Use Caliby Online
Commercially Available Caliby No-Code Web Server
Caliby: Ensemble-Conditioned Protein Sequence Design with Potts Models
Caliby is a Potts model-based method for protein sequence design that can condition on structural ensembles. It builds on ProteinMPNN and is released as open code under the Apache License 2.0 on GitHub (ProteinDesignLab).
The authors observe that sequence design models trained for native sequence recovery often produce sequences that do not reliably fold into their target backbones. They hypothesize that such models learn non-structural signals such as phylogenetic relatedness and dataset biases, which limits the foldability of designs.
Property | Detail |
|---|---|
Architecture | 3 encoder and 3 decoder MPNN layers, hidden dimension 128 |
Output | Sitewise fields and pairwise couplings (Potts model) |
Sampling | Discrete Langevin Monte Carlo with a composition perplexity restraint |
Ensemble size | K = 16 conformers |
Variants | Caliby-L and SolubleCaliby |
How Caliby Designs Sequences from Structural Ensembles
Potts model per structure: predicts sitewise fields and pairwise couplings for a single backbone.
Ensemble averaging: averages the energy terms across the Potts models of K structures, a product-of-experts formulation; once precomputed, sampling is as fast as single-structure design.
Synthetic ensembles: Protpardelle-1c partial diffusion (150 of 500 steps) yields 15 extra conformers.
Sampling: discrete Langevin Monte Carlo with a local composition perplexity restraint penalizes low-complexity sequences.
Caliby Results: Designability on Native and De Novo Backbones
Evaluation used 85 native backbones (monomers of 100 to 1024 residues) and 100 RFdiffusion backbones at each length from 100 to 500, with 8 sequences per backbone and single-sequence AlphaFold2 or ESMFold as the check. Ensemble design lowered native sequence recovery while improving pLDDT, and returns diminished beyond K = 16. Ensemble design raised Caliby-L's pLDDT to near Caliby's. SolubleCaliby showed improved designability over SolubleMPNN, mainly on binder-only metrics.
Binder success criterion | Threshold |
|---|---|
i_pTM | > 0.5 |
i_pAE | < 0.35 |
Binder pLDDT | > 0.8 |
Binder RMSD | < 3.5 A |
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Caliby on Tamarind Bio
Designing for flexible proteins: condition on conformational ensembles for multistate targets.
Designing on generated backbones: sequence de novo backbones, such as those from RFdiffusion.
Soluble binder design: use SolubleCaliby for soluble proteins and binders.
Sequence scoring: score existing sequences against a backbone or ensemble.
How to Use Caliby on Tamarind Bio
Open the tool: log in to tamarind.bio and open Caliby.
Upload a structure: a PDB or CIF backbone, optionally with side chains as context.
Choose ensembles: generate a synthetic ensemble or supply your own conformers, if exposed.
Set constraints: fixed positions, omitted amino acids, temperature or symmetry, if exposed.
Run: submit the job.
Download outputs: designed sequences with energies, and optionally self-consistency metrics.
Validate: check designs with AlphaFold2 or ESMFold before experiments.
Things to Keep in Mind
Performance drops at high rewind steps (175 or more), where ensembles become too diverse.
All evaluations are computational; no experimental validation is reported.
Single-sequence AlphaFold2 may be an ineffective filter for multistate design.
ESMFold struggles more than AlphaFold2 on RFdiffusion backbone designs.
Source: bioRxiv preprint describing Caliby, https://www.biorxiv.org/content/10.1101/2025.09.30.679633v4. Code: https://github.com/ProteinDesignLab/caliby.