Qiankang (Kant) Wang.

Genomic models have become very good at prediction. I want to know what they have learned, and whether we can read it back out as biology.

• Computational biology
• Model interpretability
• Representation learning
• Scientific computing

In the Kundaje Lab at Stanford, I compare two sequence models, AlphaGenome and ChromBPNet, to see whether they learn the same regulatory features. I look at how each one represents DNA internally, which bases drive its predictions, and which motifs it relies on. So far they find mostly the same motifs, and only a few weakly supported ones show up in just one of them. The bigger question for me is how a model's architecture and training objective shape what it learns, and how interpreting it can point to mechanisms worth testing.

Before that, I worked on the numerical side of computational biology in the Luo Lab at UC Irvine. I wrote conjugate gradient solvers for Poisson–Boltzmann electrostatics on CPUs and GPUs, and that work is now published in the Journal of Chemical Theory and Computation.

1. 2026–Present

   Regulatory genomics · Interpretability

   Kundaje Lab · Stanford

   Undergraduate Research Assistant · Stanford, CA

   Comparing what AlphaGenome and ChromBPNet learn about gene regulation, using their internal representations, base-level attributions, and motif instances on K562 and GM12878 data. So far the two models recover largely the same motifs.

   Methods & outcomes

   • Developed a framework for testing whether AlphaGenome and ChromBPNet learn the same regulatory features, comparing internal representations, base-level attributions (DeepLIFT/DeepSHAP), and motif instances (TF-MoDISco, FiNeMo) on K562 and GM12878 data, with in silico perturbation analysis as an orthogonal check.
   • Found that the two models recover largely the same regulatory motifs, with only a handful of low-support motifs unique to either model.

2. 2024–2026

   GPU scientific computing

   Luo Lab · UC Irvine

   Undergraduate Research Assistant · Irvine, CA

   Wrote the CG and BiCG solvers for AmberTorchPB, a LibTorch framework for Poisson–Boltzmann electrostatics on CPUs and GPUs, and built a Slurm pipeline that ran more than a million PBSA calculations.

   Methods & outcomes

   • Implemented conjugate gradient (CG) and biconjugate gradient (BiCG) solvers in AmberTorchPB, a LibTorch-based framework for Poisson–Boltzmann reaction field energy calculations on CPUs and GPUs.
   • Benchmarked accuracy, runtime, and memory use across numerical precisions, showing that the CG solver reproduces AMBER PBSA energies (R² = 1.00) while running over 2× faster on CPU and nearly 2× faster on GPU [1].
   • Built a Slurm-based scheduling pipeline that executed over one million PBSA energy calculations.

1. Journal of Chemical Theory and Computation · 2026

   AmberTorchPB: A Unified Framework for Poisson–Boltzmann-Based Reaction Field Energy Calculation via Tensor Computation (opens in new tab)

   Wu, Y.; Wang, Q.; Jiang, R.; Luo, R. Journal of Chemical Theory and Computation 2026, 22(7), 3554–3570. Second author

   A LibTorch framework for Poisson–Boltzmann reaction field energies on CPUs and GPUs. Its CG solver reproduces AMBER PBSA energies (R² = 1.00) while running over 2× faster on CPU and nearly 2× faster on GPU.

   My contributionImplemented the CG and BiCG solvers and benchmarked accuracy, runtime, and memory use across numerical precisions.

   Paper (opens in new tab)DOI (opens in new tab)Code (opens in new tab)

   BibTeX

   @article{wu2026ambertorchpb,
     title   = {{AmberTorchPB}: A Unified Framework for {Poisson--Boltzmann}-Based Reaction Field Energy Calculation via Tensor Computation},
     author  = {Wu, Y. and Wang, Q. and Jiang, R. and Luo, R.},
     journal = {Journal of Chemical Theory and Computation},
     year    = {2026},
     volume  = {22},
     number  = {7},
     pages   = {3554--3570},
     doi     = {10.1021/acs.jctc.6c00085}
   }

• TorchPB_DataToolBenchmark analysis for AmberTorchPB (JCTC 2026). It compares the accuracy, runtime, and memory use of the Poisson–Boltzmann solvers across numerical precisions.Jupyter Notebook (opens in new tab)
• simclr-pytorchSimCLR-style self-supervised pretraining of a ResNet-18 on 100,000 unlabeled Tiny-ImageNet images. Its frozen features reach 26.2% top-1 accuracy on the 200-class validation set with a cosine k-NN classifier.Jupyter Notebook (opens in new tab)
• personagentA conversational LLM agent with scoped memory that decides for itself when to reply. It only changes its behavior from user feedback when other evidence backs the change up, and every change is logged so it can be rolled back. Messaging platforms connect through one shared gateway.Python (opens in new tab)

Also on GitHubclaude-buddy-cyd (opens in new tab)regal (opens in new tab)shelved-ca-book-bans (opens in new tab)Decision-Tree (opens in new tab)

ML frameworks

• PyTorch
• LibTorch
• TensorFlow
• scikit-learn

ML methods

• Transformers
• Convolutional networks
• Contrastive & self-supervised learning
• Diffusion models

Interpretability

• DeepLIFT / DeepSHAP
• TF-MoDISco
• FiNeMo
• In silico perturbation

Scientific computing

• GPU optimization
• Iterative linear solvers
• Poisson–Boltzmann / PBSA
• Molecular simulation

Languages

• Python
• C++
• Java
• MATLAB
• Bash
• SQL

Tools

• Linux
• Git
• Docker
• CMake
• Slurm
• Jupyter
• LaTeX