BigWonder · KRAS inhibitor resistance research notes · Evidence through 10 October 2026
KRAS inhibitors target KRAS, a protein involved in cancer. Resistance research asks how cancer cells that responded to a drug become able to withstand it, and whether that resistant state creates another vulnerability. Genes whose RNA levels differ in public cancer-cell data can provide leads. They do not, by themselves, show that inhibiting those genes would overcome resistance.
In BigWonder, I am examining that gap for NEB, NR1D1, TECTA, USP36 and ZNF512. The research has compared recovered screening tables and RNA data with published functional work and public protein-structure predictions, keeping track of what was measured and what each candidate still needs.
There are currently zero validated new therapeutic targets. Broad essentiality led me to exclude USP36 from promotion as a selective new target; the other four remain unvalidated hypotheses. This article and the Zenodo research note cover the evidence frozen through Gate21 on 10 October 2026. The research itself remains active.
A change in RNA is not the same as a dependency
RNA data show how much RNA from a gene was measured. A therapeutic-target decision needs something else: evidence that changing the gene’s function actually changes cancer-cell survival or growth. A record of equipment running differently at the scene of a fire is not evidence that the equipment caused the fire. The analogy separates observation from causation; it is not a model of cancer biology.
I first checked whether the screening tables had measured these candidates at all. In the Hallin material and the two-drug score-table comparison, none of the five had an exact match under its native gene symbol. I retained those contexts as unmeasured, rather than turning a missing row into biological counterevidence. Results from different drugs or cell backgrounds were also kept separate instead of being assembled into one apparent resistance experiment.
There were rows to examine in the RNA data. ZNF512 had a downward direction under both normalization approaches in the Dilly source and the additional PANC-1 source, while NR1D1 had opposite directions across those sources. NEB and TECTA did not meet the expression criterion used in this comparison. These observations help identify questions to follow, but none establishes dependence on a gene in resistant cells.
The expression criterion required the larger of the parental and resistant groups’ mean CPM to be at least 1 in each source. CPM, counts per million, scales RNA counts to account for differences in total measured counts. I recorded whether a gene met that criterion in both sources; falling below it was not treated as evidence that the gene was absent or had no function.
The additional PANC-1 source has three libraries labelled parental and three labelled resistant. Libraries are preparations used to read RNA; those six labels do not establish six independently evolved cancer models. Current drug exposure and the history of resistance also differ together, so the RNA contrast cannot be attributed to resistance alone. Another source includes coculture conditions for one SNU-4646S1-TO colorectal-cancer organoid background and vascular endothelial cells. There was one RNA library per assigned condition, and physical separation and cell composition were not recovered, preventing a cell-intrinsic interpretation. I did not count it as additional validation in pancreatic cancer.
Why the five candidates received different decisions
Insufficient evidence and evidence against a particular selection criterion are different reasons to withhold promotion. The candidate decisions in this snapshot are as follows.
| Candidate | Evidence examined | Current decision |
|---|---|---|
| NEB | Unmeasured under the exact native symbol in the screening tables examined. It did not meet the expression criterion in the two PANC-1 RNA sources. Prior mouse research provides muscle-function evidence. | Retain as an unvalidated hypothesis. Low expression or a missing table row does not refute a biological role. |
| NR1D1 | RNA directions differ across the two PANC-1 sources. Prior circadian and cancer-survival research means this is not a gene of unknown function with no anticancer literature. | Retain as an unvalidated hypothesis. That literature does not establish dependency in this KRAS-inhibitor resistance context. |
| TECTA | Unmeasured under the exact native symbol in the screening tables examined. It did not meet the expression criterion in the two PANC-1 RNA sources. Prior mouse research provides auditory-function evidence. | Retain as an unvalidated hypothesis. The recovered evidence supports neither promotion nor biological refutation. |
| USP36 | Listed as common essential in DepMap Public 24Q4 v1: classified as broadly needed for survival and growth across models. Dependency appears across many cancer models, and published work also reports proliferation-related functions. | Exclude from promotion as a selective new target. Broad essentiality is not evidence of a special requirement in resistant cancer cells. |
| ZNF512 | Some RNA sources agree in a downward direction, and a public structure prediction can be inspected. Existing chromatin-function research does not establish direct functional dependency under KRAS inhibition. | Retain as an unvalidated hypothesis. The Tahoe values below justify neither promotion nor biological refutation. |
The USP36 decision concerns this study’s criterion for a selective new target. It does not rule out every possible use of USP36, and it is not a measurement of toxicity in normal tissue. Normal-cell selectivity and novelty also remain unestablished for the other candidates. Prior muscle, auditory, circadian and chromatin functions were retained as reasons to examine selectivity, not as a shortcut to a safety conclusion.
What the ZNF512 structure prediction could tell me
ZNF512 already has published evidence concerning the formation of chromatin, the structure made by DNA and its associated proteins. The 2024 study also examines ZNF512B, a different gene with a similar name. Its results cannot be substituted for ZNF512, and that work did not validate a KRAS-inhibitor response.
For the public AlphaFold model, I checked the human sequence identity and where the prediction was more or less confident. The canonical model contains 567 amino acids. Its source-reported global pLDDT summary is 62.5, with 219 positions below 50. pLDDT is a confidence score for the predicted structure at each position, not the probability that the protein is important in a cell.
Some small regions had relatively confident local predictions, while their placement relative to one another remained less certain. PAE is the metric used to examine confidence in those relative positions. A readable predicted shape therefore did not establish an inhibitable drug target or a cause of resistance. This was a check of an existing public prediction’s identity and confidence limits, not a new AlphaFold run or docking calculation.
Reading two recovered Tahoe values
The next evidence source was a processed public RNA table with drug conditions attached. At a pinned Tahoe-100M version, I recovered ZNF512 rows for the human pancreatic cancer cell lines AsPC-1 and PANC-1, plate 1, RMC-6236 at 0.05 µM. Gene metadata and the identifier ENSG00000243943 were cross-checked to avoid mixing similarly named genes.
| Human cell line | Source log2FoldChange | Source padj |
|---|---|---|
| AsPC-1 · CVCL_0152 | +0.5817095041 | 0.05191597715 |
| PANC-1 · CVCL_0480 | +0.06243937090 | 0.8971019387 |
log2FoldChange expresses the RNA difference between the source producer’s comparison groups on a logarithmic scale. padj is a p-value adjusted for multiple gene comparisons, accounting for the risk of selecting chance differences as significant. Both values are at least 0.05, so I did not use this table as evidence of a statistically significant ZNF512 change. Being close to 0.05 in AsPC-1 was not a reason to change the criterion.
A more basic problem remains: the reference group for these exact rows was not established, and control membership, aggregation and the upstream replicate design were not recovered. The original paper’s 24-hour exposure and control descriptions for other analyses cannot simply be assigned to this table. I also did not reconstruct differential expression from raw RNA counts. These values are an exact readback of a public processed table, not an independently reproduced differential-expression analysis.
The AsPC-1 value is at shard 268, native row 569790; the PANC-1 value is at shard 309, native row 621380. The pinned commit is 2dc57900b7981cfcf5e211527169a0b006546a95. A separate checker confirmed the values, positions and hashes of the retained source ranges. That verifies faithful recovery of the numbers, not drug-induced causation, acquired resistance, dose dependence or therapeutic efficacy.
The question the research continues to pursue
The next requirement is public source data from matched human parental and acquired-resistant models that can resolve how changes in gene function and KRAS-inhibitor conditions affect survival and growth. That survival-and-growth measure is often called fitness. The data need to distinguish a gene that is broadly required even before treatment from one with a particular effect in the resistant state.
The original research continues to look for those measurements and their comparison conditions. This note preserves the recovered evidence and the reasons for withholding or excluding each candidate; new evidence addressing the same question can follow in a subsequent version.