CHENWEI ZHANG | Amazon (PhD at UBC)

About Me

Life is a marathon, not a sprint.

I am Chenwei Zhang (张晨伟). I am an Applied Scientist at Amazon, working on agentic AI and LLM post-training.

I received my Ph.D. degree from Computer Science at the University of British Columbia, co-supervised by Prof. Anne Condon and Prof. Khanh Dao Duc, affiliated with the Algorithms Lab. Before that, I earned my M.Sc. (2021, supervised by Prof. Pavle Radovanovic) and my B.Sc. (2018, Dean’s Honours List) from the University of Waterloo.

My PhD research focused on developing deep generative models for biology, with an emphasis on protein structure modeling, 3D cryo-EM image processing, and DNA secondary structure representation and visualization.

I also had research experience on large language model post-training, multimodal foundation models, and AI agents. In particular, I enjoy designing multi-agent frameworks and translating them into practical, product-ready AI systems.

Research Interests

AIML: Agentic AI, Large Language Models, Generative Models, Multimodal Learning, Learning Representations and Visualization
AI4Business: Autonomous AI Agents for Ads, Forecasting Foundation Models
AI4Science: Protein Structure Modeling, Drug Discovery, Cryo-EM Image Processing, Medical Imaging, DNA Kinetics

Work Experience

Technical Career

[Sept. 2026 - Present] Full-time Applied Scientist at Amazon, building autonomous AI agents and post-training Large Language models.
[Apr. 2026 - Aug. 2026] Applied Scientist Intern at Amazon, constructing large language models for protein study.
[Jan. 2025 - Aug. 2025] Machine Learning Scientist Intern at Prenuvo, building multimodal vision-language foundation models for medical imaging and radiology report generation.
[Jun. 2023 - Jun. 2024] Machine Learning Scientist Intern at Amgen, developing deep learning approaches for protein structure modeling and cryo-EM 3D image analysis.

Teaching Assistant

CPSC 340/532M: Machine Learning and Data Mining
CPSC 330: Applied Machine Learning
CPSC 322: Introduction to Artificial Intelligence

AI Products

SmartFind AI: Multi-Agent-powered platform to discover top professors, research labs, and universities, analyze research fit, and generate personalized application materials.
Radgent AI: Resume intelligence agent for strategic job applications: resume rewrite, resume optimization, outreach email, and resume translation.
PetMinder AI: AI-powered platform for pet health monitoring and management, with internal AI assistant analyzing pet photos to automatically schedule vaccines, grooming, and care tasks.

News

[Aug. 2025] 🎉 Our paper on improving synthetic cryo-EM density maps with generative adversarial networks is accepted at Bioinformatics Advances.
[Jul. 2025] 🎉 Our paper on survey and benchmark of DL methods for atomic model building from cryo-EM density maps is accepted at Briefings in Bioinformatics.
[Jul. 2025] 🗣️ I will attend The Molecular Programming Decadal Flightplan in Seattle, USA.
[Jul. 2025] 🗣️ I will give a poster presentation at Generative AI and Biology (GenBio) workshop at ICML 2025 conference in Vancouver, Canada.
[Jun. 2025] 🎉 Our paper on enhancing 3D Cryo-EM density maps of protein structures with structure-aware multimodal U-Nets is accepted at GenBio, ICML 2025.

Previous news

[Tools] ⚒️ I made a python package to find a specified number of shortest paths (top K) in a weighted graph by Dijkstra's algorithm.
[Mar. 2025] 🏦 I did a two-month remote internship at a Venture Capital (CAS Star) for AI investments.
[Jan. 2025] 👨‍💻 I am starting a Machine Learning Scientist Internship at Prenuvo, focusing on builindg multimodal vision-language foundation models for medical imaging.
[Sept. 2024] 🗣️ I presented a poster at DNA30 conference at Johns Hopkins University, Baltimore, USA.
[Jul. 2024] 🗣️ I will attend Oxford Machine Learning Summer School (OxML 2024) at University of Oxford, Oxford, UK, from 6 - 14, July, 2024.
[Mar. 2024] 🗣️ I represented the Amgen Interns to attend the Mitacs Roundtable Discussion with new BC Minister of Post-Secondary Education and Future Skills, Lisa Beare.
[Dec. 2023] 🗣️ I will give a poster presentation at NeurIPS 2023 conference, Machine Learning in Structural Biology (MLSB) workshop, in New Orleans, USA.
[Nov. 2023] 🗣️ I will give an oral presentation at Machine Learning in Computational Biology (MLCB 2023) conference in Seattle, USA.
[Nov. 2023] 🎉 Our paper on visualizing hybridization trajectories with biophysics-informed deep graph embeddings is accepted at MLCB 2023 (PMLR), selected as an oral presentation.
[Oct. 2023] 🎉 Our paper on partial alignment of density maps and rigid body fitting using unbalanced Gromov-Wasserstein divergence is accepted at MLSB, NeurIPS 2023.
[Sept. 2023] 🗣️ I presented a poster at DNA29 conference at Tohoku University, Sendai, Japan.
[Sept. 2023] 🎉 I gave a talk at BIRS workshop in Banff, Canada.
[Jul. 2023] 🎉 Our paper on revisiting hybridization kinetics with Bayesian inference is accepted at DNA29.
[Jul. 2023] 🎉 Our abstract poster on visualizing DNA kinetics using deep graph embeddings is accepted at DNA29.
[Jun. 2023] 👨‍💻 I am starting a Machine Learning Scientist Internship at Amgen focused on employing deep learning in drug discovery.
[Jan. 2023] 🎉 Our paper on revisiting plasmonic properties of complex semiconductor nanocrystals using Magnetic Circular Dichroism Spectroscopy is accepted at The Journal of Physical Chemistry C.
[Dec. 2022] 🗣️ I presented a poster at NeurIPS 2022 conference in New Orleans, USA.
[Oct. 2022] 🎉 Our workshop paper on using semi-VAE for visualization of DNA reaction trajectories is accepted at MLSB, NeurIPS 2022.
[Oct 2022] ✍️ I PASSED my PhD qualification exam!
[Aug 2022] 🗣️ I attended DNA28 conference at University of New Mexico, Albuquerque, USA.
[Jul. 2022] 🎉 Our abstract poster on faster elementary steps in DNA reaction simulators is accepted at DNA28.
[Dec. 2021] 🎉 Our paper on the study of origin of d⁰ magnetism in transparent metal oxide nanocrystals is accepted at The Journal of Physical Chemistry C.
[Sept. 2021] 📕 I am beginning my doctoral journey in the Department of Computer Science at the University of British Columbia！
[Feb. 2021] 👨‍🎓 I obtained my MSc degree from the University of Waterloo！
[Sept. 2020] 🎉 Our paper on manipulating carrier polarization in semiconductor nanocrystals is accepted at ECS Transactions.
[Sept. 2018] 📕 I am beginning my MSc study at the University of Waterloo！
[Sept. 2015 - Aug. 2018] 👨‍🎓 I obtained a dual Bachelor's degree from the University of Waterloo (Dean’s Honours List) [BSc] and Beijing Jiaotong University [BEng] within 3 years!
[Sept. 2015] 🐣 Academic life starts!

Publications

THESIS

Applications of Deep Generative Models to DNA Reaction Kinetics and to Cryogenic Electron Microscopy

Chenwei Zhang

Ph.D. Thesis at the University of British Columbia (Ph.D. Thesis), 2026.

PDF BibTex
This dissertation explores how deep generative models can advance the analysis of challenging biological problems by integrating domain knowledge with cutting-edge deep learning techniques. It focuses on two fundamental areas: DNA reaction kinetics and cryogenic electron microscopy (cryo-EM). In the first part, we present ViDa, a biophysics-informed deep learning framework that leverages variational autoencoders (VAEs) and geometric scattering transform to generate biophysically-plausible embeddings of DNA reaction kinetics simulations. These embeddings are further reduced to a two-dimensional Euclidean space to visualize DNA hybridization and toehold-mediated three-way strand displacement reactions. By embedding simulated secondary structures and reaction trajectories into a low-dimensional representation, ViDa preserves structure and clusters trajectory ensembles into reaction pathways, making simulation results more interpretable and revealing new insights into reaction mechanisms. In the second part, we address key challenges in cryo-EM density map interpretation and protein structure modeling. We first provide a comprehensive review and benchmarking of state-of-the-art deep learning methods for protein structure modeling (i.e., atomic model building). We propose improved evaluation metrics to assess the performance of these methods and provide guidance for researchers. We then present Struc2mapGAN, a generative adversarial network (GAN) that synthesizes high-fidelity experimental-like cryo-EM density maps from protein structures. We finally present CryoSAMU, a structure-aware multimodal U-Net that enhances intermediate-resolution cryo-EM density maps by integrating density features with structural embeddings from protein large language models through cross-attention mechanisms. Overall, these contributions demonstrate the potential of deep generative models to interpret DNA reaction mechanisms and to advance cryo-EM density map analysis and protein structure modeling.

ICML

CryoSAMU: Enhancing 3D Cryo-EM Density Maps of Protein Structures at Intermediate Resolution with Structure-Aware Multimodal U-Nets

Chenwei Zhang, Khanh Dao Duc

Generative AI and Biology Workshop, ICML 2025.

PDF Code BibTex
Enhancing cryogenic electron microscopy (cryo-EM) 3D density maps at intermediate resolution (4-8 Å) is crucial in protein structure determination. Recent advances in deep learning have led to the development of automated approaches for enhancing experimental cryo-EM density maps. Yet, these methods are not optimized for intermediate-resolution maps and rely on map density features alone. To address this, we propose CryoSAMU, a novel method designed to enhance 3D cryo-EM density maps of protein structures using structure-aware multimodal U-Nets and trained on curated intermediate-resolution density maps. We comprehensively evaluate CryoSAMU across various metrics and demonstrate its competitive performance compared to state-of-the-art methods. Notably, CryoSAMU achieves significantly faster processing speed, showing promise for future practical applications. Our code is available at https://github.com/chenwei-zhang/CryoSAMU.

BA

Struc2mapGAN: improving synthetic cryogenic electron microscopy density maps with generative adversarial networks

Chenwei Zhang, Anne Condon, Khanh Dao Duc

Bioinformatics Advances, 2025.

PDF Code BibTex
Generating synthetic cryogenic electron microscopy 3D density maps from molecular structures has potential important applications in structural biology. Yet existing simulation-based methods cannot mimic all the complex features present in experimental maps, such as secondary structure elements. As an alternative, we propose struc2mapGAN, a novel data-driven method that employs a generative adversarial network to produce improved experimental-like density maps from molecular structures. More specifically, struc2mapGAN uses a nested U-Net architecture as the generator, with an additional L1 loss term and further processing of raw training experimental maps to enhance learning efficiency. While struc2mapGAN can promptly generate maps after training, we demonstrate that it outperforms existing simulation-based methods for a wide array of tested maps and across various evaluation metrics.

BIB

A comprehensive survey and benchmark of deep learning-based methods for atomic model building from cryo-electron microscopy density maps

Chenwei Zhang, Anne Condon, Khanh Dao Duc

Briefings in Bioinformatics, 2025.

PDF Code BibTex
Advancements in deep learning (DL) have recently led to new methods for automated construction of atomic models of proteins, from single-particle cryogenic electron microscopy (cryo-EM) density maps. We conduct a comprehensive survey of these methods, distinguishing between direct model building approaches that only use density maps, and indirect ones that integrate sequence-to-structure predictions from AlphaFold. To evaluate them with better precision, we refine standard existing metrics, and benchmark a subset of representative DL-methods against traditional physics-based approaches using 50 cryo-EM density maps at varying resolutions. Our findings demonstrate that overall, DL-based methods outperform traditional physics-based methods. Our benchmark also shows the benefit of integrating AlphaFold as it improved the completeness and accuracy of the model, although its dependency on available sequence information and limited training data may limit its usage.

PRX

Alignment of Partially Overlapping Cryo-EM Maps Using Unbalanced Gromov-Wasserstein Divergence

Aryan Tajmir Riahi, Chenwei Zhang, Anne Condon, James Chen, Khanh Dao Duc

PRX Life - Physical Review Journals, 2025.

PDF Code BibTex
Aligning EM density maps and fitting atomic models are essential steps in single-particle cryogenic electron microscopy (cryo-em), with recent methods leveraging various algorithms and machine learning tools. As aligning maps remains challenging in the presence of a map that only partially fits the other (e.g., one subunit), we here propose a procedure, em partial alignment with optimal transport (EMPOT), for partial alignment of three-dimensional (3D) maps. EMPOT finds a coupling between 3D point-cloud representations, which is associated with their so-called unbalanced Gromov Wasserstein divergence, and uses this coupling to find an optimal rigid body transformation. On running and benchmarking our method with experimental maps and structures, we show that EMPOT outperforms standard methods for aligning density maps and fitting atomic models to a density map, suggesting potential applications of partial optimal transport for improving cryo-em pipelines.

PMLR

ViDa: Visualizing DNA hybridization trajectories with biophysics-informed deep graph embeddings

Chenwei Zhang, Jordan Lovrod, Boyan Beronov, Khanh Dao Duc, Anne Condon

Machine Learning in Computational Biology. PMLR , 2023.

PDF Code BibTex Oral (Top 15%)
Visualization tools can help synthetic biologists and molecular programmers understand the complex reactive pathways of nucleic acid reactions, which can be designed for many potential applications and can be modelled using a continuous-time Markov chain (CTMC). Here we present ViDa, a new visualization approach for DNA reaction trajectories that uses a 2D embedding of the secondary structure state space underlying the CTMC model. To this end, we integrate a scattering transform of the secondary structure adjacency, a variational autoencoder, and a nonlinear dimensionality reduction method. We augment the training loss with domain-specific supervised terms that capture both thermodynamic and kinetic features. We assess ViDa on two well-studied DNA hybridization reactions. Our results demonstrate that the domain-specific features lead to significant quality improvements over the state-of-the-art in DNA state space visualization, successfully separating different folding pathways and thus providing useful insights into dominant reaction mechanisms.

NeurIPS

EMPOT: partial alignment of density maps and rigid body fitting using unbalanced Gromov-Wasserstein divergence

Aryan Tajmir Riahi, Chenwei Zhang, James Chen, Anne Condon, Khanh Dao Duc

Machine Learning for Structural Biology Workshop (MLSB), NeurIPS 2023.

PDF Code BibTex
Aligning EM density maps and fitting atomic models are essential steps in single particle cryogenic electron microscopy (cryo-EM), with recent methods leveraging various algorithms and machine learning tools. As aligning maps remains challenging in the presence of a map that only partially fits the other (e.g. one subunit), we here propose a new procedure, EMPOT (EM Partial alignment with Optimal Transport), for partial alignment of 3D maps. EMPOT first finds a coupling between 3D point-cloud representations, which is associated with their so-called unbalanced Gromov Wasserstein divergence, and second, uses this coupling to find an optimal rigid body transformation. Upon running and benchmarking our method with experimental maps and structures, we show that EMPOT outperforms standard methods for aligning subunits of a protein complex and fitting atomic models to a density map, suggesting potential applications of Partial Optimal Transport for improving Cryo-EM pipelines.

DNA29

Revisiting Hybridization Kinetics with Improved Elementary Step Simulation

Jordan Lovrod, Boyan Beronov, Chenwei Zhang, Erik Winfree, Anne Condon

29th International Conference on DNA Computing and Molecular Programming (DNA29), 2023.

PDF Code BibTex
Nucleic acid strands, which react by forming and breaking Watson-Crick base pairs, can be designed to form complex nanoscale structures or devices. Controlling such systems requires accurate predictions of the reaction rate and of the folding pathways of interacting strands. Simulators such as Multistrand model these kinetic properties using continuous-time Markov chains (CTMCs), whose states and transitions correspond to secondary structures and elementary base pair changes, respectively. The transient dynamics of a CTMC are determined by a kinetic model, which assigns transition rates to pairs of states, and the rate of a reaction can be estimated using the mean first passage time (MFPT) of its CTMC. However, use of Multistrand is limited by its slow runtime, particularly on rare events, and the quality of its rate predictions is compromised by a poorly-calibrated and simplistic kinetic model. The former limitation can be addressed by constructing truncated CTMCs, which only include a small subset of states and transitions, selected either manually or through simulation. As a first step to address the latter limitation, Bayesian posterior inference in an Arrhenius-type kinetic model was performed in earlier work, using a small experimental dataset of DNA reaction rates and a fixed set of manually truncated CTMCs, which we refer to as Assumed Pathway (AP) state spaces. In this work we extend this approach, by introducing a new prior model that is directly motivated by the physical meaning of the parameters and that is compatible with experimental measurements of elementary rates, and by using a larger dataset of 1105 reactions as well as larger truncated state spaces obtained from the recently introduced stochastic Pathway Elaboration (PE) method. We assess the quality of the resulting posterior distribution over kinetic parameters, as well as the quality of the posterior reaction rates predicted using AP and PE state spaces. Finally, we use the newly parameterised PE state spaces and Multistrand simulations to investigate the strong variation of helix hybridization reaction rates in a dataset of Hata et al. While we find strong evidence for the nucleation-zippering model of hybridization, in the classical sense that the rate-limiting phase is composed of elementary steps reaching a small nucleus of critical stability, the strongly sequence-dependent structure of the trajectory ensemble up to nucleation appears to be much richer than assumed in the model by Hata et al. In particular, rather than being dominated by the collision probability of nucleation sites, the trajectory segment between first binding and nucleation tends to visit numerous secondary structures involving misnucleation and hairpins, and has a sizeable effect on the probability of overcoming the nucleation barrier.

JPCC

Revisiting Plasmonic Properties of Complex Semiconductor Nanocrystals Using Magnetic Circular Dichroism Spectroscopy A Cautionary Tale

Aaron Kenny-Wilby, Gyorgy Jaics, Chenwei Zhang, Penghui Yin, Pavle V. Radovanovic

The Journal of Physical Chemistry C (J. Phys. Chem. C), 2023.

PDF BibTex
Plasmonic semiconductor nanocrystals (NCs) have emerged as an attractive alternative to noble metal nanoparticles. They typically exhibit localized surface plasmon resonance (LSPR) in the infrared range, which can be tuned by controlling the type and concentration of charge carriers. Compositionally and electronically complex semiconductor NCs have attracted significant attention due to their reported plasmonic properties, such as bipolar behavior (off-stoichiometric spinel-type gallium iron oxide or GFO), wide tunability (copper iron chalcogenide or Cu_xFe_yS_2-zSe_z), and extremely broad LSPR (oxygen-deficient WO_3-x). However, the origin of these unusual properties remains unclear. Here, we comparatively investigate the plasmonic properties of these semiconductor NC materials using magnetic circular dichroism (MCD) spectroscopy, owing to the highly specific response of LSPR to the magnetic field and light polarization. In all systems, we find indisputable evidence that the LSPR absorption bands have been mischaracterized. MCD signals that coincide with the absorption bands previously assigned to LSPR of GFO and Cu_xFe_yS_2-zSe_z NCs show rapid saturation with the magnetic field and independence on the sign of charge carriers, indicating that the corresponding absorption bands are not due to LSPR. This behavior is consistent with intra-ionic and inter-band transitions for GFO and Cu_xFe_yS_2-zSe_z NCs, respectively. The MCD signal corresponding to the absorption spectrum of WO_3-x NCs shows Brillouin function dependence on the magnetic field at high photon energies (attributed to d–d transitions of W5+) and linear dependence in the near-infrared range (attributed to LSPR), indicating that the broad absorption spectrum of WO_3-x NCs is only partly associated with LSPR. The results of this work help resolve notable discrepancies in the literature, underline the challenges in assigning absorption bands of complex semiconductor NCs to LSPR, and demonstrate that MCD spectroscopy is an invaluable tool for characterization of these materials.

NeurIPS

Visualizing DNA Rection Trajectories with Deep Graph Embedding Approaches

Chenwei Zhang, Khanh Dao Duc, Anne Condon

Machine Learning for Structural Biology Workshop (MLSB), NeurIPS 2022.

PDF Code BibTex
Synthetic biologists and molecular programmers design novel nucleic acid reactions, with many potential applications. Good visualization tools are needed to help domain experts make sense of the complex outputs of folding pathway simulations of such reactions. Here we present ViDa, a new approach for visualizing DNA reaction folding trajectories over the energy landscape of secondary structures. We integrate a deep graph embedding model with common dimensionality reduction approaches, to map high-dimensional data onto 2D Euclidean space. We assess ViDa on two well-studied and contrasting DNA hybridization reactions. Our preliminary results suggest that ViDa's visualization successfully separates trajectories with different folding mechanisms, thereby providing useful insight to users, and is a big improvement over the current state-of-the-art in DNA kinetics visualization.

JPCC

On the Origin of d⁰ Magnetism in Transparent Metal Oxide Nanocrystals

Chenwei Zhang, Penghui Yin, Whenhuan Lu, Victor Galievsky, Pavle V. Radovanovic

The Journal of Physical Chemistry C (J. Phys. Chem. C), 2021.

PDF BibTex
Controlling magnetic and magneto-optical properties of transparent metal oxide semiconductors has a significant potential for spintronics and photonics. Although ferromagnetism has been reported for several nanostructured transparent metal oxides in the absence of magnetic dopants, its origin and the nature of the exchange interactions remain controversial. Here, we report a variable-temperature–variable-field magnetic circular dichroism study of ZnO and SnO2 nanocrystals prepared under oxidizing and reducing conditions. We observe the band splitting in ZnO and SnO2 nanocrystals induced by localized doublet (S = 1/2) and triplet (S = 1) ground states, respectively. Photoluminescence measurements suggest that these states are associated with oxygen vacancies, either as isolated paramagnetic sites in ZnO or as local vacancy-based complexes in SnO2 nanocrystals. The results of this work demonstrate the ability to tune carrier polarization in metal oxide nanocrystals by in situ control of the native defect formation and attest to the anomalous Zeeman splitting of the band states, which may play an important role in generating ferromagnetism in this class of materials.

THESIS

Manipulating Carrier Polarization in Pure and Doped Metal Oxide Semiconductor Nanocrystals

Chenwei Zhang

M.Sc. Thesis at the University of Waterloo (M.Sc. Thesis), 2021.

PDF BibTex
Moore's Law will soon be inadequate to keep up with the needs of technological developments as the currently used materials have reached their natural performance limits. Hence, the development of a new technology to replace traditional semiconductor electronics is imminent. Spintronics has been emerged as a promising alternative to control charge properties as well as electron spin to improve the device performance. Metal oxide and chalcogenide semiconductor nanocrystals (NCs) have generated much research due to their unique properties and relatively low cost. In contrast to noble metal nanoparticles (NPs), the semiconductor NCs with tunable free charge carrier concentration are regarded as promising materials in technological applications including multifunctional quantum devices and next-generation electronic devices.

ECST

Manipulating Carrier Polarization in Semiconductor Nanocrystals

Chenwei Zhang, Penghui Yin, Pavle V. Radovanovic

ECS Transactions (ECST), 2020.

PDF BibTex
We investigated the role of the synthesis method and post-synthesis processing on the plasmonic properties of antimony-doped SnO2 nanocrystals. The nanocrystal samples having variable doping concentrations were prepared by coprecipitation and solvothermal methods, and subsequently thermally annealed for different durations. We found that solvothermally-synthesized Sb-doped SnO2 nanocrystals exhibit strong localized surface plasmon resonance in the near-infrared region, which is distinctly absent in the nanocrystals synthesized by the coprecipitation method. Upon thermal annealing, the plasmon absorption emerges in the nanocrystals prepared by the coprecipitation method, and increases in intensity in solvothermally-synthesized nanocrystals. Using X-ray photoelectron spectroscopy, we correlated the plasmon intensity to the oxidation of Sb3+ to Sb5+. These results demonstrate that synthesis methodology and post-synthesis treatment can dramatically influence the plasmonic properties of aliovalently doped semiconductor nanocrystals via dopant oxidation state, and can be effectively used to design semiconductor nanocrystals with targeted plasmonic properties.