Boltz-2 is now live in Aura, predict structure and affinity

We’ve just integrated Boltz-2 into Aura, Fastfold’s multi-agent AI scientist. Boltz-2 predicts 3D structure and binding affinity between proteins and ligands in a single fast pass—now accessible through a natural language prompt.

Try it with a single prompt

In Aura, you don’t need to code or install anything. Just tell the agent what you want:

Fold this protein: PQITLWQRPLVTIKIGGQLKEAL.. and this ligand: CC1CN... using Boltz-2, add affinity property to the ligand.

Aura will:

1. Predict the co-folded protein–ligand complex
2. Estimate the binding affinity
3. Visualize the result
4. Log the result to your workspace for downstream workflows

The output includes a predicted binding affinity and probability of interaction, reported per model in the ensemble. Here’s an example result:

*Affinity values are derived from IC50 in μM; lower values indicate stronger binding.
**Probability indicates the likelihood that the ligand is a true binder (0 to 1, shown here as %).

Interpreting affinity values

Boltz-2 internally predicts log(IC50), which is then converted to μM for readability. You can interpret affinity strength as follows:

• IC50 of 10⁻⁹ M ⟶ log(IC50) ≈ -3 → strong binder
• IC50 of 10⁻⁶ M ⟶ log(IC50) ≈ 0 → moderate binder
• IC50 of 10⁻⁴ M ⟶ log(IC50) ≈ 2 → weak binder or decoy

To convert log(IC50) to an approximate ΔG (binding free energy in kcal/mol):
ΔG ≈ (6 - log(IC50)) × 1.364

The probability scores range from 0 to 1 and indicate the model's confidence that the ligand is a true binder. These probabilities can be used for ranking candidates or applying filtering thresholds in large-scale screens.

Aura displays all of this automatically for each prediction—helping you move quickly from sequence to insight.

Boltz-2 represents a significant advancement in biomolecular modeling by integrating structure prediction with binding affinity estimation. Unlike its predecessor Boltz-1, which focused solely on structural outputs, Boltz-2 introduces a dedicated affinity module trained on extensive datasets, including synthetic constructs and molecular dynamics-derived data. This enables Boltz-2 to estimate binding affinities with accuracy comparable to FEP, the gold standard in free energy calculations, while operating approximately 1000× faster.

In benchmarking studies, Boltz-2 performs on par with OpenFE for FEP+ targets that were held out from training. More impressively, it achieved top performance in CASP16’s binding affinity challenge, accurately predicting affinities across 140 complexes. These results highlight Boltz-2’s ability to generalize and scale, making it a powerful tool for hit discovery, lead optimization, and generative design.

📄 Read the full Boltz-2 paper

---

Composable protein design workflows

Aura isn’t just an interface—it’s a platform for composing, optimizing, and running protein design loops. With Boltz-2, you can now build full pipelines that combine:

• Boltz-1 for structure
• ESM2 for function
• ProteinMPNN for stability
• n-gram priors for controllability
• Boltz-2 for affinity optimization

These can be tied together through multi-objective loss functions (e.g., target contact, radius of gyration, helical content), letting you optimize complex binders in minutes instead of weeks, and to take it an extra mile, we are also working on sending your candidates direclty to lab testing inside Aura.

Check this protein binder design demo.

Optmization steps and metrics:

---

Available now

Boltz-2 is now live and ready to use in the Aura Agent. Simply start typing your goal, and Aura will compose the necessary steps behind the scenes.

If you're not yet on Aura, register or contact us at if any questions or feedback hello@fastfold.ai.

We’re excited to see what you’ll build with Boltz-2—from affinity-guided screening to structure-based generative design. Aura gives scientists the power to move from idea to validation in minutes.

---

References

• Boltz-2 paper (Saro Passaro, Gabriele Corso, Jeremy Wohlwend et al)
• Boltz GitHub repository