Journal: bioRxiv

Article Title: Vibe Coding Specificity Foundation Models

doi: 10.64898/2026.06.04.730134

Figure Lengend Snippet: All six SFMs share an identical dual-encoder, symmetric InfoNCE architecture; only the encoders, paired training data, and hard-negative strategy differ. (A) Agent-to-target R@1 from pools of 512 candidates (R@1 = 0.2% random; 100 random resamplings per fold). Blue bars, in-distribution (identity_100); white bars, best out-of-distribution threshold passing leakage verification (<10%). OOD thresholds: tSFM mmseqs_060, eSFM mmseqs_080, crisprSFM hamming_080, dtSFM mmseqs_080; mhcSFM uses a 14-allele zero-shot holdout; mir-SFM uses a by-miRNA holdout. (B) Target-to-agent R@1 (inverse direction), same evaluation. Error bars ± s.d. across folds; bars without error bars denote single-fold evaluations. Fold 4 excluded from means where validation-set degeneracy was detected by the orthogonal audit. (C) mir-SFM versus seed-matching baseline, stratified by canonical and non- canonical miRNA–target interactions. (D) MS-presentation precision on the Gurung et al. cancer neoantigen benchmark (86 MS-validated peptide–HLA pairs): NetMHCpan 4.1 EL alone versus NetMHCpan ∩ mhcSFM cascade. (E) CRISPR off-target precision at ≤4 mismatches on the CRISPRoffT held-out benchmark (802 candidate sites): Hamming distance alone versus Hamming → crisprSFM re-ranking cascade. (F) SFM R@1 versus domain-specific rule-based baseline R@1 for crisprSFM (Hamming distance) and mir-SFM (seed-matching).

Article Snippet: The Gurung et al. (2023) Nature Biotechnology cancer- neoantigen benchmark was consistently labeled as “Pyke et al. 2024” in development scripts, CSVs, and the writeup — a first-author misattribution.

Techniques: Biomarker Discovery, CRISPR

Journal: Scientific Reports

Article Title: Proactive soft-failure prediction in optical transport networks via physics-inspired features and Infrastructure-as-Code orchestration

doi: 10.1038/s41598-026-52186-3

Figure Lengend Snippet: System architecture. The multi-physics stochastic simulation and the Ghosh–Adhya real-data benchmark feed a shared feature-extraction pipeline that produces 15-dimensional physics-inspired feature vectors (10 OSNR lags, velocity, acceleration, rolling mean, rolling standard deviation). The Random Forest regressor emits time-to-failure estimates; upon three consecutive sub-threshold predictions (persistence filter), the orchestration layer commits a desired-state change to a Git repository (Fig. ), triggering Kubernetes reconciliation and a Terraform-driven make-before-break migration over NETCONF/OpenROADM.

Article Snippet: The predictor is evaluated on (a) a calibrated multi-physics stochastic simulation spanning five degradation modes and (b) the Ghosh–Adhya (2025) Mendeley optical soft-failure benchmark comprising 756 real lightpaths with OSNR, BER, laser current, and received optical power across 900-sample trajectories for four failure classes.

Techniques: Extraction, Standard Deviation, Migration

Journal: Scientific Reports

Article Title: Proactive soft-failure prediction in optical transport networks via physics-inspired features and Infrastructure-as-Code orchestration

doi: 10.1038/s41598-026-52186-3

Figure Lengend Snippet: Empirical characterization of the Ghosh–Adhya (2025) real-data benchmark (training split, 3,024 trajectories). Percentage of trajectories crossing the 18 dB soft-failure alarm and the 15 dB hard-failure threshold, by class. EDFA and NLI failures produce strong OSNR signatures (52% and 82% hard-threshold crossings respectively); ECL failures are OSNR-invariant due to AGC compensation (0.5% crossings, indistinguishable from no-failure baseline), establishing the scope of an OSNR-based predictor.

Article Snippet: The predictor is evaluated on (a) a calibrated multi-physics stochastic simulation spanning five degradation modes and (b) the Ghosh–Adhya (2025) Mendeley optical soft-failure benchmark comprising 756 real lightpaths with OSNR, BER, laser current, and received optical power across 900-sample trajectories for four failure classes.

Techniques: