escalante-bio/mosaic

What it solves

Mosaic provides a unified framework for multi-objective protein design. It addresses the difficulty of combining multiple disparate protein property predictors (e.g., structure prediction, stability, and solubility models) into a single, differentiable optimization pipeline without dealing with fragmented dependencies or complex bash scripts.

How it works

Built on JAX for high-performance just-in-time (JIT) compilation, Mosaic implements a common interface for a wide variety of protein models. It uses gradient-based optimization over a continuous, relaxed sequence space, allowing users to define a composite loss function—combining learned potentials from different models—and optimize it using various gradient-based optimizers (such as simplex_APGM).

Who it’s for

It is designed for researchers and developers in protein engineering and computational biology who need to implement custom objective functions and optimization algorithms for specific applications, such as designing high-affinity protein binders.

Highlights

• Broad Model Integration: Includes a "zoo" of models including Boltz-1/2, AlphaFold2, OpenFold3, Protenix, ESMFold2, ProteinMPNN, and ESM-2/C.
• Modular Loss Functions: Allows easy combination of structure prediction losses, inverse folding likelihoods, and stability predictors into a single objective.
• Continuous Relaxation: Enables efficient gradient-based optimization of protein sequences.
• Flexible Optimizers: Provides a set of built-in optimizers with the ability to easily implement and swap in custom optimization algorithms.