Semantic Mutagenesis

Generates intermediate protein sequences by performing spherical linear interpolation (slerp) between the learned embeddings of two parent proteins. This technique navigates the continuous latent manifold to produce chimeric sequences that blend structural and functional properties. The smoothness of the latent space ensures that interpolated points decode into structurally plausible proteins rather than random amino acid strings, enabling the exploration of evolutionary paths not sampled by nature.

Applies controlled additive noise or directional shifts to a protein's embedding to alter specific biochemical properties. By identifying the vector direction in latent space that correlates with a target attribute—such as increased thermostability or altered substrate specificity—engineers can systematically mutate the semantic representation before decoding. This method transforms abstract model representations into actionable sequence-level edits without requiring explicit structural knowledge.

Leverages conditioning tags during autoregressive generation to steer output sequences toward desired functional profiles. Models like ProGen2 accept control codes specifying protein family, taxonomic lineage, or functional annotations. During decoding, the model samples from a conditional distribution P(sequence | attribute), effectively performing semantic mutagenesis by constraining the generative process to regions of sequence space associated with the specified property.

Treats the latent space as a smooth proxy for the true fitness landscape, where distance and direction correspond to functional similarity. By mapping experimental fitness data onto the latent representation, researchers can identify fitness peaks and valleys and navigate between them. This enables the discovery of high-fitness sequences that are unreachable by single-step mutagenesis, effectively bypassing local optima through semantic leaps.

Projects known evolutionary intermediates into the latent space to reconstruct and extrapolate ancestral trajectories. By tracing the path between ancestral and modern sequences, the model reveals the semantic shifts that occurred during natural evolution. Engineers can then extend these trajectories beyond observed endpoints, generating future-evolved variants with enhanced properties that nature has not yet explored.

Uses the model's learned probability distribution to score the functional plausibility of generated mutants without experimental data. The log-likelihood ratio between wild-type and mutant sequences under the model serves as a proxy for fitness, identifying which semantic perturbations are likely to preserve fold stability while altering function. This pre-screening dramatically reduces the experimental burden of variant characterization.

A dense, fixed-length vector representation of a protein sequence learned by a language model. Semantic mutagenesis operates directly on these embeddings, treating the high-dimensional space as a continuous map where linear interpolation between two points generates a smooth morphing of sequence properties. The quality of mutagenesis depends entirely on how well the embedding captures structural and functional constraints.

A conceptual mapping of all possible protein sequences to their biological fitness. Semantic mutagenesis provides a learned, differentiable proxy for this landscape. By navigating the latent space, engineers can traverse valleys of non-function to discover high-fitness peaks that are evolutionarily distant from any natural starting point, effectively bypassing the dead ends of random mutagenesis.

A metric quantifying how surprising a sequence is under the model's learned distribution. In semantic mutagenesis, perplexity acts as a quality filter:

• Low perplexity: sequence is plausible and likely folds
• High perplexity: sequence violates learned grammar Engineers use this to reject hallucinated sequences that drift into non-physical regions of latent space during perturbation.

The use of protein language models to score mutations without task-specific training. Semantic mutagenesis extends this concept from discrete point mutations to continuous latent perturbations. By measuring the likelihood shift as a representation is moved, the model implicitly predicts the functional impact of complex, multi-residue changes that would be combinatorially impossible to test experimentally.

An autoregressive generative model that produces novel protein sequences. While ProtGPT2 samples from the learned distribution, semantic mutagenesis provides a more controlled alternative by directly manipulating the latent space. This allows for targeted property optimization—such as increasing thermostability while preserving catalytic activity—rather than relying on stochastic generation and post-hoc filtering.

The task of predicting an amino acid sequence from a backbone structure. Semantic mutagenesis complements inverse folding by operating in the opposite direction: it explores sequence space to discover new backbones. When combined, these techniques enable a closed-loop design cycle where latent perturbations propose sequences, and inverse folding models like ProteinMPNN validate their structural plausibility.