Mosaic: The Functional Language for Generative Protein Design

Mosaic is a modular, high-performance framework for composite-objective protein design. Developed by Escalante Bio, Mosaic redefines the protein engineering workflow by treating neural networks as simple, differentiable functions that can be programmatically composed. By leveraging the power of JAX, Mosaic allows researchers to optimize complex, multi-objective loss functions—balancing binding affinity, stability, and expressibility—all within a single, unified optimization loop.

Unlike traditional "black-box" pipelines, Mosaic serves as an expressive language for biology, enabling the design of therapeutic candidates from the ground up through gradient-based optimization of continuous sequence space.

Key Innovations: Differentiable Multi-Objective Design

Mosaic moves beyond single-property optimization, providing the infrastructure to satisfy the competing requirements of real-world drug discovery.

• Composite Objective Functions: Seamlessly combine multiple "expert" models (e.g., Boltz-1, ESM-2, ProteinMPNN) into a single loss function to simultaneously optimize for binding, solubility, and thermal stability.

• Gradient-Based Optimization: Utilizes gradient ascent/descent over relaxed sequence space—mechanically analogous to "DeepDream"—to iteratively nudge protein sequences toward an ideal functional profile.

• Model-Agnostic Interface: A unified interface for state-of-the-art structure prediction models, including Boltz-1, Boltz-2, AF2, and Protenix, allowing users to swap backbones with a single line of code.

• Native JAX Integration: Built on JAX for high-performance XLA compilation and automatic differentiation, enabling rapid scaling on TPU and GPU hardware.

• Functional & Modular: Replaces cumbersome bash scripts and containers with a functional Pythonic API where models are treated as pure functions.

Performance & Capabilities

Mosaic provides a flexible framework where accuracy is driven by the underlying "committee" of expert models. By integrating the latest SOTA predictors, it consistently identifies high-fitness variants that satisfy complex constraints.

Scientific Breakthroughs in Protein Engineering

Custom Objective Sculpting

Mosaic allows researchers to "sculpt" their desired protein. In binder design, one can simultaneously ensure the probability of binding the intended target is maximized while the chance of binding to a similar off-target protein is minimized. This "negative design" is critical for reducing drug side effects.

Accelerated In-Silico Screening

By optimizing the input sequence directly via backpropagation, Mosaic can generate 1,000 to 50,000 potential designs in a single run. These are then ranked and filtered using Mosaic’s built-in confidence modules, narrowing the field down to the top ~10 candidates for wet-lab validation.

Beyond Non-Coding RNA: Universal Design

While initially focused on proteins, Mosaic’s modular architecture is designed to handle any biological modality that can be represented as a differentiable loss—including DNA and RNA complexes—making it a truly universal backbone for synthetic biology.

Mosaic on Tamarind Bio: Design Without Limits

Tamarind Bio democratizes access to Escalante Bio’s powerful design framework. By abstracting away the difficulties of JAX environment setup, CUDA/XLA driver versioning, and high-performance TPU orchestration, Tamarind Bio lets you start designing proteins in minutes.

• No-Code & Low-Code Deployment: Use the Tamarind dashboard to configure your "committee of experts" and launch massive design runs.

• Scalable TPU Compute: Leverage world-class hardware to run thousands of parallel optimizations without managing your own cloud infrastructure.

How to Use Mosaic on Tamarind Bio

1. Access the Platform: Log in to tamarind.bio and select the Mosaic Protein Design tool.

2. Define Your Target: Upload the PDB structure or sequence of the target protein you wish to design a binder for.

3. Build Your Objective: Choose your expert models (e.g., Boltz-1 for structure, ESM-2 for fitness, and a custom stability model).

4. Set Constraints: Define your length requirements, specific residues to fix, and any "off-target" proteins to avoid.

5. Run Optimization: The platform executes gradient-based optimization, iteratively updating the sequence to maximize your custom score.

6. Analyze & Export: View the resulting 3D structures, confidence scores (pLDDT), and sequence logos. Download the top candidates for experimental synthesis.

Source