Skip to content
View yangmei25's full-sized avatar

Block or report yangmei25

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
yangmei25/README.md

Yangmei Li, Ph.D.

AI/ML engineer and computational biologist building trustworthy AI systems for biomedical research—from AI agents and foundation models to rigorous evaluation and production-ready deployment.

Featured work

An auditable agent workflow for hereditary-cancer germline variant research that combines HGVS normalization, ClinVar, PubMed/PMC, gnomAD, gene-level context, optional TCGA cohort analysis, grounded synthesis, and human-review handoff.

  • Separates evidence for the exact variant, broader gene context, and optional somatic cohort context to prevent scope overclaiming.
  • Evaluates citation validity, scientific faithfulness, latency, cache behavior, token use, estimated cost, and recurring failure categories.
  • Uses deterministic validation, an independent Check Agent, and a persistent human-review queue rather than treating model output as scientific truth.
  • Delivered with a browser workbench, batch/history/export tools, 81 automated tests, 32/32 intake evaluation cases, and passing GitHub Actions CI.

An end-to-end protein language-model project spanning biological and classical baselines, frozen embeddings, multi-seed fine-tuning, homology-filtered external evaluation, subtype failure analysis, long-protein inference, and a tested FastAPI service.

  • Fine-tuned ESM-2 150M achieved 0.936 validation F1 and 0.982 ROC-AUC across three seeds.
  • External evaluation identified peripheral membrane proteins as the main failure mode.
  • Subtype-aware modeling improved peripheral-protein recall from 0.325 to 0.549.
  • Delivered with FastAPI, Docker, GitHub Actions, and 40 automated tests.

A patient- and sample-aware comparison of conventional Scanpy/PCA and frozen scGPT representations across 29,614 NSCLC cells from 10 metastatic lymph-node samples.

  • scGPT improved broad-cell sample-held-out kNN accuracy from 0.832 to 0.923.
  • The evaluation also found stronger platform separation and weaker performance for some fine-grained CD8 T-cell states.
  • Sample-level analysis compared cell composition, functional programs, and scGPT pooling without presenting the small cohort as a validated clinical predictor.

Technical focus

  • Bioinformatics: single-cell RNA-seq, microbial and viral genomics, metagenomics, genome assembly, comparative genomics, sequencing QC
  • AI and machine learning: PyTorch, Hugging Face Transformers, scGPT, ESM-2, LLM agents, LangGraph, grounded synthesis, evaluation and tracing, model benchmarking, error analysis
  • Scientific software: Python, R, Bash, Linux/HPC, Git, automated testing, GitHub Actions, scientific APIs, FastAPI, Docker

Current direction

I am focused on scientific AI and computational biology roles where biological domain knowledge, reproducible analysis, and rigorous evaluation are essential to building trustworthy systems.

LinkedIn · GitHub

Pinned Loading

  1. esm2-protein-localization-benchmark esm2-protein-localization-benchmark Public

    End-to-end ESM-2 protein localization benchmark with multi-seed fine-tuning, external validation, failure analysis, FastAPI, Docker, and CI

    Python 1

  2. nsclc-single-cell-foundation-model-benchmark nsclc-single-cell-foundation-model-benchmark Public

    Patient-aware benchmark of frozen scGPT versus Scanpy/PCA on 29,614 NSCLC single cells with held-out evaluation and sample-level analysis

    Python