# Reviewer-critique quantitative upgrade

Generated: 2026-07-16

## DMD signature source sensitivity

Nine signature variants were audited (full, four leave-one-source-out, and four source-only). Leave-one-source-out/source-only analyses are robustness checks, not alternative ground truths. Full-score recomputation was checked against the released candidate score in `signature_sensitivity_metadata.json`.

| variant | n_signature_genes | n_shared_with_full | n_finite_shared | spearman_vs_full | cosine_vs_full | n_positive | n_negative |
| --- | --- | --- | --- | --- | --- | --- | --- |
| full_all_sources | 17919 | 17919 | 17918 | 1 | 1 | 8935 | 8983 |
| leave_out__dmd_single_cell_baseline_pseudobulk | 17544 | 17544 | 17543 | 0.6699 | 0.9997 | 8214 | 9329 |
| source_only__dmd_single_cell_baseline_pseudobulk | 16687 | 16687 | 16687 | 0.4993 | 0.01342 | 9098 | 7589 |
| leave_out__dmd_single_cell_delta_did_final_labels | 17876 | 17876 | 17875 | 0.8234 | 1 | 9264 | 8611 |
| source_only__dmd_single_cell_delta_did_final_labels | 16058 | 16058 | 16058 | 0.1447 | 0.003322 | 7397 | 8661 |
| leave_out__sema3c_deseq2_source_state | 17153 | 17153 | 17153 | 0.7319 | 0.04034 | 9025 | 8128 |
| source_only__sema3c_deseq2_source_state | 14949 | 14949 | 14948 | 0.7442 | 0.9977 | 7214 | 7734 |
| leave_out__sema3c_formal_nichenet_target_state | 17910 | 17910 | 17909 | 0.9842 | 1 | 8861 | 9048 |
| source_only__sema3c_formal_nichenet_target_state | 11851 | 11851 | 11851 | 0.5274 | 0.2258 | 6584 | 5267 |

Pairwise source concordance (long form; Figure S1 renders Pearson and cosine matrices):

| source_a | source_b | n_shared_genes | pearson | spearman | cosine |
| --- | --- | --- | --- | --- | --- |
| dmd_single_cell_baseline_pseudobulk | dmd_single_cell_baseline_pseudobulk | 16687 | 1 | 1 | 1 |
| dmd_single_cell_baseline_pseudobulk | dmd_single_cell_delta_did_final_labels | 15980 | -0.2229 | -0.4457 | -0.2175 |
| dmd_single_cell_baseline_pseudobulk | sema3c_deseq2_source_state | 13769 | 0.01828 | 0.163 | 0.02122 |
| dmd_single_cell_baseline_pseudobulk | sema3c_formal_nichenet_target_state | 11456 | 0.1619 | 0.2183 | 0.1641 |
| dmd_single_cell_delta_did_final_labels | dmd_single_cell_baseline_pseudobulk | 15980 | -0.2229 | -0.4457 | -0.2175 |
| dmd_single_cell_delta_did_final_labels | dmd_single_cell_delta_did_final_labels | 16058 | 1 | 1 | 1 |
| dmd_single_cell_delta_did_final_labels | sema3c_deseq2_source_state | 13473 | 0.002163 | -0.06227 | 0.003237 |
| dmd_single_cell_delta_did_final_labels | sema3c_formal_nichenet_target_state | 11363 | -0.03356 | -0.04116 | -0.0324 |
| sema3c_deseq2_source_state | dmd_single_cell_baseline_pseudobulk | 13769 | 0.01828 | 0.163 | 0.02122 |
| sema3c_deseq2_source_state | dmd_single_cell_delta_did_final_labels | 13473 | 0.002163 | -0.06227 | 0.003237 |
| sema3c_deseq2_source_state | sema3c_deseq2_source_state | 14948 | 1 | 1 | 1 |
| sema3c_deseq2_source_state | sema3c_formal_nichenet_target_state | 11835 | 0.2332 | 0.4776 | 0.2336 |
| sema3c_formal_nichenet_target_state | dmd_single_cell_baseline_pseudobulk | 11456 | 0.1619 | 0.2183 | 0.1641 |
| sema3c_formal_nichenet_target_state | dmd_single_cell_delta_did_final_labels | 11363 | -0.03356 | -0.04116 | -0.0324 |
| sema3c_formal_nichenet_target_state | sema3c_deseq2_source_state | 11835 | 0.2332 | 0.4776 | 0.2336 |
| sema3c_formal_nichenet_target_state | sema3c_formal_nichenet_target_state | 11851 | 1 | 1 | 1 |

Top candidate stability across signature variants:

| perturbation | n_variants | mean_rescue | sd_rescue | min_rescue | max_rescue | mean_rank | best_rank | worst_rank | top4_frequency |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| ZNF236 | 9 | 0.04478 | 0.04434 | -0.06716 | 0.08867 | 5.111 | 2 | 21 | 0.7778 |
| ZNF133 | 9 | 0.03277 | 0.04325 | -0.04316 | 0.06754 | 6.222 | 2 | 16 | 0.6667 |
| GFOD2 | 9 | 0.03469 | 0.04974 | -0.0484 | 0.07663 | 6.444 | 1 | 19 | 0.5556 |
| ZFP69B | 9 | 0.02572 | 0.04137 | -0.04871 | 0.0591 | 8.667 | 4 | 21 | 0.5556 |
| DNAAF3 | 9 | 0.03033 | 0.02572 | -0.03549 | 0.05515 | 7.778 | 1 | 20 | 0.2222 |
| WDR4 | 9 | 0.02249 | 0.02047 | -0.0198 | 0.05536 | 9.778 | 1 | 13 | 0.2222 |
| CPEB1 | 9 | 0.02894 | 0.03563 | 0.01179 | 0.1215 | 11.33 | 1 | 17 | 0.2222 |
| EPS8L1 | 9 | 0.02611 | 0.04583 | -0.003085 | 0.1098 | 14.44 | 1 | 20 | 0.2222 |
| RAC3 | 9 | 0.03589 | 0.02592 | 0.01207 | 0.08584 | 8.667 | 2 | 12 | 0.1111 |
| ADAM10 | 9 | 0.0213 | 0.02406 | -0.01209 | 0.07556 | 11.44 | 2 | 15 | 0.1111 |
| RNASEH2C | 9 | 0.004199 | 0.03197 | -0.05224 | 0.03904 | 12.33 | 4 | 18 | 0.1111 |
| MPHOSPH6 | 9 | 0.01964 | 0.03279 | -0.0445 | 0.08487 | 12.78 | 4 | 17 | 0.1111 |

## Candidate-tier multiverse

A-tier membership was recomputed across 144 reasonable rule combinations. Missing myoblast-screen coverage and moderate dependency were each tested as either neutral/penalty-only or A-blocking conditions. The word `safe` is not used in the revised tier labels.

| perturbation | n_scenarios | a_frequency | b_frequency | c_frequency | safety_review_frequency | priority_tier | observed_rescue_score | observed_n_cells | decision_score | moderate_dependency_flag | gse293514_pos_fdr | robustness_label |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| ZNF133 | 144 | 0.6667 | 0.3333 | 0 | 0 | A_observed_priority_validate_first | 0.06382 | 84 | 41.88 | False | 1 | conditional_A |
| GFOD2 | 144 | 0.5 | 0.5 | 0 | 0 | A_observed_priority_validate_first | 0.07411 | 127 | 41.97 | False |  | conditional_A |
| ZFP69B | 144 | 0.3333 | 0.6667 | 0 | 0 | A_observed_priority_validate_first | 0.05634 | 81 | 38.15 | False |  | fragile_A |
| ZNF236 | 144 | 0.25 | 0.75 | 0 | 0 | A_observed_priority_validate_first | 0.05963 | 202 | 30.11 | True |  | fragile_A |
| MON1A | 144 | 0.1667 | 0.8333 | 0 | 0 | B_multisource_followup | 0.04651 | 94 | 23.98 | False | 1 | fragile_A |
| CALR | 144 | 0.1111 | 0.8889 | 0 | 0 | B_multisource_followup | 0.03427 | 86 | 20.11 | False | 1 | fragile_A |
| CPEB1 | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.01182 | 100 | 22.36 | False |  | not_A_in_multiverse |
| ADAM10 | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.01658 | 161 | 17.33 | False | 1 | not_A_in_multiverse |
| DDX19B | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.01813 | 145 | 17.06 | False | 1 | not_A_in_multiverse |
| DNAAF3 | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.03444 | 115 | 12.6 | False |  | not_A_in_multiverse |
| EHMT2 | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.0298 | 101 | 12.4 | False |  | not_A_in_multiverse |
| LMO2 | 144 | 0 | 1 | 0 | 0 | B_multisource_followup | 0.007237 | 96 | 12.35 | False |  | not_A_in_multiverse |
| MPHOSPH6 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.01563 | 83 | 9.215 | False | 0.7787 | not_A_in_multiverse |
| DNM1 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.00238 | 116 | 1.267 | False |  | not_A_in_multiverse |
| RAC3 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.02478 | 92 | -0.6079 | True |  | not_A_in_multiverse |
| RNASEH2C | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.02169 | 132 | -2.952 | True | 0.685 | not_A_in_multiverse |
| EPS8L1 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.003565 | 269 | -5.783 | False |  | not_A_in_multiverse |
| RNF8 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.03939 | 122 | -6.678 | True |  | not_A_in_multiverse |
| NAGLU | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.01102 | 88 | -12.04 | True | 0.685 | not_A_in_multiverse |
| INTS13 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.01927 | 157 | -19.38 | True |  | not_A_in_multiverse |
| WDR4 | 144 | 0 | 0 | 1 | 0 | C_context_or_calibration_only | 0.01852 | 108 | -24.12 | True | 1 | not_A_in_multiverse |

## 21-gene design sensitivity

This is an empirical design-sensitivity analysis, not observed post-hoc power. Positive values favor the model. The strict composite remains direction-limited because the observed cosine gain is negative for every direct head.

| model | endpoint | n_genes | mean_favorable_difference | sd_paired_difference | paired_standardized_effect_dz | favorable_gene_fraction | mde_80pct_power_n21_unadjusted | mde_80pct_power_n21_bonferroni12 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| ridge_v1 | rmse_gain_vs_zero | 21 | 0.006045 | 0.01777 | 0.3401 | 0.5238 | 0.009643 | 0.0135 |
| ridge_v1 | rmse_gain_vs_train_mean | 21 | 0.006371 | 0.0184 | 0.3462 | 0.7143 | 0.009984 | 0.01397 |
| ridge_v1 | cosine_gain_vs_train_mean | 21 | -0.04361 | 0.2135 | -0.2043 | 0.5238 | 0.1158 | 0.1621 |
| ridge_v2 | rmse_gain_vs_zero | 21 | 0.006808 | 0.01691 | 0.4025 | 0.5714 | 0.009176 | 0.01284 |
| ridge_v2 | rmse_gain_vs_train_mean | 21 | 0.007134 | 0.01817 | 0.3925 | 0.7619 | 0.00986 | 0.0138 |
| ridge_v2 | cosine_gain_vs_train_mean | 21 | -0.03469 | 0.1739 | -0.1995 | 0.381 | 0.09434 | 0.132 |
| mlp_v1 | rmse_gain_vs_zero | 21 | 0.002712 | 0.01655 | 0.1638 | 0.5714 | 0.008982 | 0.01257 |
| mlp_v1 | rmse_gain_vs_train_mean | 21 | 0.003038 | 0.01682 | 0.1806 | 0.619 | 0.009127 | 0.01277 |
| mlp_v1 | cosine_gain_vs_train_mean | 21 | -0.0395 | 0.1653 | -0.239 | 0.619 | 0.0897 | 0.1255 |
| mlp_v2 | rmse_gain_vs_zero | 21 | 0.002641 | 0.01767 | 0.1495 | 0.5714 | 0.009586 | 0.01342 |
| mlp_v2 | rmse_gain_vs_train_mean | 21 | 0.002968 | 0.01671 | 0.1775 | 0.619 | 0.009069 | 0.01269 |
| mlp_v2 | cosine_gain_vs_train_mean | 21 | -0.03793 | 0.1441 | -0.2633 | 0.619 | 0.07817 | 0.1094 |

| model | assumed_sample_size | bootstrap_replicates | alpha_test_only_bonferroni12 | rmse_both_endpoints_detection_probability | direction_endpoint_detection_probability | strict_composite_detection_probability | interpretation |
| --- | --- | --- | --- | --- | --- | --- | --- |
| ridge_v1 | 21 | 10000 | 0.004167 | 0.0035 | 0.006 | 0.0005 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v1 | 40 | 10000 | 0.004167 | 0.0481 | 0.0008 | 0.0003 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v1 | 80 | 10000 | 0.004167 | 0.4284 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v1 | 160 | 10000 | 0.004167 | 0.9285 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v2 | 21 | 10000 | 0.004167 | 0.0069 | 0.0007 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v2 | 40 | 10000 | 0.004167 | 0.1217 | 0.0003 | 0.0001 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v2 | 80 | 10000 | 0.004167 | 0.6724 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| ridge_v2 | 160 | 10000 | 0.004167 | 0.9886 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v1 | 21 | 10000 | 0.004167 | 0 | 0.0056 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v1 | 40 | 10000 | 0.004167 | 0.0004 | 0.0006 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v1 | 80 | 10000 | 0.004167 | 0.0041 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v1 | 160 | 10000 | 0.004167 | 0.0584 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v2 | 21 | 10000 | 0.004167 | 0 | 0.0014 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v2 | 40 | 10000 | 0.004167 | 0.0004 | 0.0002 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v2 | 80 | 10000 | 0.004167 | 0.003 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |
| mlp_v2 | 160 | 10000 | 0.004167 | 0.041 | 0 | 0 | Design sensitivity under empirical paired-effect distribution; not observed post-hoc power. |

## Repeated ridge effect magnitude

RMSE, raw cosine, and residual cosine are reported separately. Residual cosine remains a sensitivity endpoint.

| model | seed | n_test_genes | mean_rmse_gain_vs_zero | median_rmse_gain_vs_zero | win_fraction_vs_zero | paired_dz_vs_zero | pseudo_r2_vs_zero | mean_rmse_gain_vs_train_mean | win_fraction_vs_train_mean | paired_dz_vs_train_mean | pseudo_r2_vs_train_mean | mean_raw_cosine_gain_vs_train_mean | raw_cosine_win_fraction | mean_residual_cosine |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| ridge_v1 | 20260712 | 540 | 0.01114 | 0.002395 | 0.5648 | 0.3234 | 0.2589 | 0.005339 | 0.6407 | 0.2552 | 0.09945 | 0.000651 | 0.5426 | 0.255 |
| ridge_v1 | 20260713 | 540 | 0.01135 | 0.002817 | 0.5407 | 0.3229 | 0.2633 | 0.005343 | 0.6222 | 0.2649 | 0.1064 | -0.00315 | 0.5444 | 0.2425 |
| ridge_v1 | 20260714 | 540 | 0.0108 | 0.003597 | 0.5685 | 0.3269 | 0.2502 | 0.005514 | 0.6611 | 0.2812 | 0.09902 | -0.0008238 | 0.5481 | 0.2643 |
| ridge_v1 | 20260715 | 540 | 0.01015 | 0.0007452 | 0.5167 | 0.2832 | 0.2641 | 0.006048 | 0.6704 | 0.2728 | 0.1248 | -0.001428 | 0.5444 | 0.2669 |
| ridge_v1 | 20260716 | 540 | 0.01262 | 0.004277 | 0.5963 | 0.3645 | 0.2741 | 0.005855 | 0.6315 | 0.2875 | 0.119 | -0.004148 | 0.5241 | 0.2483 |
| ridge_v2 | 20260712 | 540 | 0.009979 | 0.003445 | 0.5574 | 0.2946 | 0.2337 | 0.004174 | 0.6685 | 0.1992 | 0.06884 | 0.002127 | 0.5593 | 0.2416 |
| ridge_v2 | 20260713 | 540 | 0.01068 | 0.003243 | 0.5463 | 0.329 | 0.2518 | 0.004667 | 0.6407 | 0.2794 | 0.09246 | 0.0004528 | 0.5722 | 0.2243 |
| ridge_v2 | 20260714 | 540 | 0.009952 | 0.003205 | 0.5574 | 0.3205 | 0.237 | 0.004669 | 0.6407 | 0.2787 | 0.08315 | 0.001359 | 0.5352 | 0.2295 |
| ridge_v2 | 20260715 | 540 | 0.009512 | 0.001542 | 0.5241 | 0.2878 | 0.251 | 0.00541 | 0.6685 | 0.2913 | 0.1093 | 0.004226 | 0.5574 | 0.247 |
| ridge_v2 | 20260716 | 540 | 0.01179 | 0.005229 | 0.5944 | 0.366 | 0.26 | 0.005029 | 0.6389 | 0.3 | 0.1018 | 0.001583 | 0.537 | 0.2301 |

## HepG2-to-muscle applicability audit

Target expression was assessed in four non-injured GSE270868 quadriceps references. Expression coverage is necessary but not sufficient for muscle-context validity.

| group | n_genes | n_detected_any | fraction_detected_any | n_expressed_ge_0_1pct_cells | fraction_expressed_ge_0_1pct_cells | n_expressed_ge_1pct_cells | fraction_expressed_ge_1pct_cells | median_fraction_cells_expressing | reference_samples | reference_cells |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| all_hepg2_targets | 2393 | 2357 | 0.985 | 2287 | 0.9557 | 2107 | 0.8805 | 0.06713 | 4 | 16431 |
| eligible_2160 | 2160 | 2131 | 0.9866 | 2070 | 0.9583 | 1908 | 0.8833 | 0.06686 | 4 | 16431 |
| candidate_matrix_21 | 21 | 21 | 1 | 19 | 0.9048 | 17 | 0.8095 | 0.0423 | 4 | 16431 |
| gse293514_overlap | 1575 | 1565 | 0.9937 | 1564 | 0.993 | 1516 | 0.9625 | 0.08052 | 4 | 16431 |

| metric | value | n_genes | boundary |
| --- | --- | --- | --- |
| spearman_hepg2_var_mean_vs_healthy_muscle_mean_umi | 0.2378 | 2169 | Cross-platform expression-rank applicability audit; not baseline-state equivalence. |
| top_quartile_expression_jaccard | 0.2285 | 2169 | Top-quartile overlap within the HepG2 target universe. |

GTEx v8 skeletal-muscle median TPM coverage (2,394 perturbation labels include one non-targeting control; the gene-target denominator is 2,393):

| group | n_gene_targets | n_gtex_symbol_matched | fraction_gtex_symbol_matched | n_tpm_gt_0_1 | fraction_all_targets_tpm_gt_0_1 | fraction_matched_targets_tpm_gt_0_1 | n_tpm_gt_1_0 | fraction_all_targets_tpm_gt_1_0 | fraction_matched_targets_tpm_gt_1_0 | n_tpm_gt_5_0 | fraction_all_targets_tpm_gt_5_0 | fraction_matched_targets_tpm_gt_5_0 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| all_gene_targets_2393 | 2393 | 2361 | 0.9866 | 2253 | 0.9415 | 0.9543 | 2133 | 0.8913 | 0.9034 | 1805 | 0.7543 | 0.7645 |
| eligible_2160 | 2160 | 2133 | 0.9875 | 2044 | 0.9463 | 0.9583 | 1937 | 0.8968 | 0.9081 | 1640 | 0.7593 | 0.7689 |
| candidate_matrix_21 | 21 | 21 | 1 | 20 | 0.9524 | 0.9524 | 16 | 0.7619 | 0.7619 | 8 | 0.381 | 0.381 |
| gse293514_overlap | 1575 | 1562 | 0.9974 | 1561 | 0.9968 | 0.9994 | 1552 | 0.9911 | 0.9936 | 1452 | 0.9272 | 0.9296 |

## Point-estimate trap

Historical mean-RMSE decisions and the publication-facing paired composite lead to different conclusions. This is a model-level audit; no unsupported candidate-level counterfactual policy was invented.

| decision_path | n_model_split_decisions | n_pass | pass_fraction | interpretation |
| --- | --- | --- | --- | --- |
| historical point-estimate RMSE-only | 8 | 8 | 1 | All four heads appear positive on validation and test means. |
| paired uncertainty + Holm + direction composite | 8 | 0 | 0 | No head has publication-facing composite support. |

## Claim boundary

These upgrades quantify robustness and applicability. They do not create guide-resolved muscle perturbation-transcriptome validation, target efficacy, or clinical evidence.