Sequence-Function Analysis

Every protein sequence carries information about what it does and how it does it. Residues critical for function—catalytic sites, binding interfaces, structural cores—are conserved across evolution because mutations at those positions are deleterious. Residues on solvent-exposed loops tolerate more variation. By aligning a protein against its homologs across species, you can map which positions matter and which are free to change.

Multiple sequence alignments (MSAs) are the foundation of this analysis. Tools like HMMER, Clustal Omega, and MUSCLE align hundreds or thousands of related sequences, revealing per-position conservation scores. Highly conserved columns (Shannon entropy near zero) typically correspond to functional or structural essentials. Variable columns suggest positions where mutations are tolerated—and therefore where engineering is more likely to succeed without disrupting fold or function.

Beyond single-position conservation, coevolution analysis examines correlated mutations between positions. If residue A and residue B consistently mutate in concert across the phylogeny, they likely make direct physical contact or participate in the same functional network. Methods like Direct Coupling Analysis (DCA) and EVcouplings extract these pairwise signals and have proven remarkably effective at predicting three-dimensional contacts directly from sequence data—a principle that AlphaFold itself leverages internally.

Protein language models like ESM2 and ProtTrans have transformed sequence-function analysis. Trained on billions of protein sequences, these models learn contextual representations that encode structural and functional information without requiring explicit alignments. A single forward pass through ESM2 produces per-residue embeddings that capture evolutionary context, secondary structure propensity, and interaction potential.

These embeddings can be used to train lightweight predictors for specific properties: binding affinity, thermostability, solubility, enzyme activity. The approach is especially powerful for protein families with limited experimental data, where traditional QSAR or Gaussian process models underperform due to sparse training sets. By leveraging the language model's pre-trained representations, you can build effective predictors from as few as 50–100 labeled examples.

In a practical engineering workflow, sequence-function analysis guides library design by identifying which positions to diversify, which to hold constant, and which combinations of mutations are likely to be additive. It also enables zero-shot variant scoring—predicting the functional impact of unseen mutations using the language model's log-likelihood ratio—which can dramatically reduce the size of a screening library while retaining the best candidates.