Variant Library
Design

Variant library design is the foundation of directed evolution and screening campaigns. The central challenge is that sequence space is astronomically large—even a 10-residue stretch with all 20 amino acids at each position represents 20^10 (over 10 trillion) possible sequences. No experimental screening technology can sample this space exhaustively, so the design of the library determines whether you find improved variants or waste screening capacity on non-functional sequences.

Combinatorial libraries introduce diversity across multiple positions simultaneously using degenerate codons (NNK, NNS) or trinucleotide phosphoramidite (TRIM) synthesis. These approaches maximize diversity but often include a large fraction of non-functional variants—sequences with stop codons, frameshifts, or destabilizing combinations. Rational design restricts diversity to positions identified through structural analysis, evolutionary conservation, or functional data, producing smaller libraries with higher functional fractions.

Site-saturation mutagenesis (SSM) targets specific positions for complete amino acid coverage, typically at CDR residues in antibodies or active site positions in enzymes. Single-site SSM libraries are small enough to screen exhaustively, but they miss epistatic interactions between positions. Combinatorial saturation at 2–4 positions simultaneously captures pairwise effects while keeping library sizes within practical screening limits (10^4 to 10^6 variants).

Focused libraries take a middle path: computational analysis identifies a small number of positions (typically 5–15) where diversity is most likely to yield functional improvements, and restricts the amino acid alphabet at each position to residues predicted to maintain structural integrity. This might mean allowing only hydrophobic residues at a buried core position while permitting charged residues at a solvent-exposed contact site. The result is a library of 10^3 to 10^5 variants where 30–70% of sequences are functional, compared to under 5% for fully random libraries.

Machine learning models trained on protein fitness landscapes can predict which variants are likely to be functional before any experimental data is collected. Protein language models (ESM2, ProtTrans) encode evolutionary knowledge from billions of natural sequences and can score variants by pseudo-likelihood—essentially asking how “natural” a given mutation looks in the context of the full sequence. When combined with structure-based scoring (Rosetta energy, ProteinMPNN log-probabilities), these models enable the design of libraries that are both diverse and enriched for functional variants.

The practical output is a defined set of DNA sequences—compatible with gene synthesis or oligo pool assembly—that covers the designed diversity while minimizing redundancy and eliminating predicted non-functional sequences. This reduces screening effort by an order of magnitude and increases the probability of finding improved variants in a single round of selection.