Use Mosaic Hallucination Online
Commercially Available Mosaic Hallucination No-Code Web Server
Mosaic: Composite-Objective Protein Design with Continuous Relaxation
Mosaic is an open-source Python framework for functional, multi-objective protein design using continuous relaxation. It wraps many protein property models behind a common JAX interface so that their outputs can be combined into a single differentiable objective. The suggested citation lists Boyd, N., Guns, S. and Escalante Bio (2025), and the code is released under the MIT license.
Designing a protein usually means satisfying several goals at once: a structure predictor should be confident in the complex, the sequence should look natural to language models, and stability and inverse-folding scores should stay favorable. Running separate tools in containers and shell scripts makes such combinations awkward and prone to dependency conflicts.
Mosaic's stated goal is to avoid containers, bash scripts and dependency conflicts by letting designers compose models in code and optimize them together.
Attribute | Detail |
|---|---|
Approach | Weighted combination of differentiable loss terms, optimized over a soft sequence |
Structure models | Boltz-1, Boltz-2, BoltzGen, AlphaFold2, OpenFold3, Protenix versions and others |
Sequence models | ProteinMPNN variants, ESM-2, ESM-C, AbLang, AbLang2 and a stability predictor |
Inputs | Target sequence, binder length, composed loss, optimizer settings |
Outputs | Designed sequences (PSSM), structure predictions, inverse-folded sequences |
License | MIT |
How Mosaic Combines Models into a Composite Design Objective
Loss terms: Each model contributes a differentiable loss, written as a JIT-compatible callable object, and custom terms can be added as classes.
Weighted composition: Terms are combined with coefficients and arithmetic. The README example mixes a Boltz-1 complex loss, a monomer loss, ESM pseudolikelihood, trigram log-likelihood, stability and ProteinMPNN terms.
Transformations: Utilities such as clipped losses, normalized gradients and fixed-position penalties adjust how individual terms behave.
Optimizers: simplex_APGM is an accelerated proximal gradient method on the probability simplex, batched_simplex_APGM adds vectorized throughput, and gradient_MCMC is a discrete sampler for fine-tuning designs.
Reported Experimental Results from Mosaic Designs
Case study | Reported outcome |
|---|---|
Minibinder walkthrough, PD-L1 | 8 of 10 designs validated experimentally |
Minibinder walkthrough, IL7Ra | 7 of 10 designs validated experimentally |
Adaptyv Nipah competition | Single-digit nanomolar binders and a 90% wet-lab success rate |
Adaptyv x Muni TREM2 | Top-affinity binder overall (1.11 nM) |
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 Mosaic on Tamarind Bio
Minibinder design: Design small binders against targets such as PD-L1 and IL7Ra, as in the README walkthrough.
Multi-objective optimization: Balance structure confidence, language-model likelihood, stability and inverse-folding terms in one objective.
Competition-style binder campaigns: The README cites a Boltz-2, soluble ProteinMPNN and simplex_APGM script used in the Adaptyv Nipah competition.
Custom objectives: Add new loss terms to encode design goals specific to a project.
How to Use Mosaic on Tamarind Bio
Open the tool: Log in to tamarind.bio and select Mosaic Hallucination.
Provide the target: Enter the target chain sequence, and a template structure if exposed.
Set the binder length: Choose the length of the protein to design.
Choose objectives: If exposed, select which models contribute loss terms and their weights.
Set optimizer parameters: If exposed, adjust step size and number of steps; the README suggests a starting step size of about 0.1 times the square root of the binder length.
Run the job: Submit and wait, noting that the first JIT-compiled call can be slow.
Download designs: Retrieve the designed sequences and predicted structures.
Filter and validate: Apply standard in-silico filters before ordering designs.
Things to Keep in Mind
The authors warn that Mosaic may require substantial hand-holding, such as tuning learning rates, unlike some other design tools.
It often produces proteins that fail simple in-silico tests, so standard filtering should be combined with it.
Optimization is hard, and some objectives may need custom heuristics.
For AF3-style models, design-time binder predictions lack sidechains; re-predicting with target-only features restores them.
Source: Mosaic README, escalante-bio/mosaic (suggested citation: Boyd N., Guns S. and Escalante Bio, 2025); no paper DOI is listed. Code: https://github.com/escalante-bio/mosaic