Head-to-head against NetMHCpan-4.2c
Four independent benchmarks, temporal-gated to first-public date ≥ 2025-08-08 (day after NetMHCpan-4.2 release), overlap-audited against a 17.6 M-row union training catalogue. Primary result below is the multi-study mono-allelic HLA-I EL benchmark. Interactive charts — hover, zoom, or export any panel as PNG. Full methodology + data sources →
ROC curve — per allele interactive · toggle allele
DeepNeo's presentation (EL) head vs NetMHCpan-4.2c on the mono_el_v2 candidates, per HLA allele. Pick an allele from the dropdown; hover for FPR / TPR values.
Precision–Recall per allele
Same allele + comparison as above. AUPR shown in the legend.
Per-allele AUROC scatter DeepNeo EL vs NetMHCpan-4.2
Each dot is one HLA allele; point size ∝ √n_positives. Points above the y = x line are DeepNeo wins. Hover for allele name and Δ AUROC.
Per-allele Δ AUROC DeepNeo EL − NetMHCpan-4.2, sorted
Positive bars = DeepNeo wins. Every supported allele bar is positive. HLA-B*07:01 is unsupported by NetMHCpan-4.2, so it has no bar (DeepNeo covers it at AUROC 0.79 standalone).
Per-allele table sorted by Δ, all 13 alleles
| HLA allele | n | n_pos | DeepNeo EL AUROC | NetMHCpan-4.2 AUROC | Δ AUROC | DeepNeo EL AUPR | NetMHCpan AUPR |
|---|
Correlation with binder label
Spearman and Pearson between predictor score and binary binder label, on the pooled 496,240-row set.
Top-k precision EL head
Precision at k of the highest-scoring predictions. DeepNeo tops the precision at every k ≥ 50; NetMHCpan matches at k = 10.
Decontamination audit every benchmark, all 4
Overlap between each benchmark's positives and the 17.6 M-row union training catalogue (BA + EL + IEDB + CEDAR partitions from NetMHCpan-4.2's training). mono_el_v2 is 86 % exact-disjoint — the honest novel-generalization test set. mono_el_v1 is superseded because its 74 % training leak on positives inflated absolute AUROC (win still holds on the disjoint recompute).
v1 exact-disjoint recompute the win survives decontamination
On the 540-positive exact-disjoint subset of the (superseded) mono_el_v1 benchmark, DeepNeo EL widens its lead over NetMHCpan-4.2 from +0.012 to +0.036 AUROC with non-overlapping CIs. The v1 result is not memorization; it survives removing the training-set leak.