feat: add functional tracks pipeline notebook

.gitattributes

@@ -36,3 +36,11 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 assets/paper_summary.jpg filter=lfs diff=lfs merge=lfs -text
 assets/paper_summary.png filter=lfs diff=lfs merge=lfs -text
 assets/output_tracks.png filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/K562_DNAse.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/K562_H3k4me3.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/K562_RNA_seq.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/HepG2_CTCF.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/HepG2_DNAse.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/HepG2_H3k4me3.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/HepG2_RNA_seq.bw filter=lfs diff=lfs merge=lfs -text
+notebooks_pipelines/bigwig_outputs/K562_CTCF.bw filter=lfs diff=lfs merge=lfs -text

README.md

@@ -3,7 +3,7 @@ title: NTv3 — Foundation Models for Long-Range Genomics
 emoji: 🧬
 colorFrom: indigo
 colorTo: blue
-sdk: gradio
+sdk: static
 pinned: false
 ---
 

app.py

@@ -1,11 +0,0 @@
-"""
-Main Gradio app entry point for NTv3 Space.
-This file imports the track prediction demo from app_tracks.py.
-"""
-from app_tracks import demo_interface
-
-# For Hugging Face Spaces with Gradio SDK, the 'demo' variable must be named 'demo'
-demo = demo_interface
-
-if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", share=False)

app_tracks.py

@@ -1,158 +0,0 @@
-"""
-Gradio app for NTv3 track prediction demo.
-This module contains the interactive track prediction interface.
-"""
-import gradio as gr
-import torch
-from transformers import pipeline
-import os
-
-# Initialize the pipeline (will be loaded on first use)
-ntv3_tracks = None
-
-def load_pipeline():
-    """Load the pipeline on first use (lazy loading)."""
-    global ntv3_tracks
-    if ntv3_tracks is None:
-        model_name = "InstaDeepAI/NTv3_650M_pos"
-        ntv3_tracks = pipeline(
-            "ntv3-tracks",
-            model=model_name,
-            trust_remote_code=True,
-            device=0 if torch.cuda.is_available() else -1,
-        )
-    return ntv3_tracks
-
-def predict_tracks(chrom, start, end, species):
-    """Run track prediction on the specified genomic region."""
-    try:
-        # Validate inputs
-        if not chrom or not start or not end or not species:
-            return "❌ Please fill in all fields."
-        
-        start = int(start)
-        end = int(end)
-        
-        if start >= end:
-            return "❌ Start position must be less than end position."
-        
-        if end - start > 1_000_000:
-            return "❌ Region size cannot exceed 1 Mb (1,000,000 bp)."
-        
-        # Load pipeline
-        pipe = load_pipeline()
-        
-        # Run prediction
-        out = pipe({
-            "chrom": chrom,
-            "start": start,
-            "end": end,
-            "species": species.lower()
-        })
-        
-        # Format output
-        result = f"""✅ Prediction completed successfully!
-
-📊 Output Shapes:
-• BigWig tracks logits: {tuple(out.bigwig_tracks_logits.shape)}
-  → {out.bigwig_tracks_logits.shape[1]} functional tracks over the center region
-
-• BED tracks logits: {tuple(out.bed_tracks_logits.shape)}
-  → {out.bed_tracks_logits.shape[1]} genomic elements over the center region
-
-• Language model logits: {tuple(out.mlm_logits.shape)}
-  → MLM predictions for the entire sequence
-
-📝 Note: Predictions are made over 37.5% of the center region of the input sequence.
-"""
-        return result
-        
-    except Exception as e:
-        return f"❌ Error: {str(e)}"
-
-# Create the track prediction demo interface (embedded in HTML)
-def create_demo_interface():
-    """Create the Gradio interface for track prediction."""
-    with gr.Blocks(title="NTv3 Track Prediction Demo", theme=gr.themes.Soft()) as demo_interface:
-        gr.Markdown("""
-        # 🧬 NTv3 Interactive Track Prediction Demo
-        
-        This demo allows you to run the NTv3 650M post-trained model to predict functional tracks and genomic elements for any genomic region.
-        
-        **Model:** `InstaDeepAI/NTv3_650M_pos`
-        """)
-        
-        with gr.Row():
-            with gr.Column():
-                chrom = gr.Textbox(
-                    label="Chromosome",
-                    placeholder="e.g., chr19",
-                    value="chr19",
-                    info="Chromosome name (e.g., chr1, chr19)"
-                )
-                start = gr.Number(
-                    label="Start Position",
-                    placeholder="e.g., 6700000",
-                    value=6_700_000,
-                    info="Start position in base pairs"
-                )
-                end = gr.Number(
-                    label="End Position",
-                    placeholder="e.g., 6831072",
-                    value=6_831_072,
-                    info="End position in base pairs"
-                )
-                species = gr.Dropdown(
-                    label="Species",
-                    choices=[
-                        "human", "mouse", "rat", "chicken", "zebrafish",
-                        "fruitfly", "worm", "yeast", "arabidopsis", "rice",
-                        "maize", "soybean", "tomato", "potato", "grape",
-                        "poplar", "medicago", "lotus", "brachypodium", "sorghum",
-                        "barley", "wheat", "oats", "rye"
-                    ],
-                    value="human",
-                    info="Select the species (24 supported species)"
-                )
-                predict_btn = gr.Button("🚀 Run Prediction", variant="primary")
-            
-            with gr.Column():
-                output = gr.Textbox(
-                    label="Results",
-                    lines=15,
-                    interactive=False,
-                    placeholder="Results will appear here after running prediction..."
-                )
-        
-        gr.Markdown("""
-        ### 📝 Notes:
-        - The model predicts ~7k functional tracks and 21 genomic elements
-        - Predictions are made over 37.5% of the center region of the input sequence
-        - Maximum region size: 1 Mb (1,000,000 base pairs)
-        - First run may take longer as the model loads
-        """)
-        
-        predict_btn.click(
-            fn=predict_tracks,
-            inputs=[chrom, start, end, species],
-            outputs=output
-        )
-        
-        gr.Examples(
-            examples=[
-                ["chr19", 6_700_000, 6_831_072, "human"],
-                ["chr1", 100_000, 200_000, "human"],
-                ["chr2", 50_000, 150_000, "mouse"],
-            ],
-            inputs=[chrom, start, end, species]
-        )
-    
-    return demo_interface
-
-# Create the demo interface
-demo_interface = create_demo_interface()
-
-# If running this file directly (for local testing)
-if __name__ == "__main__":
-    demo_interface.launch(server_name="0.0.0.0", share=False)
-

index.html

@@ -273,7 +273,8 @@
     <!-- Tab Navigation -->
     <div class="tabs">
       <button class="tab-button active" data-tab="home">🏠 Home</button>
-      <button class="tab-button" data-tab="demo">💻 Code Demo</button>
+      <button class="tab-button" data-tab="functional_tracks">💻 Code Demo</button>
+      <button class="tab-button" data-tab="annotation">🧬 Genome Annotation</button>
     </div>
 
     <!-- Home Tab (Content loaded from tabs/home.html) -->
@@ -281,10 +282,15 @@
       <!-- Content will be loaded dynamically -->
     </div>
 
-    <!-- Code Demo Tab (Content loaded from tabs/demo.html) -->
-    <div id="demo" class="tab-content">
+    <!-- Functional Tracks Tab (Content loaded from tabs/functional_tracks.html) -->
+    <div id="functional_tracks" class="tab-content">
       <!-- Content will be loaded dynamically -->
-        </div>
+    </div>
+
+    <!-- Genome Annotation Tab (Content loaded from tabs/annotation.html) -->
+    <div id="annotation" class="tab-content">
+      <!-- Content will be loaded dynamically -->
+    </div>
           
     <!-- <div class="paper-summary">
 		<h2>📄 A foundational model for joint sequence-function multi-species modeling at scale for long-range genomic prediction</h2>
@@ -301,7 +307,8 @@
     // Tab content mapping
     const tabFiles = {
       'home': 'tabs/home.html',
-      'demo': 'tabs/demo.html'
+      'functional_tracks': 'tabs/functional_tracks.html',
+      'annotation': 'tabs/annotation.html'
     };
 
     // Cache for loaded tab content

notebooks_pipelines/02_genome_annotation.ipynb

@@ -11,7 +11,7 @@
     "\n",
     "The pipeline abstracts away all the underlying steps: running inference with the model, retrieving and processing the predicted probabilities, and applying the HMM to generate a consistent annotation. It returns a ready-to-use GFF file that can be visualized in any genome browser for the sequence of interest.\n",
     "\n",
-    "If you’re interested in exploring the intermediate probabilities, please refer to the [track-prediction notebook](https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks/01_tracks_prediction.ipynb). These probabilities can be useful for assessing model confidence and identifying potentially interesting biological regions. This notebook focuses on the higher-level task of producing gene annotations directly from raw DNA.\n",
+    "If you’re interested in exploring the intermediate probabilities, please refer to the [track-prediction notebook](https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_tutorials/01_tracks_prediction.ipynb). These probabilities can be useful for assessing model confidence and identifying potentially interesting biological regions. This notebook focuses on the higher-level task of producing gene annotations directly from raw DNA.\n",
     "\n",
     "> 📝 **Note for Google Colab users:** This notebook is compatible with Colab! For faster inference, make sure to enable GPU: Runtime → Change runtime type → GPU (T4 or better recommended)."
    ]
@@ -184,7 +184,7 @@
    "id": "190ff65e",
    "metadata": {},
    "source": [
-    "## 4) 📁 Save a GFF file"
+    "## 4) 📁 Save as GFF file"
    ]
   },
   {
@@ -266,7 +266,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 8,
+   "execution_count": null,
    "id": "0904a5cb",
    "metadata": {},
    "outputs": [
@@ -325,7 +325,7 @@
    ],
    "source": [
     "config = {\n",
-    "    \"genome\": \"hg38\",  # built-in hg38\n",
+    "    \"genome\": assembly,\n",
     "    \"locus\": f\"{chrom}:{start}-{end}\",\n",
     "}\n",
     "\n",

notebooks_pipelines/bigwig_outputs/HepG2_CTCF.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa369433dcd408740311d0a0c5209fbb8889402a82fdad222cc11413dcaf1f1a
+size 380493

notebooks_pipelines/bigwig_outputs/HepG2_DNAse.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a59a1a90a9337b0862c27f4bc492e72f7d78ad0d3ffa5fdf7a981faca8d55cc
+size 387370

notebooks_pipelines/bigwig_outputs/HepG2_H3k4me3.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:349d8edea3908f828dbb1946d9ab16f7220b61a5922d6a7b75ac4fa55b5f359a
+size 381439

notebooks_pipelines/bigwig_outputs/HepG2_RNA_seq.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a116151b3b1550916f49093ca128326338911ee070481480e994a0baa1b00d4f
+size 381391

notebooks_pipelines/bigwig_outputs/K562_CTCF.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf84025d20d7ec59efbf02e7f0d20d67e3bb9564ffe3538e2397c0a46a576aea
+size 379394

notebooks_pipelines/bigwig_outputs/K562_DNAse.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8333fc13cfabe984cc303c575c2e65da20908240f8c733658b5b309a4191cb07
+size 381686

notebooks_pipelines/bigwig_outputs/K562_H3k4me3.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:074971cc028eab364a6cb7642bcc0cf70603de6f3f3f425c600b3b6f90699f32
+size 383184

notebooks_pipelines/bigwig_outputs/K562_RNA_seq.bw

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e4118a761250629de70de55c935b38f53a0c0eb6dc2156dcb632335c1ef7f42
+size 380637

requirements.txt

@@ -1,7 +0,0 @@
-gradio>=4.0.0
-torch>=2.0.0
-transformers>=4.55.0
-accelerate>=0.20.0
-safetensors>=0.3.0
-huggingface_hub>=0.23.0
-

tabs/annotation.html

@@ -0,0 +1,134 @@
+<div class="summary">
+  <h2>🧬 NTv3 Post-Trained Genome Annotation</h2>
+  <p>This notebook demonstrates how to use the NTv3 post-trained model to perform genome annotation directly from a DNA sequence. It relies on a pipeline that applies a Hidden Markov Model (HMM) to the per-base probabilities returned by NTv3, converting them into a coherent gene model that respects biological constraints and valid transitions between genomic elements.</p>
+  <p>The pipeline abstracts away all the underlying steps: running inference with the model, retrieving and processing the predicted probabilities, and applying the HMM to generate a consistent annotation. It returns a ready-to-use GFF file that can be visualized in any genome browser for the sequence of interest.</p>
+  <p>If you're interested in exploring the intermediate probabilities, please refer to the <a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks/01_tracks_prediction.ipynb" target="_blank" rel="noopener">track-prediction notebook</a>. These probabilities can be useful for assessing model confidence and identifying potentially interesting biological regions. This notebook focuses on the higher-level task of producing gene annotations directly from raw DNA.</p>
+  <p><strong>📝 Note for Google Colab users:</strong> This notebook is compatible with Colab! For faster inference, make sure to enable GPU: Runtime → Change runtime type → GPU (T4 or better recommended).</p>
+</div>
+
+<div class="grid">
+  <div class="card" style="grid-column: span 12;">
+    <h2>0) 📦 Imports + setup</h2>
+    <p>Install dependencies:</p>
+    <div class="code"><pre><code class="language-bash">pip -q install "transformers>=4.55" "huggingface_hub>=0.23" safetensors torch pyfaidx requests seaborn matplotlib igv_notebook</code></pre></div>
+    
+    <p style="margin-top: 20px;">Import required libraries:</p>
+    <div class="code"><pre><code class="language-python">import re
+import time
+import torch
+import requests
+from transformers import pipeline</code></pre></div>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>1) 📦 Configuration</h2>
+    <p>Set your NTv3 model and genomic window here:</p>
+    <div class="code"><pre><code class="language-python"># Define the model and genomic window
+model_name = "InstaDeepAI/NTv3_650M_pos"
+assembly = "hg38"
+chrom = "chr19"
+start = 6_700_000
+end = 6_831_072</code></pre></div>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>2) 📥 Fetch chromosome sequence for the chosen window</h2>
+    <div class="code"><pre><code class="language-python"># Get the sequence from the UCSC API
+url = f"https://api.genome.ucsc.edu/getData/sequence?genome={assembly};chrom={chrom};start={start};end={end}"
+seq = requests.get(url).json()["dna"].upper()
+print(f"Original sequence length: {len(seq)}")
+
+# Crop to multiple of 128 (the pipeline will crop again, but this is a no-op once divisible)
+seq = seq[:int(len(seq) // 128) * 128]
+print(f"Cropped sequence length: {len(seq)}, {len(seq) / 128} transformer tokens")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      <strong>Example output:</strong><br>
+      Original sequence length: 131072<br>
+      Cropped sequence length: 131072, 1024.0 transformer tokens
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>3) ⚡ Genome annotation pipeline (pre-processing, inference, post-processing)</h2>
+    <div class="code"><pre><code class="language-python"># Build NTv3 GFF pipeline
+ntv3_gff = pipeline(
+    "ntv3-gff",
+    model=model_name,
+    trust_remote_code=True,
+    device=0 if torch.cuda.is_available() else -1,
+)
+
+# Run pipeline: DNA -> NTv3 -> HMM -> GFF3
+inputs = {
+    "sequence": seq,
+    "chrom": chrom,
+    "start": start,
+    "end": end,
+    "assembly": assembly,
+}
+
+# Run the pipeline
+start_time = time.time()
+gff_text = ntv3_gff(inputs)
+end_time = time.time()
+print(f"Inference + decoding time: {end_time - start_time:.2f} seconds")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      The pipeline performs all the necessary steps: running inference with the model, retrieving and processing the predicted probabilities, and applying the HMM to generate a consistent annotation.
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>4) 📁 Save a GFF file</h2>
+    <div class="code"><pre><code class="language-python"># Save GFF3 file
+short_model_name_match = re.search(r"[^/]+$", model_name)
+short_model_name = short_model_name_match.group() if short_model_name_match else model_name
+
+output_filename = f"{short_model_name}_{assembly}_{chrom}_{start}_{end}.gff3"
+with open(output_filename, "w") as output_file:
+    output_file.write(gff_text)
+
+print(f"Saved GFF file to {output_filename}")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      <strong>Example output:</strong> Saved GFF file to NTv3_650M_pos_hg38_chr19_6700000_6831072.gff3
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>5) 🌐 Create an IGV Browser</h2>
+    <div class="code"><pre><code class="language-python">import igv_notebook
+
+igv_notebook.init()
+
+config = {
+    "genome": "hg38",  # built-in hg38
+    "locus": f"{chrom}:{start}-{end}",
+}
+
+gff_track = {
+    "name": "NTv3 annotations",
+    "format": "gff3",
+    "type": "annotation",
+    "url": output_filename,  # just the filename
+}
+
+browser = igv_notebook.Browser(config)
+browser.load_track(gff_track)
+
+# Re-center on the region, just to be sure
+browser.search(f"{chrom}:{start}-{end}")
+browser  # <- just return the object, no .show()</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      This creates an interactive IGV browser visualization of the annotations. The GFF file can also be visualized in any genome browser.
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>📓 Full Notebook</h2>
+    <p>To view and run the complete notebook interactively:</p>
+    <ul>
+      <li><a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_pipelines/02_genome_annotation.ipynb" target="_blank" rel="noopener">View notebook on Hugging Face</a></li>
+      <li>Download and run in Jupyter, Google Colab, or any notebook environment</li>
+    </ul>
+  </div>
+</div>
+

tabs/demo.html

@@ -1,88 +0,0 @@
-<div class="summary">
-  <h2>💻 Interactive Code Demo</h2>
-  <p>Run the NTv3 650M post-trained model interactively to predict functional tracks and genomic elements for any genomic region.</p>
-  <p><strong>Model:</strong> <code>InstaDeepAI/NTv3_650M_pos</code></p>
-</div>
-
-<div class="grid">
-  <div class="card" style="grid-column: span 12;">
-    <h2>🚀 NTv3 Track Prediction Pipeline</h2>
-    <p>Enter a genomic region to get predictions for functional tracks and genomic elements. The model will predict ~7k functional tracks and 21 genomic elements over the center 37.5% of your input region.</p>
-    
-    <!-- Gradio app embedded here -->
-    <!-- Note: With Gradio SDK, the app.py serves as the main interface -->
-    <!-- The HTML interface can still be accessed, but the Gradio demo is the primary interface -->
-    <div id="gradio-container" style="margin-top: 20px; min-height: 600px;">
-      <p style="color: var(--muted); margin-bottom: 15px;">
-        <strong>Note:</strong> With Gradio SDK enabled, the interactive demo is now the main interface of this Space. 
-        You can interact with it directly, or use the code example below to run predictions programmatically.
-      </p>
-      <div style="background: rgba(0,0,0,0.3); padding: 20px; border-radius: 12px; border: 1px solid var(--border);">
-        <p style="color: var(--link); margin: 0;">
-          💡 The Gradio interactive demo is now available as the main interface of this Space. 
-          Refresh the page to see it, or use the code example below.
-        </p>
-      </div>
-    </div>
-    
-    <p style="margin-top: 20px; color: var(--muted); font-size: 13px;">
-      <strong>Note:</strong> The first run may take longer as the model loads. Maximum region size: 1 Mb (1,000,000 base pairs).
-    </p>
-  </div>
-  
-  <div class="card" style="grid-column: span 12;">
-    <h2>📝 Code Example</h2>
-    <p>Here's the Python code that powers the demo above. You can run this in a notebook or Python script:</p>
-    <div class="code"><pre><code class="language-python">from transformers import pipeline
-import torch
-
-model_name = "InstaDeepAI/NTv3_650M_pos"
-
-ntv3_tracks = pipeline(
-    "ntv3-tracks",
-    model=model_name,
-    trust_remote_code=True,
-    device=0 if torch.cuda.is_available() else -1,
-)
-
-# Run track prediction
-out = ntv3_tracks(
-  {
-    "chrom": "chr19",
-    "start": 6_700_000,
-    "end": 6_831_072,
-    "species": "human"
-  }
-)
-
-# Print output shapes
-# 7k human tracks over 37.5 % center region of the input sequence
-print("bigwig_tracks_logits:", tuple(out.bigwig_tracks_logits.shape))
-# Location of 21 genomic elements over 37.5 % center region of the input sequence
-print("bed_tracks_logits:", tuple(out.bed_tracks_logits.shape))
-# Language model logits for whole sequence over vocabulary
-print("language model logits:", tuple(out.mlm_logits.shape))</code></pre></div>
-    <p style="margin-top: 15px;">To run the interactive Gradio app locally:</p>
-    <div class="code"><pre><code class="language-bash">pip install -r requirements.txt
-python app.py</code></pre></div>
-  </div>
-</div>
-
-<script>
-// Try to detect if Gradio app is available
-window.addEventListener('load', function() {
-  const iframe = document.getElementById('gradio-iframe');
-  iframe.onerror = function() {
-    // If iframe fails to load, keep showing the instructions
-    document.getElementById('gradio-loading').style.display = 'block';
-    iframe.style.display = 'none';
-  };
-  // Set a timeout to show instructions if iframe doesn't load
-  setTimeout(function() {
-    if (iframe.style.display === 'none') {
-      document.getElementById('gradio-loading').style.display = 'block';
-    }
-  }, 2000);
-});
-</script>
-

tabs/functional_tracks.html

@@ -0,0 +1,220 @@
+<div class="summary">
+  <h2>🧬 NTv3 Post-Trained Functional Track Prediction</h2>
+  <p>This notebook demonstrates how to use the NTv3 post-trained model to predict functional tracks and genome annotation directly from a DNA sequence.</p>
+  <p>The pipeline abstracts away all the underlying steps: running inference with the model and plotting the predictions per tracks.</p>
+  <p>If you're interested in exploring the intermediate probabilities, please refer to the <a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_tutorials/01_tracks_prediction.ipynb" target="_blank" rel="noopener">track-prediction notebook</a>.</p>
+  <p><strong>📝 Note for Google Colab users:</strong> This notebook is compatible with Colab! For faster inference, make sure to enable GPU: Runtime → Change runtime type → GPU (T4 or better recommended).</p>
+</div>
+
+<div class="grid">
+  <div class="card" style="grid-column: span 12;">
+    <h2>0) 📦 Imports + setup</h2>
+    <p>Install dependencies:</p>
+    <div class="code"><pre><code class="language-bash">pip -q install "transformers>=4.55" "huggingface_hub>=0.23" safetensors torch pyfaidx requests seaborn matplotlib igv_notebook pyBigWig</code></pre></div>
+    
+    <p style="margin-top: 20px;">Import required libraries:</p>
+    <div class="code"><pre><code class="language-python">import re
+import time
+import os
+import torch
+import requests
+import numpy as np
+import pyBigWig
+from transformers import pipeline, AutoConfig</code></pre></div>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>1) 📦 Configuration</h2>
+    <p>Set your NTv3 model and genomic window here:</p>
+    <div class="code"><pre><code class="language-python"># Define the model and genomic window
+model_name = "InstaDeepAI/NTv3_650M_pos"
+
+species = "human"  # will use for condition the model on species
+assembly = "hg38"  # will use for fetching the chromosome sequence
+
+chrom = "chr19"
+start = 6_700_000
+end = 6_831_072</code></pre></div>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>2) 📥 Fetch chromosome sequence for the chosen window</h2>
+    <div class="code"><pre><code class="language-python"># Get the sequence from the UCSC API
+url = f"https://api.genome.ucsc.edu/getData/sequence?genome={assembly};chrom={chrom};start={start};end={end}"
+seq = requests.get(url).json()["dna"].upper()
+print(f"Original sequence length: {len(seq)}")
+
+# Crop to multiple of 128 (the pipeline will crop again, but this is a no-op once divisible)
+seq = seq[:int(len(seq) // 128) * 128]
+print(f"Cropped sequence length: {len(seq)}, {len(seq) / 128} transformer tokens")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      <strong>Example output:</strong><br>
+      Original sequence length: 131072<br>
+      Cropped sequence length: 131072, 1024.0 transformer tokens
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>3) ⚡ Functional track prediction pipeline (pre-processing, inference, plotting)</h2>
+    <div class="code"><pre><code class="language-python"># Build NTv3 tracks pipeline
+ntv3_tracks = pipeline(
+    "ntv3-tracks",
+    model=model_name,
+    trust_remote_code=True,
+    device=0 if torch.cuda.is_available() else -1,
+)
+
+# Select tracks to plot
+tracks_to_plot = {
+    "K562 RNA-seq": "ENCSR056HPM",
+    "K562 DNAse": "ENCSR921NMD",
+    "K562 H3k4me3": "ENCSR000DWD",
+    "K562 CTCF": "ENCSR000AKO",
+    "HepG2 RNA-seq": "ENCSR561FEE_P",
+    "HepG2 DNAse": "ENCSR000EJV",
+    "HepG2 H3k4me3": "ENCSR000AMP",
+    "HepG2 CTCF": "ENCSR000BIE",
+}
+elements_to_plot = ["protein_coding_gene", "exon", "intron", "splice_donor", "splice_acceptor"]
+
+# Run pipeline: DNA -> NTv3 -> Tracks -> plot
+start_time = time.time()
+
+ntv3_predictions = ntv3_tracks(
+    {"chrom": "chr19", "start": 6_700_000, "end": 6_831_072, "species": species},
+    plot=True,
+    tracks_to_plot=tracks_to_plot,
+    elements_to_plot=elements_to_plot,
+)
+
+end_time = time.time()
+
+print(f"Inference + decoding time: {end_time - start_time:.2f} seconds")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      The pipeline performs all the necessary steps: running inference with the model and plotting the predictions for the specified tracks and genomic elements.
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>4) 📁 Save as BigWig file</h2>
+    <div class="code"><pre><code class="language-python"># Load config to get track names and find indices for tracks_to_plot
+cfg = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+all_bigwig_names = cfg.bigwigs_per_file_assembly[assembly]
+
+# Find indices of tracks we want to save
+# Use display names (keys) for filenames, but track IDs (values) to find indices
+track_data_list = []  # List of (display_name, track_id, index) tuples
+for display_name, track_id in tracks_to_plot.items():
+    try:
+        idx = all_bigwig_names.index(track_id)
+        track_data_list.append((display_name, track_id, idx))
+    except ValueError:
+        print(f"Warning: Track '{track_id}' ({display_name}) not found in config. Skipping...")
+
+print(f"Found {len(track_data_list)} tracks to save from tracks_to_plot")
+
+# Get predictions (shape: (49152, 7362))
+bigwig_logits = ntv3_predictions.bigwig_tracks_logits
+if isinstance(bigwig_logits, torch.Tensor):
+    bigwig_logits = bigwig_logits.detach().cpu().numpy()
+
+# Calculate genomic coordinates for the center 37.5% region
+# The predictions cover the center 37.5% of the input sequence
+input_length = end - start
+center_start_offset = int(input_length * 0.3125)  # (1 - 0.375) / 2 = 0.3125
+center_length = int(input_length * 0.375)
+center_start = start + center_start_offset
+center_end = center_start + center_length
+
+print(f"Input region: {chrom}:{start}-{end} (length: {input_length:,} bp)")
+print(f"Prediction region: {chrom}:{center_start}-{center_end} (length: {center_length:,} bp)")
+print(f"Number of positions: {bigwig_logits.shape[0]}")
+
+# Create output directory
+output_dir = "bigwig_outputs"
+os.makedirs(output_dir, exist_ok=True)
+
+# Save each track as a separate BigWig file
+print(f"\nSaving BigWig files to '{output_dir}/' directory...")
+for i, (display_name, track_id, track_idx) in enumerate(track_data_list):
+    # Get track data (logits for this track)
+    track_data = bigwig_logits[:, track_idx].astype(np.float32)
+    
+    # Create BigWig file using display name (key) for filename
+    # Clean the display name for use as filename (replace spaces, special chars)
+    track_clean_name = display_name.replace(" ", "_").replace("/", "_").replace("-", "_")
+    bw_filename = os.path.join(output_dir, f"{track_clean_name}.bw")
+    bw = pyBigWig.open(bw_filename, "w")
+    
+    # Add header (chromosome and size)
+    bw.addHeader([(chrom, end)])
+    
+    # Add entries (intervals with values)
+    # Each position in track_data corresponds to one base pair
+    starts = np.arange(center_start, center_start + len(track_data), dtype=np.int64)
+    ends = starts + 1
+    values = track_data.tolist()
+    
+    bw.addEntries(
+        chroms=[chrom] * len(starts),
+        starts=starts.tolist(),
+        ends=ends.tolist(),
+        values=values
+    )
+    
+    bw.close()
+    
+    print(f"  Saved {i + 1}/{len(track_data_list)}: {display_name} ({track_clean_name}.bw)")
+
+print(f"\n✅ Successfully saved {len(track_data_list)} BigWig files to '{output_dir}/'")</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      This saves each selected functional track as a separate BigWig file that can be visualized in genome browsers. The files are saved with user-friendly display names (e.g., "K562_RNA_seq.bw").
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>5) 🌐 Create an IGV Browser</h2>
+    <div class="code"><pre><code class="language-python">import igv_notebook
+
+igv_notebook.init()
+
+# Build tracks array with all BigWig files we saved
+tracks = []
+for track_display_name, track_id in tracks_to_plot.items():
+    # Clean the display name to match the filename we saved
+    track_clean_name = track_display_name.replace(" ", "_").replace("/", "_").replace("-", "_")
+    bigwig_path = os.path.join(output_dir, f"{track_clean_name}.bw")
+    bigwig_track = {
+        "name": track_display_name,
+        "format": "bigwig",
+        "url": bigwig_path,
+        "height": 70,
+        "autoscale": True,
+        "displayMode": "EXPANDED",
+    }
+    tracks.append(bigwig_track)
+
+config = {
+    "genome": assembly,
+    "locus": f"{chrom}:{center_start}-{center_end}",
+    "tracks": tracks,
+    "theme": "dark",
+}
+
+browser = igv_notebook.Browser(config)
+browser  # <- just return the object, no .show()</code></pre></div>
+    <p style="margin-top: 15px; color: var(--muted); font-size: 13px;">
+      This creates an interactive IGV browser visualization with a dark theme showing all the predicted functional tracks. The BigWig files can also be visualized in any genome browser.
+    </p>
+  </div>
+
+  <div class="card" style="grid-column: span 12;">
+    <h2>📓 Full Notebook</h2>
+    <p>To view and run the complete notebook interactively:</p>
+    <ul>
+      <li><a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_pipelines/01_functional_track_prediction.ipynb" target="_blank" rel="noopener">View notebook on Hugging Face</a></li>
+      <li>Download and run in Jupyter, Google Colab, or any notebook environment</li>
+    </ul>
+  </div>
+</div>
+

tabs/home.html

@@ -92,7 +92,7 @@
     <div class="card">
       <h2>📓 Pipelines notebooks (browse <a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/tree/main/notebooks_pipelines" target="_blank" rel="noopener">folder</a>)</h2>
       <ul>
-        <li> 🎯 01 — Generate bigwig predictions for certain tracks</li>
+        <li><a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_pipelines/01_functional_track_prediction.ipynb" target="_blank" rel="noopener">🎯 01 — Generate bigwig predictions for certain tracks</a></li>
         <li><a href="https://huggingface.co/spaces/InstaDeepAI/ntv3/blob/main/notebooks_pipelines/02_genome_annotation.ipynb" target="_blank" rel="noopener">🏷️ 02 — Genome annotation / segmentation</a></li>
         <li>🎯 03 — Fine-tune on bigwig tracks</li>
         <li>🔍 04 — Interpret a given genomic region</li>