Study
Enformer Recognizes cis-Regulated Genes by Prediction Magnitude but Fails to Predict Effect Direction: A Magnitude–Direction Dissociation in Individual-Level Expression Prediction
View full titleShow less
We cross-reference Enformer's individual-level per-gene Pearson R (Sasse et al.
Abstract
We cross-reference Enformer's individual-level per-gene Pearson R (Sasse et al. 2023, Nature Genetics, ROSMAP cohort n=839) with GTEx v8 Brain Cortex eGenes (n=9,082) across 6,808 genes, and report a fundamental dissociation: prediction magnitude (|R|) monotonically predicts cis-eQTL gene identity (eGene rate 31.6% → 82.5%; |R| model AIC=8274 vs R+R² AIC=8378, ΔAIC=104), but prediction direction for eGenes is near-random (54.0% correct; max D10=62.5%). Direction errors are irreversible by fine-tuning on ROSMAP (Mann–Whitney p≈10⁻¹⁰⁰), yet PrediXcan recovers correct direction for R<0 eGenes (PrediXcanR=0.295), consistent with the hypothesis that Enformer's non-brain ENCODE training corpus lacks the brain-specific regulatory programs needed to correctly orient directional effects.
This page presents preserved, source-backed released materials.
Source inventory
Deterministically extracted immutable study materials.
Extracted study record
Authoritative metadata and released result artifacts.
The preserved source contains 5 result tables; 5 have complete deterministic representation coverage.
Regulatory driver landscape
1,091 released driver loci · signed distance from transcription start site, not magnitude.
Evidence visualization
Regulatory driver landscape
1,091 released driver loci · signed distance from transcription start site, not magnitude.
TSS-centered regulatory evidence
Driver loci around transcription start sites
Released driver locus · ◎ Primary driver · ● Selected locus · point area encodes carrier prevalence
Selected released entity
NANOS1
| Distance | Position | Score | Carriers | Prevalence | Primary |
|---|---|---|---|---|---|
| −75.37 kb | 118,954,350 | −0 | 376 | 32.4% | No |
| −70.4 kb | 118,959,316 | −0.03 | 379 | 32.6% | No |
| −52.99 kb | 118,976,731 | +0.02 | 484 | 41.7% | No |
| −50.94 kb | 118,978,778 | −0 | 398 | 34.3% | No |
| −50.53 kb | 118,979,186 | −0 | 962 | 82.9% | No |
| −50.43 kb | 118,979,285 | −0.01 | 960 | 82.7% | No |
| −50.12 kb | 118,979,592 | +0 | 802 | 69.1% | No |
| −49.78 kb | 118,979,937 | +0 | 810 | 69.8% | No |
| −47.66 kb | 118,982,059 | −0.01 | 95 | 8.2% | No |
| −46.72 kb | 118,982,999 | +0 | 1,122 | 96.6% | No |
| −46.12 kb | 118,983,597 | +0.01 | 554 | 47.7% | No |
| −45.73 kb | 118,983,984 | +0 | 554 | 47.7% | No |
Paired released correlations
Driver versus aggregate correlation
Signed distance is an explicit released coordinate relative to the transcription start site. Driver score, prevalence, flags, and correlations are copied from the immutable table; no effect size or biological conclusion is inferred.
Source & verification
- File
- Data/SupplementaryTable2.tsv
- Version
afac6658- Integrity
1618b841…3fa458
Ranked signed feature weights
5,314 released model terms · positive and negative coefficients shown around zero.
Evidence visualization
Ranked signed feature weights
5,314 released model terms · positive and negative coefficients shown around zero.
Strongest released feature contributions
Negative ← zero → Positive
CHIP:H3K36me3:layer of hippocampus female adult (75 years)
- Weight
- −0.21295
- Direction
- Negative
- Group
- CHIP
- Feature type
- H3K36me3
- Context
- layer of hippocampus female adult (75 years)
- Source row
- 2,336
CHIP:H3K36me3:layer of hippocampus female adult (75 years)
Source & verification
- File
- fine_tuned_model/GTExWeights.csv
- Version
afac6658- Integrity
68e68810…ebd21c
Gene-level model concordance
6,820 entities · observed versus model-predicted released values.
Evidence visualization
Gene-level model concordance
6,820 entities · observed versus model-predicted released values.
Observed versus predicted
Fine Tuned parity plot
Released entity · equality (y = x)
Selected released entity
MALAT1
| Series | Mean | SD | Correlation | p-value |
|---|---|---|---|---|
| Observed | 12.59 | 0.33 | Not released | Not released |
| Fine Tuned | 9.04 | 0.015 | -0.03 | 0.391 |
| CAGE Adult Brain | 1.68 | 0.006 | -0.025 | 0.477 |
Per-entity correlation
Fine Tuned versus CAGE Adult Brain
Position and identifiers remain categorical context. Means, standard deviations, correlations, and p-values are copied from the immutable released table; no effect size is inferred.
Source & verification
- File
- Data/SupplementaryTable1.tsv
- Version
afac6658- Integrity
86a0dbe9…cd4009
Gene Ids And Positions
Genomic loci from Data/gene-ids-and-positions.tsv, placed on the released reference assembly.
Evidence visualization
Genomic loci in the released assessment set
58,302 loci mapped to hg38 · chromosome is a category, not a magnitude.
22 additional released contigs are available in the selector.
| Gene / locus | Start | End | Strand |
|---|---|---|---|
| DDX11L1 | 11,869 | 14,409 | + |
| AL591163.1 | 6,767,954 | 6,770,038 | + |
| CD24P1 | 15,614,643 | 15,614,867 | - |
| LYPLA2 | 23,790,970 | 23,795,539 | + |
| AL451070.1 | 31,333,067 | 31,346,799 | - |
| MYCL | 39,895,426 | 39,902,256 | - |
| AL135960.1 | 47,225,797 | 47,230,750 | + |
| AC093424.1 | 59,289,303 | 59,289,640 | - |
| MIGA1 | 77,779,624 | 77,879,539 | + |
| ARHGAP29 | 94,148,988 | 94,275,068 | - |
The interactive view uses 584 deterministic representatives from 58,302 source rows. The complete immutable source remains linked.
Source & verification
- File
- Data/gene-ids-and-positions.tsv
- Version
afac6658- Integrity
1c1bf881…a9a24f
Deterministic extraction method
No generative model selected, summarized, or transformed the results.
Evidence classified preserved repository files, parsed bounded structured tables, inferred column types, and selected registered components using versioned rules.
Data and provenance
Immutable publication and repository sources.
Reuse boundary
Only released and preserved source rows are displayed.
No model is rerun and no missing value is inferred.
Pinned scientific commit: afac6658d57bfe6e2c1f17d9a96e3593843000a2
Limitations
Deterministic coverage boundaries.
Every accepted released table cell is represented by registered Evidence components; unstructured claims are not converted into numerical results.
Citation
Cite the scholarly record and pinned analysis source.
Scholarly publication
DongKoo Lee. Enformer Recognizes cis-Regulated Genes by Prediction Magnitude but Fails to Predict Effect Direction: A Magnitude–Direction Dissociation in Individual-Level Expression Prediction. 10.5281/zenodo.20754855.
Pinned analysis source
https://github.com/mostafavilabuw/EnformerAssessment, commit afac6658d57bfe6e2c1f17d9a96e3593843000a2.