Inspiration

I wanted to learn biology by working with real research data. But getting started with single-cell research meant navigating unfamiliar terminology, public datasets, scientific tools, and AI models—all at once.

I built CellCast AI to make that starting point more approachable: a workspace where people can explore real cells, ask questions about gene expression, and learn what the results mean while using the tools.

What it does

CellCast brings public single-cell datasets, measured expression comparisons, and browser-based AI inference into one research workspace.

Users can:

  • Explore published plant and human studies from the EMBL-EBI Single Cell Expression Atlas.
  • Inspect cell landscapes and individual cell records.
  • Compare gene expression between populations or observed control and treatment groups.
  • Run scPlantLLM cell embeddings and scGPT expression reconstruction locally in their browser.
  • Reopen saved comparisons and export results.

A nine-step interactive tour highlights the actual controls and explains the biology behind each action. Users learn by conducting an investigation, rather than reading a separate tutorial.

How I built it

I built the application with Next.js, React, TypeScript, and Bun, using shadcn/ui for interface components, deck.gl for cell visualization, Recharts for expression summaries, and TanStack Virtual for large tables.

Public dataset requests pass through constrained Next.js routes. AI inference runs in a browser worker through ONNX Runtime Web, using WASM by default with optional WebGPU support.

I adapted pretrained model inference paths into browser-compatible ONNX exports and compared their outputs against PyTorch references. Deployment downloads pinned model artifacts and verifies their SHA-256 hashes before building.

I used Pentool to create the editable logo, banner, and supporting graphics, and Reactour to build the educational walkthrough. The project is open source on GitHub and deployed on Vercel.

What I learned

The biggest lesson was that scientific software needs to explain both its results and their limits.

Gene expression measures detected RNA; it does not directly measure protein activity. Nearby cells in an embedding are not automatically biologically equivalent. A difference between group means does not establish statistical significance or causation.

I also learned that numerical agreement between a browser model and its original implementation verifies the conversion—not the model’s biological accuracy.

Those distinctions shaped CellCast’s interface, documentation, and guided lessons.

Challenges

Keeping research claims honest. Observed treatment responses and predictions of unseen interventions are different capabilities. CellCast keeps them separate: its current models do not predict unmeasured knockout or drug responses.

Bringing models into the browser. Large weights, preprocessing, model architecture details, and runtime differences required careful handling. I used shortened, measured gene panels for exploratory runs and checked numerical consistency against reference outputs.

Working with live scientific data. Studies differ in identifiers, metadata, available conditions, and size. The application needs clear loading states, source links, and recoverable errors instead of assuming every dataset supports every analysis.

Teaching inside a working interface. The tour had to follow navigation, wait for controls to appear, support real interactions, and remain readable on mobile. It also needed to let users pause, skip, and return without losing their research.

What’s next

I want to expand the guided investigations, improve dataset compatibility, and make model preparation more representative of established scientific workflows.

The goal is to help more people move from “I don’t understand this dataset” to “I can ask a clear question, inspect the evidence, and explain what I still don’t know.”

Built With

  • bun
  • deck.gl
  • next.js
  • onnx
  • onnx-runtime-web
  • python
  • pytorch
  • radix-ui
  • react
  • reactour
  • recharts
  • scgpt
  • scplantllm
  • shadcn/ui
  • tailwind-css
  • tanstack-virtual
  • typescript
  • webassembly
  • webgpu
Share this project:

Updates

Submission history