miRNA Target Prediction with CLIP-Seq Data

ISEF Category: Cellular and Molecular Biology

Subcategory: Molecular Biology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A single miRNA can change how a cell behaves, but finding its true targets is hard. Many prediction tools guess from sequence alone, then miss what happens in real tissues. You can test whether a contrastive transformer trained on CLIP-seq data does better than older methods. That makes this a real bioinformatics project, not just a coding exercise.

What Is It?

MicroRNAs, or miRNAs, are short RNA molecules that help control gene expression. Think of them like dimmer switches. They do not turn genes fully on or off, but they can lower how much protein a gene makes. The hard part is figuring out which genes each miRNA actually regulates.

TargetScan is a classic prediction tool. It looks for sequence patterns that suggest a miRNA might bind to a gene. That works as a starting point, but biology is messier than a pattern match. Proteins, cell type, and RNA structure all change what really happens inside a tissue.

CLIP-seq data helps with that problem. CLIP stands for crosslinking and immunoprecipitation. These datasets show where RNA-binding proteins, including AGO proteins that work with miRNAs, touch RNA in real cells. A contrastive transformer can learn from those examples and compare true binding signals against false ones. Your project asks whether that newer model improves tissue-specific prediction over TargetScan.

Why This Is a Good Topic

This is a strong science fair topic because it gives you a real comparison to test. You have a clear baseline, TargetScan, and a clear upgrade, a transformer trained on public AGO-CLIP data. You can measure performance with standard metrics and focus on tissue-specific predictions, which connects to cancer, development, and gene regulation. You can also do the whole project with public data and coding tools, so you do not need a wet lab.

Research Questions

• How does a contrastive transformer trained on AGO-CLIP data change miRNA target prediction accuracy compared with TargetScan?
• What is the effect of training on tissue-specific CLIP-seq datasets versus mixed-tissue datasets on prediction performance?
• Does adding sequence context around the binding site improve precision for tissue-specific miRNA targets?
• To what extent do model gains differ across tissues with different levels of CLIP-seq coverage?
• Which feature type, sequence motif, evolutionary conservation, or CLIP signal, contributes most to correct target prediction?
• How does the model perform on held-out miRNA families that were not seen during training?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Stable internet access for downloading public datasets.
• Python installed with a local environment manager such as conda or venv.
• Spreadsheet software for tracking samples, labels, and results.
• External storage or cloud backup for large data files.
• Text editor or notebook interface for code and notes.
• Access to public CLIP-seq and miRNA annotation files from GEO, ENCODE, or similar repositories.
• PubMed account or browser access for reading review papers and methods.

Advanced Materials

• Linux workstation or university server access for larger model runs.
• GPU access for training and tuning the transformer.
• Python environment with PyTorch, scikit-learn, pandas, NumPy, and matplotlib.
• Genome annotation files and reference transcriptome downloads.
• Public AGO-CLIP datasets with matched tissue labels.
• Benchmark datasets for miRNA target validation.
• Version control system such as git for tracking model changes.
• Bioinformatics workflow tools for reproducible preprocessing.
• JupyterLab or similar notebook interface for analysis and visualization.

Software & Tools

• Python: Runs data cleaning, model training, and evaluation scripts for the prediction pipeline.
• PyTorch: Builds and trains the transformer model on sequence and CLIP features.
• scikit-learn: Calculates classification metrics and supports baseline comparisons.
• pandas: Organizes large biological datasets and sample metadata.
• matplotlib: Makes clear plots for ROC curves, precision-recall curves, and tissue comparisons.

Experiment Steps

1. Define the prediction task and decide what counts as a true miRNA target in your benchmark set.
2. Select public AGO-CLIP datasets and matching target labels, then plan how you will split tissues and samples.
3. Build a baseline pipeline that reproduces or approximates TargetScan-style predictions.
4. Design the transformer input format so the model can compare sequence context and CLIP signal.
5. Choose evaluation metrics that reward correct positives and penalize false positives across tissues.
6. Plan ablation tests that remove one feature type at a time so you can measure what the model really learned.

Common Pitfalls

• Using a benchmark set that overlaps too much with training data, which makes the model look better than it is.
• Mixing tissues with very different data quality, which hides whether the method truly improves tissue-specific prediction.
• Treating CLIP peaks as perfect ground truth, which can label noisy binding events as true targets.
• Comparing the new model only against a weak baseline, which does not prove real improvement over TargetScan-like methods.
• Forgetting to hold out entire miRNA families or tissues, which lets the model memorize patterns instead of learning general rules.

What Makes This Competitive

A competitive project here needs more than a model that scores well once. You want clean train, validation, and test splits that block data leakage. You also want a careful comparison against TargetScan and at least one other baseline, plus a few ablation tests that explain why your model works. Strong entries often add a tissue-specific analysis, a failure analysis, or a novel way to score predictions across different CLIP datasets.

Project Variations

• Focus on one tissue type, such as brain, liver, or breast, and test whether prediction quality changes with tissue-specific AGO-CLIP data.
• Compare sequence-only predictions with sequence plus RNA secondary structure features to see whether structure adds value.
• Test whether the model handles noncanonical miRNA binding sites better than a standard baseline.

Learn More

• PubMed: Search review articles on microRNA target prediction, AGO-CLIP, and transcriptome-wide binding methods.
• NCBI Gene and GEO: Find gene annotations and public CLIP-seq datasets for model building and validation.
• ENCODE Portal: Explore public RNA-binding protein and CLIP-based datasets with standardized metadata.
• MIT OpenCourseWare: Look for free courses in machine learning, computational biology, and genomics analysis.
• Nature Methods: Read methods papers on CLIP-seq, model benchmarking, and RNA-protein interaction analysis through your school or public library access.