Predict Fibroblast to Heart Cell Reprogramming

ISEF Category: Biomedical Engineering

Subcategory: Cell and Tissue Engineering  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Turning a skin cell into a heart cell sounds like sci-fi, but researchers already do pieces of it in the lab. The hard part is picking the right transcription factors, the proteins that switch genes on and off. A graph-neural network can help you score those combinations before anyone runs an expensive experiment. That makes this a strong in silico project with real biomedical stakes.

What Is It?

Fibroblasts are common connective-tissue cells. Cardiomyocytes are heart muscle cells. Trans-differentiation means pushing one mature cell type to become another without going back to a stem-cell state. In this project, you would build a machine-learning model that predicts which transcription-factor cocktails might push fibroblasts toward a cardiomyocyte-like state.

A graph-neural network is a model that learns from networks. Think of it like a smart map that learns which genes and proteins interact with each other. Human Cell Atlas data can give you cell-type gene-expression patterns, and STRING-DB can give you protein interaction links. You then compare your model’s scores with published reprogramming benchmarks, so you can see whether the model ranks known good cocktails above weaker ones.

Why This Is a Good Topic

This is a good science fair topic because you can test a clear prediction without a wet lab. You can measure model accuracy, ranking quality, and how well different biological feature sets improve performance. The project connects to heart disease, regenerative medicine, and faster drug discovery. You can also show real research skills, like cleaning data, building a baseline model, and validating against published results.

Research Questions

• How does adding protein interaction data from STRING-DB change the model’s ability to rank known cardiomyocyte reprogramming cocktails?
• What is the effect of using Human Cell Atlas cell-type signatures instead of generic gene lists on prediction accuracy?
• Does a graph-neural-network model outperform a simpler baseline model for predicting transcription-factor cocktail success?
• To what extent do the model’s top-ranked cocktails match published reprogramming-efficiency benchmarks?
• Which gene-network features most strongly influence the model’s score for fibroblast to cardiomyocyte trans-differentiation?
• How does removing low-confidence interactions from the network affect ranking stability?
• What is the effect of training on one benchmark study and testing on another benchmark set?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Stable internet access.
• Spreadsheet software for tracking samples and results.
• Python installed with Jupyter Notebook.
• Free account access to public datasets, including Human Cell Atlas resources and STRING-DB.
• Reference manager or notes app for saving paper citations.
• Digital notebook for experiment logs.

Advanced Materials

• High-memory workstation or university server access.
• Python environment with PyTorch or TensorFlow.
• Graph machine-learning library such as PyTorch Geometric or DGL.
• Local copy of Human Cell Atlas data relevant to fibroblasts and cardiomyocytes.
• STRING-DB interaction exports.
• Published benchmark dataset of transcription-factor reprogramming outcomes.
• Version control system such as Git.

Software & Tools

• Python: Runs data cleaning, model training, and evaluation scripts for the project.
• Jupyter Notebook: Lets you document code, figures, and results in one place.
• PubMed: Helps you find review articles and benchmark studies on cardiac reprogramming.
• STRING-DB: Provides protein-protein interaction data for building the gene network.
• Human Cell Atlas portal: Gives cell-type expression data you can use to define fibroblast and cardiomyocyte signatures.
• Graphviz: Helps you visualize the interaction network and inspect model inputs.

Experiment Steps

1. Define the exact prediction task, including what counts as a successful reprogramming cocktail and which published benchmarks you will score against.
2. Assemble your input features from Human Cell Atlas signatures, STRING-DB interactions, and any curated transcription-factor lists.
3. Build a simple baseline first, then build the graph-neural-network model so you can compare them fairly.
4. Decide how you will split training and test data to avoid leakage between similar cocktails or related papers.
5. Plan your evaluation metrics, such as rank correlation, top-k hit rate, and calibration of predicted scores.
6. Prepare a validation set from published studies, then check whether the model elevates known successful cocktails above weaker ones.

Common Pitfalls

• Mixing up fibroblast and cardiomyocyte labels across datasets, which can flip the biological meaning of the model.
• Training and testing on overlapping benchmark papers, which makes the accuracy look much better than it really is.
• Using raw gene symbols without normalizing aliases, which creates duplicate nodes and missing matches in the network.
• Treating STRING-DB links as all equal, which can hide the difference between weak and high-confidence interactions.
• Reporting only classification accuracy, which can miss whether the model actually ranks the best transcription-factor cocktails near the top.

What Makes This Competitive

A class-level version of this project just trains a model and reports accuracy. A stronger version compares multiple feature sets, checks ranking quality, and tests whether the model generalizes across different published benchmarks. You can also add careful error analysis, like asking which transcription factors the model keeps promoting and whether those choices match biology. That kind of analysis shows real judgment, not just code running.

Project Variations

• Swap the target cell type and predict fibroblast to neuron reprogramming instead of cardiomyocytes.
• Compare a graph-neural-network model with a random forest, logistic regression, or gradient boosting baseline.
• Use single-cell RNA-seq signatures from a different public atlas and test whether the ranking changes across datasets.

Learn More

• Human Cell Atlas: Search the public atlas portal for fibroblast and cardiomyocyte single-cell expression data and cell-type markers.
• STRING-DB: Search the STRING database site for protein interaction networks and confidence scores.
• PubMed: Search for review articles on direct cardiac reprogramming and transcription-factor cocktails.
• NIH NCBI Gene Expression Omnibus: Find published expression datasets for fibroblasts, cardiomyocytes, and reprogramming studies.
• MIT OpenCourseWare: Search for free courses on machine learning, bioinformatics, and graph models that can help with the coding side.