AI for Science 路 Antibody and protein design 路 Reliable scientific machine learning
I am a researcher at Xidian University working on machine learning for protein engineering and molecular modeling. My research combines protein foundation models, three-dimensional structure, evolutionary information, and scientific constraints to build reliable and auditable methods for antibody design and mutation-effect prediction.
My current work focuses on a practical question: how can complementary sources of biological information be combined without obscuring where a model's improvement comes from?
Framework-Conditioned Refinement for Structure-Guided Antibody CDR Design
AbRefine combines a frozen antibody inverse-folding model with a frozen antibody language model through a lightweight Information-Aware Fusion (IAF) module. IAF uses structural uncertainty, framework uncertainty, expert disagreement, and framework informativeness to determine the framework contribution at each residue.
- The fusion module contains only 49 trainable parameters.
- Native CDR identities are never supplied as model inputs.
- Recovery improves by +2.25 percentage points on the public benchmark, +3.08 points on a recent external panel, and +1.78 points on a family-disjoint reserve.
- The formulation transfers across both AntiFold and AbMPNN.
- Matched-framework controls establish that the useful signal is target-specific rather than only a generic CDR amino-acid prior.
Manuscript and public code release are in preparation. The audited research repository is currently private during submission.
Physics-Constrained Planning of Reliable Antibody Maturation Trajectories
PhyloPath treats antibody maturation as path planning over measured fitness landscapes. It prioritizes productive final variants while penalizing unfavorable intermediates and enforcing physically meaningful trajectory behavior.
- Evaluated across four antibody backgrounds and eight antibody鈥揳ntigen tasks.
- Path-aware beam search improves fully monotonic trajectory rate over matched greedy search by 0.584.
- The project includes ablations, noise-robustness analysis, frozen machine-readable results, and focused reproducibility tests.
Manuscript and code release are in preparation.
Thermodynamically Consistent Protein鈥揚rotein Mutation-Effect Prediction
DiMA-PPI represents mutation effects through a shared state potential, making reverse-mutation antisymmetry and path additivity consequences of the model formulation rather than post-hoc penalties.
Public code, manuscript, and reproducibility bundle
| Theme | Current questions |
|---|---|
| Antibody design | Structure-guided CDR sequence design, evolutionary priors, maturation trajectories |
| Protein foundation models | Combining sequence representations with local geometry and experimental context |
| Geometric learning | Structure-conditioned prediction, molecular graphs, and equivariant representations |
| Physics-informed learning | Antisymmetry, state potentials, path consistency, and thermodynamic constraints |
| Reliable scientific ML | Protein-disjoint evaluation, uncertainty, calibration, and target-level statistics |
| Generative modeling | Inverse folding, conditional sequence generation, and structure-aware sampling |
- Evaluate at the scientific unit of generalization. Protein-level claims require protein- or family-disjoint evaluation and target-level statistics.
- Separate information sources. Ablations and matched controls should show whether gains arise from structure, sequence context, prior knowledge, or data overlap.
- Build constraints into the formulation. Scientific consistency is most useful when it follows from the model rather than a post-hoc correction.
- Keep claims auditable. Frozen splits, machine-readable summaries, paired uncertainty estimates, and reproducible commands accompany reported results.
Python 路 PyTorch 路 protein language models 路 antibody inverse folding 路 geometric deep learning 路 Gaussian processes 路 scientific data pipelines 路 reproducible experimentation
Center for Biomedical-photonics and Molecular Imaging, School of Life Science and Technology, Xidian University 路 Institute of Interdisciplinary Artificial Intelligence, Xidian University 路 Xi'an, China
For research discussions or collaboration, please open a GitHub issue in the relevant public project or contact me through my GitHub profile.