/564	0.87 (0.82)	0.81	urinary_tract 6/39	0.71
MK-2206	Phase 2	17/555	0.86 (0.78)	0.81	gastric 3/17	0.38
neratinib	Phase 3	48/541	0.85 (0.79)	0.77	upper_aerodigestive 10/48	0.61
canertinib	Phase 3	16/559	0.85 (0.71)	0.76	upper_aerodigestive 3/16	0.15
alisertib	Phase 3	67/549	0.84 (0.79)	0.75	bone 7/67	0.13
GDC-0941	Phase 2	14/564	0.81 (0.64)	0.71	uterus 3/14	0.44
OSI-027	Phase 1	36/560	0.77 (0.68)	0.70	gastric 4/36	0.52
cyt387	Phase 3	7/560	0.75 (0.55)	0.71	bone 1/7	1.69
WZ-4002	Preclinical	5/556	1.00 (0.99)	0.98	lung 3/5	0.32
PD-318088	Preclinical	65/559	0.94 (0.92)	0.82	skin 20/65	0.20
I-BET151	Preclinical	18/541	0.93 (0.87)	0.73	rhabdoid 2/18	2.16
RITA	Preclinical	221/530	0.92 (0.89)	0.79	colorectal 21/221	0.23
erastin	Preclinical	50/560	0.90 (0.86)	0.75	lung 14/50	1.31
PLX-4720	Preclinical	15/556	0.89 (0.77)	0.95	skin 11/15	0.87
nutlin-3	Preclinical	19/541	0.89 (0.80)	0.79	uterus 3/19	2.00
GDC-0879	Preclinical	23/556	0.88 (0.77)	0.76	skin 13/23	0.72
AT7867	Preclinical	11/559	0.87 (0.77)	0.86	uterus 2/11	0.68
WZ8040	Preclinical	11/556	0.85 (0.66)	0.81	urinary_tract 3/11	0.78
ku-0063794	Preclinical	26/556	0.82 (0.72)	0.70	liver 4/26	0.87
GSK-3-inhibitor-IX	Preclinical	13/554	0.79 (0.62)	0.80	kidney 2/13	2.26
SB-225002	Preclinical	156/554	0.76 (0.71)	0.71	central_nervous_system 19/156	2.02

Methods, pre-registration and limitations

Discovery and confirmation used the PRISM Repurposing 19Q4 primary screen (nominal 2.5 µM, 5 days) and secondary screen (8 doses, 4-fold dilution from 10 µM), from the Broad Institute (CC BY 4.0). CTRP v2 (Broad Institute, NCI CTD² Network) supplied a different assay. Selected primary concentrations are not one exact value: 646 of 710 candidates are recorded at 2.5 µM, with a range of 1.04–5.0 µM. Ten lines that failed STR profiling were excluded. Compounds were matched by Broad ID and CTRP lines by CCLE name; ambiguous names and combinations were excluded.

The initial procedure, thresholds, code and tests were frozen before data acquisition in commit bc4ed8c. After reading provider documentation and before response-value access, v1a (78c4032) prioritized recommended secondary-screen redos (MTS010). After inventory stopped on row-label handling and before response statistics were computed, v1b (4070bc2) corrected failed-STR parsing. The record does not guarantee that v1b preceded all response-value access. Thresholds were unchanged. After input registration (93103b7) and inventory (75c860a), the evaluation ran once. The 20 tests are local runs.

CTRP shares the Broad Institute and many cell lines with PRISM; this is not validation at another laboratory or in unused cells. The descriptive 95% intervals from 500 bootstraps resample lines within hit/non-hit groups. They do not cluster related lineages or include uncertainty from shared assay conditions. A single shuffle does not provide candidate-specific p-values or FDR.

The pre-registered use flag matches only the string oncology, so hematologic-malignancy drugs remain in the JSON's launched_non_oncology_tiers. The 47 displayed above also exclude malignancy, a post-result display grouping that changes no frozen decision. Hub annotations do not establish current country-specific approvals, target species, formulations or absence of cancer indications. Compound identity and purity were not re-verified. Historical frozen JSON field names such as kill_* remain intact; their display and interpretation are corrected here.

MAEN computed compound-level aggregates from the PRISM Repurposing 19Q4 Dataset (Broad Institute, CC BY 4.0); source tables are not redistributed. The results published here are in whole or part based upon data generated by the Cancer Target Discovery and Development (CTD²) Network (https://ocg.cancer.gov/programs/ctd2/data-portal) established by the National Cancer Institute’s Office of Cancer Genomics.

Repurposing summary JSON PRISM Repurposing 19Q4 (CC BY 4.0)CTRP v2 / NCI CTD²Frozen before results: bc4ed8c / 78c4032 / 4070bc220 local tests

02 — PROTEIN

Change carries clues
about the treatment.

One replicate provides the query; the other provides the candidates. Across R1 → R2 and R2 → R1, we measure the rank of the original treatment condition for 156 queries.

Schematic · Reverse search

Reorder conditions by how similar their change is.

Q: change measured in replicate 1 (query)

Same treatment condition measured in replicate 2
Similarity

A protein up/down pattern from one replicate is the query; the 78 conditions in the other replicate are ranked by similarity. The original condition ranked first in 109/156 queries (69.9%) and in the top five in 149/156 (95.5%). Missingness positions alone reached 75.6% top five, so measurement-group cues contribute. This is not performance on unknown substances.

69.9%

Original condition ranked first

109 / 156 queries

Top 1 is the share of queries that rank the original condition first. The search is limited to measured conditions.

What information was used under the same evaluation?

Observed protein abundance69.9%

Missingness pattern only16.7%

Values permuted within masks0.6%

0%Uniform-rank reference: 1.3%100%

The limitation

Dependence from shared DMSO controls and TMT measurement groups remains. A high retrieval rate does not demonstrate prediction of unknown substances or treatment effects.

Showing the published aggregate results.

CONTROL / MISSINGNESS

Without using values,
16.7% still rank first.

Top 5: observed values 95.5%, missingness only 75.6%, permuted values 6.4%. The positions of missing values alone can identify some conditions. Measurement-group characteristics enter the search, so this control is presented alongside the main result.

EXTERNAL / ANOTHER STUDY

11 mapped signatures
were evaluated externally.

6 / 11ranked first

7 / 11in the top 5

Of 20 signatures, 19 had candidates and one had none. Eleven mapped by source names were evaluated; nine with uncertain identity or absent from the catalog were excluded from identity accuracy. This is a limited diagnostic, not an independent final validation.

01 — RNA

Does a cell’s RNA
help prediction?

The same 44 cells, six drugs and 20 splits were compared. Error is MAE in pGI50. Lower values mean predicted concentrations are closer to the experimental measurements.

0.55093Mean MAE of the RNA model

Same 20 splits, including added controls / pGI50 MAE ↓
Prediction method	Mean MAE
Drug meanAverage response for each drug	0.60794
Tissue, fixedTissue tendency / shrinkage 10	0.59465
Tissue, inner CVTissue model tuned internally	0.61016
RNA, inner CVCell-specific molecular features	0.55093
Permuted RNA controlCell-to-RNA correspondence permuted	0.60766

RNA had lower MAE than the tuned tissue model and the permuted RNA control in 19/20 splits each. These are descriptive comparisons that reuse the same cells, not independent trials, evidence of significance or a causal effect of RNA.

21 local tests passed1 CPU thread / 17.237 sAdded-control summary JSON

EXTERNAL / RNA — FIRST EXTERNAL EVALUATION

Learn from NCI.
Test in other cells.

Predictions learned from 44 NCI cells are transferred to GDSC1 and compared with sensitivity ranks within each drug. This transfer crosses different endpoints: GI50 and IC50.

Schematic · Rank correlation ρ

ρ measures how well two rankings agree.

Rank ARank B

Parallel lines: the two rankings being compared agree completely.

Like diagnostic Q2: replacing RNA with another lab’s measurement kept the frozen model’s output ranking (actual 0.94083).

Like diagnostic Q1: even the same cells measured at NCI and GDSC agreed only weakly in rank (actual 0.27041).

Like the primary result: in external GDSC1, predicted and measured ranks barely agreed (actual 0.06165).

A schematic with eight lines (the ρ shown is computed for these eight). The actual values in the notes are means of the six drug-specific correlations. ρ runs from −1 to 1; it is not a hit rate or treatment effect.

External RNA advantage unconfirmed

Primary: rank correlation strictly within screened ranges

Spearman ρ / −1 to 1
340 cells / 1,047 pairs / 6 drugs

RNA model

0.0616595% interval −0.02069 to 0.12753

Tissue, fixed

0.1031795% interval 0.02903 to 0.17933

Tissue, inner CV

0.1031795% interval 0.02903 to 0.17933

Unweighted mean of within-drug ρ (macro ρ) / 6 drugs. 95% percentile intervals from 500 paired bootstraps of annotated origin groups. Uncertainty is descriptive and conditional on the frozen models and cohort. NCI inner CV selected shrinkage 10, so fixed and CV tissue predictions are identical.

Primary analysis: strictly inside screened concentration ranges. n is the per-drug cell count; ρ compares predicted pGI50 ranks with measured pIC50 ranks.
Drug	In-range n	RNA ρ	Tissue, fixed ρ	Tissue, CV ρ
Doxorubicin	287	−0.03813	−0.05402	−0.05402
Cisplatin	115	0.16145	−0.00727	−0.00727
Paclitaxel	128	−0.18131	0.07657	0.07657
Fluorouracil	161	0.06228	0.15065	0.15065
Methotrexate	111	0.30304	0.37095	0.37095
Etoposide	245	0.06259	0.08212	0.08212

On narrow screens, the “Tissue, CV” column is hidden because it is identical to “Tissue, fixed”.

Reading the result

The RNA point estimate is lower than tissue. Across the same six drugs, the macro ρ difference is −0.04152 (95% interval −0.11405 to 0.01943). The difference interval includes zero; this external evaluation does not establish an RNA advantage.

SENSITIVITY / INCLUDING OUT-OF-RANGE FITS

The point-estimate direction changes with the range rule.

The primary cohort is 340 cells and 1,047 drug–cell pairs strictly inside screened concentrations. Among all 387 response-matched cells and 1,962 pairs, 915 out-of-range fits were excluded: 914 above the maximum and one below the minimum. This compares fitted values with screened ranges; it is not a true censoring flag.

Secondary RNA ρ 0.18051Tissue ρ 0.13146Difference +0.04906

The prespecified secondary analysis includes out-of-range fits and has a higher RNA point estimate, but no secondary bootstrap interval was computed. Because the direction changes, an external advantage depends on the range condition and remains unconfirmed. All per-drug secondary results are retained in the summary JSON.

ρ measures agreement in ranks, not prediction accuracy or treatment effects. The drug-mean model gives every cell the same prediction for a given drug, so its within-drug ρ is undefined. It is not reported as zero.

External cohort, frozen procedure and limitations

Before external response values were inspected, 418 RNA-eligible cells were selected. All 60 NCI-60 models and known same-origin or derived models were excluded. The 418 count is eligibility before response matching; it differs from the evaluated n for each drug.

Models were refitted on 44 NCI development cells. Procedures and predictions were frozen at 21677bb before external responses were retrieved. Final design-metadata mappings, extracted without response values, were additionally frozen at c961cdb. Of 500 features selected in NCI, 492 were observed in external RNA. Missing-data handling and scaling follow the frozen NCI-side procedure. The 2019 Sanger RNASeq FPKM member is transformed once to log2(FPKM+1). The FPKM interpretation relies on the legacy filename and official description; it is not assumed equivalent to current RNA processing. Eight unobserved features use NCI training medians. Unseen tissues receive zero tissue effect in the tissue models.

NCI 48-hour growth GI50 and GDSC1 72-hour viability IC50 are different endpoints. pIC50 = 6 − LN_IC50 / ln(10) aligns concentration units from µM to M; it does not turn IC50 into GI50. This evaluation measures within-drug rank correlation and is not directly compared with the original MAE. Cohort eligibility was frozen at at least 20 origin groups per drug and three drugs, but no formal pass/fail threshold for effect size was predefined. NCI-side inner CV selected tissue shrinkage 10, making the fixed and CV tissue predictions identical.

Drug mapping uses source single-agent names; salt forms and chemical structures were not independently matched exactly. Sites, culture conditions and RNA processing differ, and unannotated same-origin or derived relationships cannot be ruled out. Tests comprised 56 under Python 3.14.8 / NumPy 2.5.3 and 20 under the protein Python 3.12.14 environment; these are not a single-environment or CI result. Normal-cell toxicity was not evaluated. This evaluation cannot establish safety, toxicity or treatment effects in people.

External evaluation summary JSON GDSC1 release 8.5 / assay documentation Sanger RNA 2019-11-01 / documentation Sanger data usage policyFrozen before results: 21677bb / c961cdb76 local test executions passed / two environments

DIAGNOSTICS / RNA — DEVELOPMENT DIAGNOSTICS

Separate the clues
behind weak transfer.

Measurements, RNA from the same cells, tissue mapping and screened ranges. Starting from the previous external evaluation, separate these questions to examine disagreements.

These diagnostics use data whose results have already been seen.

The original GDSC evaluation has been inspected; this follow-up is not an independent performance confirmation. A separate bridge cohort overlapping NCI is used for measurement comparisons and is not added to the original external evaluation. These diagnostics do not establish a principal cause or a causal mechanism.

Q1Response across assays
Q2RNA in the same cells
Q3Tissue mapping
Q4Out-of-range fits

Q1Low-agreement tendency; tentative band

Does response in the same cells
agree across measurements?

Compare GI50 and IC50 within each drug for models strictly matched between NCI and GDSC. This diagnoses measurement agreement, separately from correlations of model predictions.

Candidate matched models: 58
Candidate origin groups: 55

In-range response macro ρ

0.2704195% interval 0.06319 to 0.44541

The in-range primary diagnosis covers 48 models, 158 drug–model pairs and all six drugs. Response ranks show limited agreement, but the heuristic low-agreement band remains tentative given the interval width.

Measurement correlation cannot be converted into an upper bound or mathematical ceiling for model accuracy. Cell matching, assay conditions and range rules remain limitations.

Every drug: cross-measurement comparison and status

Per-drug model counts, origin groups, response rank correlation and status. Insufficient counts and undefined results remain visible.
Drug	Models n	Origins n	Response ρ	Status
Doxorubicin	46	46	0.35753	Estimable
Cisplatin	18	18	0.18060	Estimable
Paclitaxel	15	15	0.10903	Estimable
Fluorouracil	25	25	0.52846	Estimable
Methotrexate	12	12	0.18182	Estimable
Etoposide	42	42	0.26505	Estimable

All-finite results cover 52 models and 279 pairs: macro ρ 0.48620 (interval not computed). No in-range rows meet the NCI top-dose proxy, so removing it leaves 0.27041 unchanged. The proxy is not a true censoring flag.

Q2High output-rank agreement, not accuracy

In the same cells,
do RNA and outputs agree?

Compare the same cells with RNA in both NCI and Sanger, including outputs after feeding each profile to the frozen model. No new model fitting or external retuning is performed.

Candidate RNA models: 18
Effective RNA models: 18

Frozen-output macro ρ

0.9408395% interval 0.86481 to 0.96237

Output ranks agree strongly in the same 18 models. Yet mean Sanger-input outputs are lower for all six drugs, with shifts of −0.22047 to −0.02023 pGI50. Rank agreement does not establish absence of input differences or platform equivalence. Overlap with the 44 training cells: 16 of 18 models are in training and 2 are outside training (18 origin groups). The two outside-training models are not treated as an independent accuracy cohort.

RNA or output agreement is not drug-response accuracy or a guarantee of RNA causality. Overlapping-cell results are not added to independent external performance.

Agreement metrics, counts and status

Metric definitions and effective counts are reported together. Undefined results and insufficient counts are not replaced with zero.
Comparison	Effective n	Summary	Status
RNA input agreement Per-gene, matched-cell Spearman ρ	442 genes / 18 models	Median 0.76891 / quartiles 0.62965 to 0.85637	442 evaluated / 8 missing, 50 constant excluded
Frozen output agreement Within-drug output-rank correlation	18 models	Macro ρ 0.94083 / six drugs	18 models / 108 drug–model pairs / estimable

Of 500 selected features, 493 map and 442 genes have estimable correlations (8 missing and 50 constant excluded). Same-cell identity retrieval is 17/18. This diagnoses cell matching, not drug-response accuracy. The 1/n=1/18 (5.6%) uniform-rank reference is mathematical, not a permutation-null test. Identity and per-PC correlations are also retained in the summary JSON.

Shifts are Sanger-input minus NCI-input outputs, in pGI50. All six drugs retain both rank and level differences.
Drug	n	Output ρ	Mean shift	Mean absolute difference
Doxorubicin	18	0.89267	−0.12382	0.12568
Cisplatin	18	0.96904	−0.02023	0.03875
Paclitaxel	18	0.94840	−0.18602	0.19533
Fluorouracil	18	0.93602	−0.06492	0.06854
Methotrexate	18	0.94427	−0.22047	0.22047
Etoposide	18	0.95459	−0.13662	0.14105

Q3Stratified description, cause unassigned

Do mapped tissues and
Unknown cells differ?

Describe the previous external cohort separately for Known cells mapped to NCI tissue classes and unmapped Unknown cells. Report each stratum’s counts, drug set and undefined results.

Known / mapped tissue

0.09191

RNA macro ρ / Drugs summarized 6

Unknown / unmapped tissue

0.00034

RNA macro ρ / Drugs summarized 6

Known / 147 models, 455 pairs / six drugs

RNA

0.0919195% interval −0.02427 to 0.19717

Tissue, fixed

0.1546295% interval 0.03690 to 0.26094

RNA−tissue difference

−0.0627295% interval −0.14476 to 0.02674

Unknown / 193 models, 592 pairs / six RNA drugs

RNA

0.0003495% interval −0.09606 to 0.10309

Unknown fixed/CV tissue predictions are constant: macro ρ, intervals and RNA differences are undefined. Zero tissue drugs are summarized.

Known includes 147 models and 455 pairs. The RNA–tissue difference is −0.06272; its 95% interval, −0.14476 to 0.02674, includes zero. Unknown includes 193 models and 592 pairs: six RNA drugs but zero estimable tissue drugs. The tissue comparison difference is also undefined.

Unknown receives zero tissue effect under the fixed rule. Constant predictions have undefined within-drug ρ. A stratum difference does not establish tissue mapping as the main cause.

Known / Unknown results for every drug

Known: stratified counts, rank correlations and status from the original in-range evaluation.
Drug	Models n	RNA ρ	Tissue, fixed ρ	Status
Doxorubicin	117	−0.00275	−0.07882	RNA/tissue estimable
Cisplatin	45	0.09315	0.07230	RNA/tissue estimable
Paclitaxel	58	0.13876	0.14099	RNA/tissue estimable
Fluorouracil	80	−0.03821	0.23530	RNA/tissue estimable
Methotrexate	55	0.11508	0.39239	RNA/tissue estimable
Etoposide	100	0.24541	0.16557	RNA/tissue estimable

Unknown: stratified counts, rank correlations and status from the original in-range evaluation.
Drug	Models n	RNA ρ	Tissue, fixed ρ	Status
Doxorubicin	170	−0.06121	Undefined	RNA estimable / tissue constant
Cisplatin	70	0.27509	Undefined	RNA estimable / tissue constant
Paclitaxel	70	−0.38686	Undefined	RNA estimable / tissue constant
Fluorouracil	81	0.09131	Undefined	RNA estimable / tissue constant
Methotrexate	56	0.13903	Undefined	RNA estimable / tissue constant
Etoposide	145	−0.05531	Undefined	RNA estimable / tissue constant

Q4Out-of-range description; uncertain difference

What do ranks show
among out-of-range fits?

Describe only fitted IC50 values outside screened concentration ranges. Keep this cohort separate from the previous all-finite secondary analysis and preserve per-drug counts.

Out-of-range drug–cell pairs: 915
Out-of-range models: 348

RNA

0.1839895% interval 0.10980 to 0.25785

Tissue, fixed

0.0964395% interval 0.00547 to 0.18481

RNA−tissue difference

0.0875595% interval −0.01389 to 0.19757

RNA has a higher out-of-range point estimate, but the difference, +0.08755, has a 95% interval of −0.01389 to 0.19757, including zero. This description also reflects a different range-conditioned population; it does not confirm an external advantage.

Out-of-range fits extrapolate beyond screened concentrations; they are not true censoring flags. Their correlations do not establish measured treatment effects or an external advantage.

Every drug: out-of-range-only results and status

Out-of-range fits only. Do not mix the denominator with the original in-range primary or all-finite secondary analysis.
Drug	Models n	RNA ρ	Tissue, fixed ρ	Status
Doxorubicin	46	0.46864	0.12276	Estimable
Cisplatin	258	0.16767	0.08119	Estimable
Paclitaxel	58	−0.02581	0.12804	Estimable
Fluorouracil	199	0.13683	0.10669	Estimable
Methotrexate	261	0.32693	0.15387	Estimable
Etoposide	93	0.02959	−0.01400	Estimable

What must follow the diagnostics

Treat these differences as development clues. Define the next procedures prospectively and reserve different unused data to confirm performance. Retain every drug’s results rather than selecting only favorable drugs.

Diagnostic cohorts, limitations and missing results

This follow-up comes after the original evaluation was inspected. Freezing the procedure records changes; it does not make this an unseen independent confirmation. Origin annotations may be incomplete, and GI50/IC50 endpoints, sites, RNA processing and screened ranges differ. Undefined, insufficient and unavailable comparisons keep those states; they are not replaced with zero or success.

Macro correlations and differences use matched drug sets, with descriptive 95% percentile intervals from 500 paired origin-group bootstraps. Intervals condition on the frozen models and each diagnostic cohort. All six drugs remain, and undefined tissue correlations are not replaced with zero.

v1 was not executed. v2 accessed raw bridge responses and RNA programmatically before a preserved 65.188-second soft-cap failure; no new Q1/Q2 statistics were calculated. v3 retains the numerical rules and removes duplicate I/O only, re-registered at b621429. It ran in 20.989 seconds on one CPU thread. Original external aggregates had already been viewed; this is not a new blind external test.

Follow-up diagnostic summary JSON The same primary sources as the external evaluation aboveProtocol freezing and execution record: v3 / b621429 / 20.989 s / I/O-only change

CONNECTED WORKFLOW / EXPLORATORY INTEGRATION

Connect a change
to its evidence.

A research CLI now retrieves measured treatment conditions from a requested protein-abundance change and references existing PRISM and CTRP aggregates. Similarity and response evidence remain separate; response evidence does not change the retrieval order. This does not establish the efficacy of a new treatment candidate.

What is MAEN doing?

Read a change, find similar records, and keep the unanswered questions. All bars, arrows and conditions X/Y/Z are invented examples. Actual counts are shown separately below.

Schematic

Down to advance · up to rewind · stop to readScroll controls the explanation

01 / Which change is requested?

Choose what to increase or decrease

Compare the current and requested states using the same measurement scale. Identify what should increase, decrease or stay the same.

Current amount
Requested amount

A+ More

B− Less

C0 Same

A ↑
B ↓
C →

A/B/C are invented proteins on an arbitrary scale. The requested state is not necessarily a healthy state.

02 / Has a similar change been measured?

Find similar changes

Compare protein-abundance changes with all 78 conditions measured in cultured MCF7 cancer cells 12 hours after treatment. The three examples below are invented to explain the method.

Requested change (schematic)

A ↑
B ↓
C →

Condition XA / B / C↑ ↑ →
Condition YA / B / C↑ ↓ →
Condition ZA / B / C↓ ↑ →

This does not predict the changes caused by an unknown substance.

03 / Can another record be attached?

Attach records from other experiments

Order conditions by pattern similarity first. Then attach existing drug-response records where source names correspond.

Condition Y ↑ ↓ →Provisional record: name correspondence only
Condition X ↑ ↑ →Correspondence unresolved: condition retained
Condition Z ↓ ↑ →Correspondence unresolved: condition retained

Keep conditions without references. Attaching a record does not change the similarity order.

A shared name does not establish the same chemical structure, salt or sample.

04 / What is still unknown?

Keep unanswered questions visible

Finding a similar measured pattern leaves several questions before the requested effect can be established.

Is it the same substance?Check structure, salt and sample
What happens in normal cells?Compare with cancer cells
Can that exposure be reached?Check concentration and duration
Does it hold in another study?Use independent data and experiments

Unassessed items remain unassessed. This diagram does not demonstrate treatment efficacy or restoration of a healthy state.

Scrolling controls the explanation, not experimental time.

Actual counts from the connection demo

A workflow check using a published, previously used 8-azaguanine signature (35 genes). These counts are separate from the invented bars and ranking above.

All 78 conditions: can they be scored?

74Ranked
4Unscoreable; reasons saved

The same 78: do source names correspond?

30Provisional name correspondence
48Unresolved; retained

These are two classifications of the same 78 conditions, not a filter from 74 to 30. Conditions without references and unscoreable conditions are all retained.

How to read these counts

Of the 74 ranked conditions, 29 have name correspondence and 45 are unresolved. Of the 4 unscoreable conditions, 1 has name correspondence and 3 are unresolved. Name correspondence does not establish chemical identity. Response references use different cells, doses and durations and do not establish treatment efficacy. The RNA model is not a candidate gate because its external advantage was not confirmed.

Connection demo counts and provenance (JSON)

The next checks · 15 October 2026

We checked substance references and measurement conditions before extending retrieval to another study. When comparison requirements are unresolved, we record the missing information before numerical analysis. No new independent retrieval result has been produced.

What the audit found, and what is needed next

Public lapatinib reference structures agree. The old experiment lacks source-substance identifiers, so existing name correspondences have not been upgraded to confirmed substance identity.
For BRP, biological replicates and reinjections can now be distinguished. Matrix normalization and sample-loading correction remain unconfirmed, and no contemporaneous 12-hour control was found. No numerical retrieval was run.
The separate decryptE study has 16 case-insensitive exact name overlaps with the old catalog and processed files for individual drugs. It measured Jurkat cells at 18 hours, differing from the old MCF7 catalog at 12 hours. Sample mapping and replicate checks remain incomplete.
The KSR1 study provides a normalized protein-ratio column for drug alone. Control timing, handling and biological replicate count remain unclear, so its numerical data have not been analyzed either.

Next, resolve control, sample-mapping and normalization records, then fix eligible conditions and methods before viewing outcomes. The connection-demo counts and ranks, and all existing evaluation results, remain unchanged.

Primary sources: BRP · decryptE · KSR1 study · lapatinib reference structure

01Define the input
Subtract current from target
Use protein profiles declared to share an assay, normalization and units. Compute target minus current on measured genes shared by both profiles, without zero-filling missing values. A target is not a validated healthy state.
02Keep all 78 conditions
Retrieve known changes
Rank similar or opposite changes in the existing MCF7, 12-hour catalog. Preserve source doses and unscoreable conditions with reasons. Changes caused by unknown substances are not predicted.
03Provisional label references
Attach response references
For 30 conditions, source labels match apart from letter case and provide references to existing PRISM/CTRP evidence. Chemical, salt and sample identity remain unverified. The other 48 conditions are retained as unresolved.
04Gaps remain explicit
Record what needs testing
Normal-cell effects, achievable human exposure, current approvals, formulations, target species, mechanism and independent replication remain unassessed. The RNA model has no established external advantage and is not used as a pass filter.

This is a connection demo using previously viewed data.The first run uses the first published S3 signature, 8-azaguanine (35 genes), without selecting a query for its new result. It is not a healthy-state target and has no mapped exact target condition, so retrieval accuracy is not computed. Matching of times, doses and cells across the query, catalog and response references is not established.

The source export contains 710 selective and 202 broad-primary-activity records, not an individual inventory of all 4,532 compounds. Absence from this export does not mean inactivity or lack of screening. A viable-cell proxy does not establish cell death. Public connection-demo aggregate (JSON). Searches run in a separate research CLI; this page does not accept inputs or run the analysis.

NEXT / THE DIRECTION

Before an answer,
test the question.

This evaluation did not establish an external RNA advantage. The added diagnostics do not identify a principal cause. Preserve cross-measurement disagreement and input/output level shifts as development clues. Prioritize the next performance check in different unused data with the same endpoint; if this cohort is used for future tuning, treat it as development data and use other unused data for performance evaluation. For repurposing candidates, next pre-register checks of current approvals, formulations, target species and exposure, and tests of mechanism hypotheses.

01Completed
Retrieve known conditions
Measure retrieval across replicates and compare it with missingness and value-permutation controls.
02Development diagnostics recorded
Move to unused data
Use the diagnostic clues to plan the next procedures prospectively. Prioritize same-endpoint performance confirmation; reevaluating these diagnostic cohorts is not independent confirmation.
03Search CLI implemented
Rank a desired cellular change
A research CLI ranks measured conditions with similar or opposite signatures to an input gene-change CSV. Whether they produce the desired effect remains unverified.
04Future stage
Test the effect experimentally
Have experts evaluate growth, effects on normal cells and reproducibility through experiments.

This website displays and compares aggregate results. CSV searches run in a separate research CLI. Changes caused by unknown substances, recovery toward a normal state, safety in people and treatment effects remain untested.

EVIDENCE / THE RECORD

Make the numbers
traceable.

Sources, evaluation methods, controls and previous results are preserved. This site publishes aggregates. Computation code, original matrices and cell-level raw data remain private.

Protein retrieval: evaluation and limitations

MCF7 breast-cancer cells, 12-hour treatments, 78 conditions and two replicates. Because metformin appears at five concentrations, this means 78 conditions for 74 named substances in the source table, rather than 78 different chemicals. Measurements concern protein abundance, not activity or phosphorylation.

Positive measurements in Table S2 were log2-transformed. Cosine similarity used pairs with at least 100 jointly observed accessions. Missing values were neither imputed nor equated with zero. Interpreting the values as linear relative abundance is an assumption informed by the paper’s methods; the workbook itself does not declare units. This is not a reconstruction of the paper’s gene aggregation, minimum-value imputation or ComBat-corrected matrix.

57.8% of measurement entries are missing. Missingness patterns are shared within TMT groups, as are DMSO controls. Jaccard similarity of observed accession sets (a value-free control of missingness positions) and value permutations that preserve masks serve as controls. The uniform-rank reference is a mathematical comparison, not independent probabilistic trials or a p-value. Separating replicates does not remove measurement dependence.

External Table S3 uses gene-level Log2FC signatures. Of 20 signatures, 19 had candidates and one had none. Eleven mapped using source names were evaluated; nine with uncertain identity or absent from the catalog were excluded from identity accuracy. Name matching does not prove exact chemical identity, and cell type, time and concentration differ.

RNA prediction: evaluation and limitations

Official CellMiner 2.15, HG-19, date label 09-11-2025, log2(FPKM+1). 23,808 gene records × 60 cells, no missing values. Zero is not treated as missing. Provider-side FPKM processing is used as supplied; raw data is not reprocessed.

Related cell lineages are reduced to representatives; 44 development cells are compared across the same 20 splits. Outer training has 28–30 cells, calibration 6–8 and evaluation 7–8, shared by all six drugs. Median imputation, top-500 variance selection, scaling and PCA are fitted only within each training partition. PCA 2/5 and Ridge alpha 1/10/100 are selected by inner 3-fold MAE. Added controls use tissue shrinkage 0/1/10/100 with inner CV and one fixed RNA-row permutation. The permutation also breaks tissue correspondence. Four tissue candidates and six molecular candidates do not equalize model-class complexity. Evaluation results are not used for tuning.

In the original three-model comparison, RNA had lower MAE than the drug mean in 19/20 splits and fixed tissue in 18/20. The added tuned tissue model was worse than fixed tissue in 18/20. These splits overlap and are not independent trials.

Molecular six-drug simultaneous interval coverage averaged 87.2%, but half-widths were broad: 1.548–2.533 pGI50. Coverage is descriptive; exchangeability is unestablished. “55/55” refers to input-name matching, not prediction accuracy. GI50 censoring and measurement conditions remain unresolved. Independent confirmation on the same GI50 endpoint is incomplete; GDSC1 transfer is shown in the external evaluation above, and the eight excluded held-out cells were also used previously.

Drugs: doxorubicin, cisplatin, paclitaxel, fluorouracil, methotrexate and etoposide.

Protocol freezing, environment and terms of use

Protein-search code, configuration and input hashes were frozen before scoring at commit 012e2c5. The 41 passing tests are a local execution record. Input parsing took 6.498 s and computation 1.558 s, excluding download and plotting.

RNA added controls froze code, inputs, configuration and environment before scoring at commit ce5a4fb. There were 21 passing local tests and a 17.237 s run on one CPU thread. In the original RNA evaluation, the protocol was frozen before scoring, but imported code and environment hashes were additionally pinned after results. We do not claim every dependency was frozen in advance. That original run recorded 14 local tests and 11.224 s. These records are not CI or a validation of the whole system.

Original data and full transformed matrices are not published. The protein paper is CC BY-NC-ND 4.0; a separate data license has not been confirmed. No blanket redistribution permission is assumed for this or NCI data. Each provider’s terms apply.

Primary data sources and aggregate JSON

Primary MCF7 protein studyA proteomic landscape of pharmacologic perturbations for functional relevance / Tables S1–S3
ProteomeXchange / PXD043263Mass-spectrometry data registration associated with the study
NCI Developmental Therapeutics ProgramProvider of GI50 data
CellMiner / official downloadsRNA expression / terms of use
GDSC1 release 8.5 / drug responseResponses for the external evaluation (Wellcome Sanger Institute and partners) / direct download of the source table (Excel, about 29 MB)
Cell Model Passports / RNA-seq 2019-11-01RNA expression for the external evaluation / direct download of the source archive (zip)
Sanger data usage policyTerms of use for GDSC and Cell Model Passports data
PRISM Repurposing 19Q4 DatasetUsed for repurposing discovery and confirmation (Broad Institute, CC BY 4.0)
Corsello et al. 2020, Nature CancerOriginal PRISM Repurposing paper
CTRP v2 / NCI CTD² NetworkIndependent assay (Broad Institute, CTRPv2.0_2015_ctd2_ExpandedDataset)
Repurposing summary JSONCompound level: steps, separations and positive controls for 710 compounds
Protein retrieval summary JSONAggregates by replicate direction, metric and control
Follow-up v3 diagnostic summary JSONQ1–Q4 / viewed-data development diagnostics, every drug and undefined states retained
External RNA transfer summary JSONGDSC1 / strict in-range primary and all-finite secondary evaluations
RNA added-control summary JSONFive models across the same 20 splits
Original RNA evaluation summary JSONThree models / all 20 splits

Previous records and all 20 original RNA splits

Previous results are preserved with qualifications. Related lineages crossed partitions in 18/20 of the old splits. Do not directly compare the old tissue MAE of 0.58265 with the current molecular model.

The original three-model results for the current 44-cell cohort are below. Green marks the smallest MAE in each row.

Seed	Drug mean	Tissue, fixed	RNA
20262001	0.60676	0.61848	0.59497
20262002	0.59834	0.61469	0.54597
20262003	0.59920	0.57384	0.50564
20262004	0.63684	0.59126	0.52255
20262005	0.68002	0.66629	0.66227
20262006	0.66339	0.59873	0.63236
20262007	0.58108	0.51833	0.51763
20262008	0.67575	0.64204	0.60010
20262009	0.58791	0.61086	0.60458
20262010	0.63202	0.60213	0.56232
20262011	0.54645	0.54093	0.43990
20262012	0.64457	0.61010	0.55536
20262013	0.55410	0.58195	0.52009
20262014	0.59156	0.65358	0.57023
20262015	0.62850	0.59973	0.55066
20262016	0.61919	0.63110	0.56594
20262017	0.62450	0.57631	0.60239
20262018	0.59086	0.58553	0.50056
20262019	0.59853	0.63635	0.56310
20262020	0.49926	0.44086	0.40197

Disclaimer

Personal record run with AI
No expert supervision or peer review
Not medical advice
No warranty of accuracy
No liability for damage from use
Not affiliated with data providers

Use of this site — read the full disclaimer

1. A personal record run with AI

MAEN is a personal hobby project. The code, the computations, the interpretation of results and the text on this page are produced mainly with AI (large language models). AI can produce errors that sound plausible, and the operator is not in a position to verify the content professionally. Nothing here has been supervised or peer-reviewed by experts in medicine, pharmacology or statistics.

2. Not medical advice

Nothing here may be used for diagnosis, treatment, prevention or medication decisions, and nothing here recommends or discourages any drug or therapy. The data comes from cultured cancer cells, which is different from how drugs act in the human body. For any illness or treatment, always consult a physician, pharmacist or other qualified professional. Do not start, change or stop any treatment because of this site.

3. No warranty of accuracy

Computations, data processing and descriptions may contain errors. The operator makes no warranties, express or implied, regarding accuracy, completeness, usefulness, timeliness or fitness for a particular purpose. Source data reflects the time it was retrieved and may since have been updated.

4. No liability

The operator accepts no liability for any damage arising from the use of, or inability to use, information on this site — including health or financial harm and errors in research or business decisions. Use the information at your own judgment and risk. If you cite or use it in research, papers or work, always verify against primary sources (official data from each provider and peer-reviewed literature).

5. Not affiliated with data providers

This site uses data and materials published by the U.S. National Cancer Institute (NCI), GDSC and Cell Model Passports (run by the Wellcome Sanger Institute and partners), the Broad Institute (PRISM Repurposing and CTRP), the NCI CTD² Network, the authors and journal of the MCF7 proteomics study, and ProteomeXchange. It has no affiliation with these institutions, authors or the U.S. government, and is not endorsed, recommended or approved by them. Data use is subject to each provider’s terms.

6. Subject to change without notice

Content may be corrected, updated or removed without notice. Past results are kept as records, but errors or caveats may be found later. We are not responsible for the content of linked external sites.

This English page is a translation for convenience. If it differs from the Japanese page, the Japanese version prevails.

From cellular changeto the next hypothesis.

Explore response.Read the change.

Predict drug response from a cell’s features.

Retrieve treatment conditions from protein changes.

Find differences in cell sensitivity among existing compounds.

Among existing compounds,look for different sensitivities.

In pictures: what was tested

From 4,532 to 50.

Three classes by primary-hit count.

Are the same lines on the sensitive side again?

A three-step filter

Read the 47 compounds with their annotations.

Vemurafenib’s primary hits leaned toward skin-derived lines.

Change carries cluesabout the treatment.

Reorder conditions by how similar their change is.

Without using values,16.7% still rank first.

11 mapped signatureswere evaluated externally.

Does a cell’s RNAhelp prediction?

Learn from NCI.Test in other cells.

ρ measures how well two rankings agree.

Primary: rank correlation strictly within screened ranges

The point-estimate direction changes with the range rule.

Separate the cluesbehind weak transfer.

Does response in the same cellsagree across measurements?

In the same cells,do RNA and outputs agree?

Do mapped tissues andUnknown cells differ?

Known / mapped tissue

Unknown / unmapped tissue

What do ranks showamong out-of-range fits?

What must follow the diagnostics

Connect a changeto its evidence.

What is MAEN doing?

Choose what to increase or decrease

Find similar changes

Attach records from other experiments

Keep unanswered questions visible

Actual counts from the connection demo

All 78 conditions: can they be scored?

The same 78: do source names correspond?

Subtract current from target

Retrieve known changes

Attach response references

Record what needs testing

Before an answer,test the question.

Retrieve known conditions

Move to unused data

Rank a desired cellular change

Test the effect experimentally

Make the numberstraceable.

Disclaimer

1. A personal record run with AI

2. Not medical advice

3. No warranty of accuracy

4. No liability

5. Not affiliated with data providers

6. Subject to change without notice

From cellular change
to the next hypothesis.

Explore response.
Read the change.

Among existing compounds,
look for different sensitivities.

Change carries clues
about the treatment.

Without using values,
16.7% still rank first.

11 mapped signatures
were evaluated externally.

Does a cell’s RNA
help prediction?

Learn from NCI.
Test in other cells.

Separate the clues
behind weak transfer.

Does response in the same cells
agree across measurements?

In the same cells,
do RNA and outputs agree?

Do mapped tissues and
Unknown cells differ?

What do ranks show
among out-of-range fits?

Connect a change
to its evidence.

Before an answer,
test the question.

Make the numbers
traceable.