> ## Documentation Index
> Fetch the complete documentation index at: https://docs.revilico.bio/llms.txt
> Use this file to discover all available pages before exploring further.

# Boltz2 Cofolding

> Predict binding poses and activity metrics from a protein sequence and SMILES string using Boltz2's AI co-folding engine

## Overview

**Boltz2 Cofolding** is Rev-Bind's AI-powered structure prediction engine. Unlike classical docking, which requires a pre-determined 3D protein structure and a defined binding pocket, Boltz2 takes just two inputs — a **protein sequence** and a **SMILES string** — and generates a co-folded 3D structure showing the ligand placed within its most biologically relevant binding pocket.

This makes Boltz2 an especially powerful early-stage tool: you can generate binding pose predictions and activity metrics without needing a crystal structure or experimental pocket data.

<div style={{position: "relative", paddingBottom: "54.48%", height: 0}}>
  <iframe src="https://www.loom.com/embed/1b00be061f524e3ba446ba122ec24e99" frameBorder="0" webkitallowfullscreen="true" mozallowfullscreen="true" allowFullScreen={true} style={{position: "absolute", top: 0, left: 0, width: "100%", height: "100%"}} />
</div>

[Running Boltz2 Protein-Ligand Cofolding Pipeline — Watch Video](https://www.loom.com/share/1b00be061f524e3ba446ba122ec24e99)

***

## How Boltz2 Cofolding Works

Boltz2 runs an AI engine that concatenates the protein sequence and small molecule structure, then simultaneously folds and docks them together. The resulting **co-folded structure** is a prediction of the binding pose — specifically, how the ligand sits within the pocket that drives the target's biological activity.

For example, with a kinase target the engine will co-fold the ligand into the ATP-competitive hinge region pocket — placing the compound where it would need to bind to produce a therapeutic effect — without requiring you to first define that pocket manually.

### Comparison to Classical Docking

|                            | Classical Docking (RevScreen)                 | Boltz2 Cofolding                                 |
| -------------------------- | --------------------------------------------- | ------------------------------------------------ |
| Input                      | 3D protein structure + ligand 3D conformation | Protein sequence + SMILES                        |
| Pocket definition          | Required (manual or RevPocket)                | Automatic                                        |
| Conformational flexibility | Limited                                       | Full co-fold                                     |
| Best for                   | Large-scale library screening                 | Early-stage pose prediction, novel targets       |
| Output                     | Docking scores, poses                         | Co-folded structure, confidence, affinity, pIC50 |

***

## Output Metrics

Each co-folded result returns three primary metrics:

| Metric                   | Description                                                                                  |
| ------------------------ | -------------------------------------------------------------------------------------------- |
| **Confidence Score**     | The model’s confidence in the predicted binding pose — reflects structural reliability       |
| **Affinity Probability** | A heuristic estimate of binding affinity — useful for relative ranking across a compound set |
| **Predicted pIC50**      | The primary activity outcome — a predicted potency metric derived from the co-folded pose    |

<Note>
  Click the **i** icon next to any metric in the results panel for a plain-language explanation of what it means and how to interpret it.
</Note>

These metrics together give you a rapid, hypothesis-generating view of how well a compound is likely to engage your target — useful for triaging compound sets before committing to more compute-intensive methods.

***

## Running a Boltz2 Pipeline

<Steps>
  <Step title="Navigate to Boltz2">
    From the Revilico OS dashboard, go to **Rev-Bind → Boltz2**.
  </Step>

  <Step title="Open Data Engineering">
    Click into **Data Engineering**. You’ll see two input fields: one for the **protein sequence** and one for the **SMILES string**.

    If you don’t have a campaign running yet, use the **demo files** to follow along — they contain all the required inputs pre-formatted.
  </Step>

  <Step title="Input Your Protein Sequence">
    The protein sequence should be provided as a **CSV file** with a single chain column (e.g., `AXLChainA`) containing only the raw amino acid sequence string.

    You have two options:

    * **Drag-drop** your CSV file directly into the data engineering pane
    * Click **Add Manually** and paste the sequence string directly into the input field

    <Tip>
      The demo file format shows exactly how the CSV should be structured. Reference it if you are building your own input file for the first time.
    </Tip>
  </Step>

  <Step title="Input Your SMILES String">
    The small molecule is defined by its SMILES string. You can provide this as:

    * A **CSV file** containing one or more SMILES strings for batch cofolding
    * **Manual entry** — paste a single SMILES string directly into the input field

    For a bulk campaign (e.g., 1,000 compounds), prepare a CSV with one SMILES per row and drag-drop it in.
  </Step>

  <Step title="Name Your Pipeline">
    Enter a descriptive name for the pipeline (e.g., `AXL-boltz2-type2-inhibitors`). This name will identify the run in your pipeline history.
  </Step>

  <Step title="Select the Pipeline Type">
    Set the **Pipeline Type** to **Affinity** for standard protein-ligand cofolding. Additional co-folding types are available for more complex analyses:

    * **Affinity** — Protein + ligand cofolding for binding pose and activity prediction *(standard)*
    * **Multimer** — Co-fold multiple protein chains together
    * Additional types for varied protein-small molecule configurations

    <Note>
      Make sure to select the pipeline type that matches your experimental setup. Using the wrong type will produce mismatched output.
    </Note>
  </Step>

  <Step title="Create and Run the Pipeline">
    Click **Create Pipeline**. The engine will queue and run your co-folding job. You’ll see a confirmation message when the pipeline is created successfully.

    Track all active and completed pipelines from the **central pipeline hub** in the navigation bar.
  </Step>
</Steps>

***

## Reading and Navigating Your Results

Once the pipeline completes, navigate to your results from the pipeline hub.

### Viewing Co-folded Structures

The primary output is an interactive 3D viewer showing the co-folded protein-ligand complex. The ligand will be placed within the highest-fidelity pocket — the site most likely to drive biological activity for that target class.

For kinase targets, for example, you’ll see the compound positioned within the **hinge region** — competing with ATP in the canonical ATP-binding site, which is exactly the mechanism a type II competitive inhibitor would engage.

### Navigating Large Result Sets

When running campaigns with many compounds, use the results panel to browse your full ligand list. You can jump directly to any specific compound by number (e.g., **Ligand 450**) rather than scrolling through the entire set.

### Exporting Data

For downstream analysis of large batches, click **Export CSV** to download your full results table. This includes all confidence scores, affinity probabilities, and predicted pIC50 values for every compound in the run — ready for statistical analysis, filtering, or integration into your own workflows.

***

## Calibrating Boltz2 Outputs

Boltz2's predicted metrics are powerful heuristics for **relative ranking** within a compound set. They are not absolute experimental measurements.

Best practice is to:

1. Run Boltz2 on a set of compounds that includes known actives and known inactives from your target
2. Validate that the predicted pIC50 and confidence scores rank-order consistently with your experimental data
3. Use this calibration to interpret predictions for novel compounds with greater confidence

<Warning>
  Boltz2 outputs should always be calibrated against experimental results at each stage of your campaign. Treat predicted pIC50 as a relative ranking tool rather than an absolute potency prediction until calibrated for your target system.
</Warning>

***

## Triaging to Protein-Ligand MD

Once you have identified your highest-confidence co-folded poses, you can triage them directly into **Protein-Ligand MD** for more robust validation.

From the results panel, select your compound(s) of interest and click **Triage to Protein-Ligand MD**. This will:

* Export the co-folded pose as the starting structure for MD
* Queue an MD simulation to assess binding stability and residence time
* Enable MMPBSA and MMGBSA rescoring for more rigorous free energy estimates

This workflow is covered in depth in the [**RevMD-Bind tutorial**](/docs/revmd-bind).

***

## Next Steps

After Boltz2 cofolding, your compound list is ready for:

* [**Protein-Ligand MD — RevMD-Bind**](/docs/revmd-bind) — Triage top poses directly to MD for stability validation and MMPBSA/MMGBSA rescoring
* [**RevFEP**](/docs/revfep) — Rigorous alchemical free energy calculations for final-stage candidate ranking
* [**RevScreen - Static & Flexible Docking**](/tutorials/rev-bind/static-flexible-docking) — Run classical docking screens on the same compound set for orthogonal validation
* [**RevScreen - Ensemble Docking**](/tutorials/rev-bind/ensemble-docking) — Use MD-sampled protein conformations for maximum-accuracy virtual screening on shortlisted leads