面向蛋白质工程与分子建模的 AI4Science 研究，关注几何表示、生成建模、物理约束学习、不确定性估计与跨蛋白泛化。
I work on machine learning methods for protein engineering and molecular modeling, with an emphasis on geometric representation learning, scientific priors, uncertainty, and out-of-distribution generalization.
- Protein representation learning — combining protein language models, structure-aware embeddings, local geometry, and mutation descriptors.
- Scientific machine learning — encoding symmetry, antisymmetry, consistency, and other domain constraints into model objectives.
- Generative modeling — inverse folding and flow-based approaches for sequence and structure generation.
- Reliable prediction — probabilistic learning, uncertainty decomposition, protein-disjoint evaluation, and auditable experiments.
A structure-aware learning framework for single- and double-mutation stability prediction on unseen proteins.
- Fuses ESM-2 sequence deltas, SaProt structural context, physicochemical changes, and local geometry.
- Uses Pooling by Multihead Attention (PMA) for permutation-invariant mutation-set encoding.
- Models single and double mutants with separate sparse variational Gaussian process (SVGP) heads.
- Aligns training with scientific evaluation through protein-balanced sampling and within-protein RankNet pretraining.
- Evaluated on protein-disjoint MegaScale splits across three seeds.
| Task | Protein-wise Spearman | Pooled Spearman |
|---|---|---|
| Single mutants | 0.7654 ± 0.0033 | 0.7152 ± 0.0035 |
| Double mutants | 0.6753 ± 0.0200 | 0.5533 ± 0.0030 |
Code and reproducibility · Manuscript and evidence package
A validated trajectory-planning framework for navigating measured antibody fitness landscapes while penalizing unfavorable intermediates. The frozen evaluation spans four antibody backgrounds and eight antibody--antigen tasks, with ablations, noise robustness analysis, machine-readable results, and a one-command audit covering 22 focused tests and manuscript regeneration.
- Path-aware beam search improved fully monotonic trajectory rate over matched greedy search by 0.584 (antibody-cluster 95% CI 0.371--0.796).
- Final measured gain improved by 0.087 (95% CI 0.058--0.108).
- Negative results and the failed global-oracle gate are retained explicitly to keep the public claim boundary auditable.
Code, frozen results, and reproducibility package
A compact manuscript and reproducibility bundle for a shared four-state potential model. The formulation enforces reverse-mutation antisymmetry and path additivity by construction, and includes strict-split, external-test, ablation, bootstrap, and path/cycle audits.
Code, manuscript, and reproducibility bundle
The table distinguishes validated public work from research explorations so that project maturity is explicit.
| Area | Methods explored | Scientific use | Status |
|---|---|---|---|
| Protein representations | ESM-2, SaProt, mutation-delta features, local structural descriptors | Mutation-effect prediction and protein generalization | Validated in public project |
| Set and ranking learning | PMA, RankNet, protein/quantile-balanced sampling | Variable-order mutations and imbalanced measurements | Validated in public project |
| Probabilistic learning | SVGP, predictive mean/variance, uncertainty-aware ranking | Risk-aware mutation prioritization | Validated in public project |
| Physics-constrained learning | Antisymmetric ΔΔG objectives, state potentials, cycle consistency, path-aware planning | Consistent mutation prediction and antibody evolutionary trajectories | Validated in public projects |
| Geometric learning | Molecular graphs, geometric GNNs, SE(3)-aware/equivariant representations | Structure-conditioned molecular learning | Research exploration |
| Conditional generation | Antibody inverse folding, ProteinMPNN-style sequence design | Structure-to-sequence generation | Research exploration |
| Flow-based generation | Flow matching and structure-aware generative models | Molecular sequence/structure generation | Research exploration |
| Distributional geometry | Optimal transport on molecular representations and surfaces | Comparing structured molecular distributions | Research exploration |
| Reliability | Uncertainty decomposition, structural identifiability, OOD evaluation | Detecting ambiguous or unreliable predictions | Research exploration |
- Evaluate at the scientific unit of generalization. Use protein-disjoint splits and protein-level metrics when the goal is performance on unseen proteins.
- Match the objective to the deployment question. Ranking, calibration, and balanced sampling can matter as much as architecture choice.
- Encode trustworthy structure. Scientific constraints should improve consistency without hiding empirical limitations.
- Keep claims auditable. Separate multi-seed results, ablations, controls, and exploratory findings.
Python · PyTorch · protein language models · graph and geometric deep learning · Gaussian processes · scientific data pipelines · reproducible experimentation
I am developing AI4Science methods that connect protein foundation models with geometric and physics-informed learning for reliable protein design and mutation-effect prediction.