Rajan Gyawali

Rajan Gyawali

Ph.D. candidate in Computer Science, University of Missouri

I develop machine learning methods for scientific data, where signals are faint and trustworthy labels are expensive. My research asks how to combine complementary sources of evidence, how to adapt pretrained foundation models to low-signal data, and how to learn when supervision is scarce or unreliable. Most of this work is in cryo-electron microscopy (cryo-EM), where these problems limit how quickly protein structures can be determined. I work in the Bioinformatics and Machine Learning Lab, advised by Dr. Jianlin Cheng.

I have built MICA, which combines cryo-EM maps with AlphaFold3 predictions to build atomic models of protein complexes; particle pickers built on foundation models; and CryoPPP and CryoVirusDB, expert-labeled benchmarks for particle picking, all with open code and data. As an Applied Scientist II Intern at Amazon in 2026, I built a training-free system in which LLM agents improve through adversarial self-play and detect real-world scam and phishing pages without retraining.

I expect to complete my Ph.D. in May 2027 and am on the job market.

Email CV Google Scholar GitHub LinkedIn
Portrait of Rajan Gyawali
A simulated cryo-EM micrograph. Protein particles are faint blobs buried in noise; a picker finds them one by one.

News

Jun 2026
Joined Amazon in San Diego as an Applied Scientist II Intern, working on self-evolving multi-agent LLM systems for scam and phishing detection.
May 2026
Received the Outstanding Doctoral Student Award, EECS Department, University of Missouri.
May 2026
CryoFSL published in Briefings in Bioinformatics.
Apr 2026
Book chapter “AI tools for cryo-EM” published in Progress in Molecular Biology and Translational Science.
Feb 2026
CryoVirusDB published in Viruses.
Oct 2025
MICA published in Communications Chemistry.

Research

My research follows four connected themes: multimodal learning and the adaptation of foundation models to low-signal scientific data; learning from scarce and unreliable supervision; benchmarks and evaluation that let a field measure its progress; and LLM agents that improve themselves without retraining. The projects below put each theme into practice, from atomic models of protein complexes to phishing detection.

Multimodal protein structure determination

MICA, Communications Chemistry, 2025

MICA fuses cryo-EM density maps with AlphaFold3 predictions in a multi-task encoder–decoder that predicts backbone atoms, Cα atoms, and amino acid types, then refines them into an atomic model of the protein complex.

MICA reaches an average TM-score of 0.93 on recently released high-resolution cryo-EM maps and significantly outperforms other state-of-the-art deep learning methods in accuracy and completeness, while staying robust to protein size and map resolution.

PaperCodeData

Self-evolving LLM agents for phishing detection

Applied Scientist II Intern, Amazon, 2026

In this training-free framework, an attacker agent synthesizes new phishing pages and URL attack patterns, an oracle certifies their labels, a detector classifies them, and a reflector distills reusable detection strategies, so the system improves through adversarial self-play.

The learned strategies transferred to real-world scam URLs and phishing pages without retraining, outperforming existing baselines on multiple public and internal benchmarks while keeping false-positive rates low.

Foundation models for cryo-EM particle picking

CryoSegNet, CryoTransformer, CryoFSL, 2024–2026

CryoSegNet couples Meta’s Segment Anything Model with an attention-gated U-Net, CryoTransformer uses a ResNet and transformer detector, and CryoFSL adapts foundation models with few-shot learning.

CryoSegNet surpasses the widely used pickers crYOLO and Topaz, CryoTransformer outperformed state-of-the-art methods as the first picker trained on CryoPPP, and CryoFSL picks robustly from only a handful of annotated micrographs.

CryoSegNetPaperCodeCryoTransformerPaperCodeCryoFSLPaperCode

Benchmarks for machine learning in cryo-EM

CryoPPP, Scientific Data, 2023; CryoVirusDB, Viruses, 2026

CryoPPP and CryoVirusDB are expert-curated datasets for training and testing particle pickers, with CryoPPP’s labels validated by 2D particle class validation and 3D density map validation against the gold standard.

CryoPPP contains 2.6 TB and 9,893 micrographs from 34 EMPIAR protein datasets, and CryoVirusDB adds 9,941 micrographs and 339,398 labeled virus particles from nine EMPIAR datasets.

CryoPPPPaperCodeDataCryoVirusDBPaperCodeData

Publications

* equal contribution

2026

  1. CryoFSL: a few-shot learning framework for robust protein particle picking in cryo-EM micrographs

    Biplab Poudel, Rajan Gyawali, Ashwin Dhakal, Jianlin Cheng, D. Xu

    Briefings in Bioinformatics, 27(3), bbag285, 2026

    PaperCodeJournal article

  2. CryoVirusDB: An Annotated Dataset for AI-Based Virus Particle Identification in Cryo-EM Micrographs

    Rajan Gyawali, Ashwin Dhakal, Liguo Wang, Jianlin Cheng

    Viruses, 18(2), 224, 2026

    PaperCodeDataJournal article, dataset

  3. AI tools for cryo-EM: Protein particle picking, density map enhancement, and atomic model building

    Ashwin Dhakal*, Rajan Gyawali*, J. Selvaraj*, Jianlin Cheng

    Progress in Molecular Biology and Translational Science, 223, 71–113, 2026

    Book chapter

2025

  1. Multimodal deep learning integration of cryo-EM and AlphaFold3 for high-accuracy protein structure determination

    Rajan Gyawali, Ashwin Dhakal, Jianlin Cheng

    Communications Chemistry, 8(1), 320, 2025

    PaperCodeDataJournal article

  2. Artificial intelligence in cryo-EM protein particle picking: recent advances and remaining challenges

    Ashwin Dhakal*, Rajan Gyawali*, Liguo Wang, Jianlin Cheng

    Briefings in Bioinformatics, 26(1), bbaf011, 2025

    PaperReview article

2024

  1. CryoSegNet: accurate cryo-EM protein particle picking by integrating the foundational AI image segmentation model and attention-gated U-Net

    Rajan Gyawali, Ashwin Dhakal, Liguo Wang, Jianlin Cheng

    Briefings in Bioinformatics, 25(4), bbae282, 2024

    PaperCodeJournal article

  2. CryoTransformer: a transformer model for picking protein particles from cryo-EM micrographs

    Ashwin Dhakal, Rajan Gyawali, Liguo Wang, Jianlin Cheng

    Bioinformatics, 40(3), btae109, 2024

    PaperCodeJournal article

  3. Adapting Segment Anything Model (SAM) through prompt-based learning for enhanced protein identification in cryo-EM micrographs

    F. He, Z. Yang, M. Gao, B. Poudel, N. S. E. S. Dhas, R. Gyawali, A. Dhakal, et al.

    IEEE International Conference on Medical Artificial Intelligence (MedAI), 9–20, 2024

    PaperConference paper

2023

  1. A large expert-curated cryo-EM image dataset for machine learning protein particle picking

    Ashwin Dhakal*, Rajan Gyawali*, Liguo Wang, Jianlin Cheng

    Scientific Data, 10(1), 392, 2023

    PaperCodeDataJournal article, dataset

  2. Predicting protein–ligand binding structure using E(n) equivariant graph neural networks

    Ashwin Dhakal*, Rajan Gyawali*, Jianlin Cheng

    bioRxiv, 2023.08.06.552202, 2023

    PaperPreprint

2019

  1. An Approach for the Employee Face Recognition by RPN and Faster R-CNN Techniques

    Rajan Gyawali, Dibakar Raj Pant

    Proceedings of IOE Graduate Conference, 2019-Summer, 6, 231–237, 2019

    PaperConference paper

Software and datasets

MICA

Automated atomic model building from cryo-EM density maps and AlphaFold3 predictions.

CodeData

CryoSegNet

Protein particle picking with the Segment Anything Model and an attention-gated U-Net, with output for RELION and CryoSPARC.

Code

CryoTransformer

Particle picking with a ResNet and transformer detector trained on CryoPPP.

Code

CryoPPP

Expert-curated benchmark of 9,893 labeled micrographs from 34 EMPIAR protein datasets.

CodeData

CryoVirusDB

Labeled virus particles in 9,941 cryo-EM micrographs for training and testing virus particle pickers.

CodeData

Experience

Jun – Aug 2026

Applied Scientist II Intern

Amazon, San Diego, CA

Built a training-free, self-evolving framework for URL-based scam and phishing detection, driven by adversarial self-play among four LLM agents: an attacker that synthesizes novel phishing pages and URL attack patterns, an oracle that certifies ground-truth labels, a detector, and a reflector that distills reusable detection strategies. Strategies learned from synthetic attacks transferred to real-world scam URLs and phishing pages, outperforming existing baselines on public and internal benchmarks without retraining while keeping false-positive rates low.

Aug 2022 – present

Graduate Research Assistant

Bioinformatics and Machine Learning Lab, University of Missouri, Columbia, MO

Deep learning for cryo-EM and protein structure: multimodal atomic model building (MICA), foundation-model-based particle picking (CryoSegNet), and benchmark datasets (CryoPPP, CryoVirusDB), plus contributions to transformer-based detection, few-shot learning, macromolecule identification in cryo-ET tomograms, and protein–ligand binding prediction with E(n)-equivariant graph neural networks. Trains models with PyTorch DDP on multi-node, multi-GPU clusters (Slurm, LSF).

May 2018 – Jun 2022

Telecommunications Engineer

Nepal Telecom, Kathmandu, Nepal

Built the Wireless KPI Analytics Tool, a Django application with REST APIs that automated analysis across 5,000+ network elements and cut analysis time from days to minutes. Designed Python and SQL pipelines for 10+ GB of daily network data, and LSTM models for anomaly detection and forecasting that improved network reliability by 15%.

Nov 2016 – May 2022

Lecturer

Himalaya College of Engineering, Tribhuvan University, Kathmandu, Nepal

Promoted from Assistant Lecturer to Lecturer within a year, then continued as Adjunct Lecturer alongside full-time engineering work. See Teaching and mentoring.

Teaching and mentoring

2026 – present

Undergraduate research mentor

Bioinformatics and Machine Learning Lab, University of Missouri

Mentors undergraduate researchers in deep learning–based image segmentation.

2016 – 2022

Instructor of record

Himalaya College of Engineering, Tribhuvan University, Kathmandu, Nepal

Instructor of record for undergraduate engineering courses, with 36–48 students each. Designed Python training and supervised final-year projects.

Education

2022 – May 2027

Ph.D. in Computer Science

University of Missouri, Columbia, MO

Advisor: Dr. Jianlin Cheng. Focus: deep learning, computational biology and bioinformatics. GPA 4.0. Expected May 2027.

2017 – 2019

M.Sc. in Electronics and Communication Engineering

Tribhuvan University, Kathmandu, Nepal

Specialization in information and communication, graduated with distinction. Thesis: employee face recognition with region proposal networks and Faster R-CNN, 96.0% accuracy on the ChokePoint dataset.

2012 – 2016

B.E. in Electronics and Communication Engineering

Himalaya College of Engineering, Tribhuvan University, Kathmandu, Nepal

Graduated with distinction. Final project: a brain-controlled wheelchair using a NeuroSky MindWave brain–computer interface.

Honors and awards

2026
Outstanding Doctoral Student Award, EECS Department, University of Missouri
2024
EECS Excellence Travel Grant, University of Missouri
2022
Dean's Fellowship, University of Missouri
2018
First rank, engineer recruitment examination, Public Service Commission, Nepal
2017
Graduate Scholarship, Tribhuvan University
2012
Undergraduate Scholarship, Tribhuvan University
Golden Jubilee Scholarship for undergraduate study, Embassy of India, Kathmandu

Talks

2024
CryoSegNet: AI-based method for protein particle picking from cryo-EM micrographs. Invited talk, Cryo-EM Super Group, University of Missouri.
2024
A large expert-curated cryo-EM image dataset for machine learning protein particle picking. Poster, RECOMB 2024, Cambridge, MA.

Service and leadership