RESEARCH / METHODS & EVIDENCE
My research spans peptide and molecular learning, clinical modeling, bioinformatics and the mechanisms of machine intelligence. Every study opens with its question, method, evidence and limits.
Peptide–protein binding ΔΔG prediction via information channel decomposition.
Pooled Spearman correlation: zero-shot PepDDG 0.619; separately calibrated PepDDG-Cal 0.691.
Author-provided V4 · accepted NeurIPS 2026 · Source ↗
Multi-view heterogeneous graphs for peptide representations across natural and non-natural amino acids.
Noncanonical affinity PCC reaches 0.873 after two-stage training. Canonical CPP AUC is 0.838 versus ESM 0.885. These are 2023 manuscript values, not the 2025 journal tables.
Author manuscript · 2023 · Tables 2, 3, 5 · Source ↗
Sequence–structure contrastive learning for peptide representation.
Table 1 reports CPP AUC 0.874 versus ESM2 0.864; affinity Spearman 0.447 trails fine-tuned ESM at 0.473. Final 2026 journal results are not substituted for these v1 values.
Author manuscript v1 · 2024 · Tables 1, 3 · Source ↗
Quantitative definition and benchmarking of activity cliffs in antimicrobial peptides.
The v2 abstract reports 33-layer ESM2 Spearman 0.4669. Figures 5–6 show that ranking, regression error and top-50 recall respond differently to cliff thresholds.
Author manuscript v2 · 3 November 2024 · Figures 5–6 · Source ↗
Few-shot and contrastive learning for scarce and imbalanced molecular property labels.
10-shot AUC exceeds Pre-PAR on all four datasets; 1-shot results trail Pre-PAR on Tox21 and SIDER.
Published article · 2024 · Table 3 · Source ↗
Multimodal molecular graphs and chemical knowledge, using ChatGPT-generated knowledge in the reported implementation.
FreeSolv RMSE is 0.796 ± 0.09; external solubility RMSE is 0.651 ± 0.046. The comparison retains stronger ESOL/Lipophilicity and HIV baselines.
Final article · online 2025, volume 2026 · Source ↗
Recognizing chemical structures from images.
Repository SMI scores are 0.650202, 0.604444, 0.859930 and 0.746284 respectively. These are implementation-reported values, separate from the final journal tables.
Public repository README · benchmark table · Source ↗
Helper-enhanced feature selection via Pareto-optimized genetic search.
Table 2 improves DLBCL BHOA accuracy from 97.50% to 99.88%, but Semeion FTGGA falls from 97.80% to 96.62%.
Author manuscript v1 · 21 October 2025 · §§4–5 / Table 2 · Source ↗
Interpretable prognostic feature selection and compression for electronic health records.
Table 2 reports mortality AUROC 0.9054 and decompensation AUROC 0.9143; selected baseline comparisons appear below.
Author manuscript v2 · 20 June 2025 · Table 2 · Source ↗
Adversarial heterogeneous graph attention for fair electronic-health-record representations.
The final abstract reports AUROC 0.9021 for mortality and 0.9217 for decompensation.
Final article · 2024 · Abstract · Source ↗
RESEARCH / METHODS & CONNECTIONS
Tools, algorithms and public validation records.
Combine generative structure models and sequence design to propose candidates for further testing.
Tool interfaces implemented; public execution validation pending.
Compare mutation candidates and structure–activity relationships to study links between sequence, structure and binding.
Ranking and analysis code implemented; public validation pending.
Compare evidence for binding, selectivity, stability and developability to study interpretable trade-offs.
Evaluation code implemented; optimization strategies remain under study.
Use protein language and structure models for representation learning, prediction and design, while testing the limits of transfer.
ESM interface implemented; AlphaFold 3 has an actual public structure example.
Public structure example · 2026-10-13 · 17 aa ↗
