Exhaustive single-point mutagenesis across a protein's entire sequence, coupled to functional selection and deep sequencing to chart the complete fitness landscape
Deep mutational scanning begins with a deceptively simple engineering goal: generate every possible single amino-acid substitution across an entire protein, in one pooled library, so that downstream selection and sequencing can measure the functional consequence of each variant simultaneously. For a protein of length L, exhaustive saturation requires 19×L missense variants plus L nonsense (stop-codon) variants — for a typical 320-residue domain that is 6,080 missense + 320 nonsense = 6,400 designed sequences, every one of which must be represented, uniquely trackable, and free of synthesis bias before selection ever begins.
Two dominant strategies generate saturation libraries. The first, NNK/NNS degenerate-codon mutagenesis, uses primers with a randomized codon (N=any base, K=G/T, S=G/C) at each targeted position during PCR-based or Kunkel mutagenesis; NNK encodes all 20 amino acids plus 1 stop codon using only 32 codon combinations (vs. 64 for NNN), reducing but not eliminating codon-usage bias and enabling multiplexed reactions across many positions in parallel, typically using pooled oligo mixes and Golden Gate or Gibson assembly to reintroduce the mutagenized fragment into a full-length expression construct. The second, and now dominant, strategy is direct oligo-pool synthesis: a company such as Twist Bioscience or Agilent (SurePrint) synthesizes tens of thousands of distinct 200–300 nt oligonucleotides on a DNA microarray, each encoding one designed missense or nonsense variant with defined flanking homology arms, at a cost of roughly $0.01–0.05 per base. This approach eliminates most degenerate-codon bias because every variant is independently specified, and allows combinatorial or non-saturating "rational" libraries (targeting only specific interface residues, for example) as easily as full saturation.
Because functional selection assays typically read out through a short sequence tag rather than the full-length gene, each variant clone must be linked to a random barcode (18–26 nt) integrated adjacent to the coding sequence. This variant-barcode map is established once, prior to selection, using long-read sequencing — PacBio circular consensus sequencing (CCS) or Oxford Nanopore — which reads the full insert plus barcode from thousands of individual molecules and builds a lookup table associating each barcode with its designed variant, typically requiring 3–5 CCS passes per molecule for >99.9% base-call accuracy. All downstream selection timepoints are then sequenced only at the short barcode locus, converting an expensive full-length resequencing problem into a cheap, highly multiplexable amplicon problem.
Library quality control before selection includes: (1) verifying by CCS that ≥90–98% of designed variants are represented at all, since synthesis failures and cloning bottlenecks predictably drop a fraction of rare variants; (2) confirming a roughly log-normal, non-bimodal frequency distribution across variants (Gini coefficient <0.3 preferred) so no handful of clones dominate the pool; (3) transforming or electroporating at ≥30-fold molar excess of cells relative to library complexity, since insufficient transformation efficiency is the single most common cause of an irrecoverable pre-selection bottleneck — once diversity is lost at this step, no amount of sequencing depth downstream can recover it.
A pooled genotype library is only informative once it is subjected to a phenotypic assay that differentially enriches or depletes variants according to function. The choice of selection modality — growth-based, display-based, or abundance-based — determines exactly what "fitness" means in the resulting dataset: catalytic activity, binding affinity, protein stability, or some composite thereof, and each requires its own bottleneck-control discipline to keep sampling noise below the true biological signal.
Growth-based (competitive) selection links variant function directly to organismal fitness: a variant that restores an essential activity (e.g., a drug-resistance enzyme such as APH(3')-II kinase, or a complementing metabolic gene in an auxotrophic strain) allows the host cell to grow or survive antibiotic/chemical challenge, while loss-of-function variants are outcompeted or killed. Populations are sampled at an initial (pre-selection, T0) timepoint and after one or more rounds of competitive outgrowth (T1…Tn, typically 4–10 generations apart); the fold-change in each variant's frequency directly estimates its relative fitness.
Display-based selection is used when the phenotype of interest is binding affinity or protein abundance/stability rather than organismal growth. In yeast surface display DMS (the method underlying the Bloom lab's SARS-CoV-2 receptor-binding-domain scans), each variant is expressed as an Aga2p fusion on the yeast cell wall, incubated with a fluorescently labeled ligand (e.g., soluble ACE2 for the RBD-ACE2 interaction), and cells are sorted by fluorescence-activated cell sorting (FACS) into 4–6 log-spaced fluorescence bins. Variant frequency across bins is reconstructed by sequencing each bin separately and fitting a mean fluorescence/Kd estimate per variant from the bin distribution. VAMP-seq (Variant Abundance by Massively Parallel sequencing, Matreyek et al. 2018) uses an analogous FACS strategy but couples the variant to a fluorescent reporter whose steady-state level tracks cellular protein abundance, directly measuring folding stability for proteins such as PTEN and TPMT.
Regardless of format, the critical experimental discipline is bottleneck control: the effective population size carried through selection must stay at least 100–200× the library complexity at every timepoint, since sampling fewer cells/particles than that introduces genetic-drift noise indistinguishable from true fitness signal. A 6,400-variant library therefore requires sorting or growing at least 1.3–2.5 million cells per replicate per timepoint. Three or more independent biological replicates (separate transformations carried through selection and sequencing independently) are standard, both to estimate measurement noise directly and to flag variants whose apparent fitness is driven by a single outlier clone rather than reproducible biology.
Every selection timepoint is converted into a quantitative variant-frequency table by high-throughput amplicon sequencing of the barcode locus. Because fitness is inferred purely from the ratio of counts before and after selection, sequencing depth and error correction directly set the statistical resolution of the entire experiment — an under-sequenced variant cannot be distinguished from a moderately deleterious one, and PCR amplification artifacts can masquerade as real enrichment.
The short barcode locus from each selection timepoint (input pool plus every sorted bin or growth output) is PCR-amplified with sample-specific Illumina indices and pooled for sequencing on a NovaSeq or NextSeq instrument, typically as 150 bp paired-end reads, targeting 500–2,000 reads per variant per timepoint — for a 6,400-variant library sampled at 4 timepoints across 3 replicates, that is on the order of 40–150 million total reads per experiment, comfortably within a single mid-output NovaSeq SP flow cell. Before amplification, an 8–12 nt unique molecular identifier (UMI) is frequently incorporated adjacent to the barcode; because early PCR cycles can stochastically amplify one template molecule far more than others ("PCR jackpotting"), collapsing reads sharing a UMI down to a single count corrects for this artifact and prevents a handful of over-amplified molecules from distorting the frequency table.
Raw reads are processed through a computational pipeline — commonly Enrich2 (Rubin et al. 2017), dms_tools2, DiMSum, or mutscan — that (1) demultiplexes reads by sample index, (2) matches each barcode read against the PacBio-derived variant-barcode lookup table (allowing 1–2 mismatches to tolerate sequencing error), (3) collapses UMI duplicates, and (4) tabulates a raw count per variant per timepoint. Variants falling below a minimum count threshold at the pre-selection timepoint (commonly ≥10–20 reads) are flagged as insufficiently sampled and excluded from downstream fitness calculation rather than reported with spuriously extreme scores. Finally, pairwise Pearson correlation of variant frequencies between biological replicates is computed as a hard quality gate — r > 0.85 is the typical publication threshold — below which the entire selection round is repeated rather than the analysis proceeding on unreliable data. Concordant replicates are then averaged (often as a precision-weighted mean) to produce the final per-variant read-count table that feeds fitness scoring.
The defining output of a deep mutational scan is a single number per variant — a fitness score — assembled across every position and every possible substitution into the position-by-amino-acid heatmap that the field calls the "fitness landscape." Converting raw counts into this comparable, normalized score requires careful anchoring against internal controls, because raw enrichment ratios are confounded by differences in sequencing depth, generation number, and selection stringency between experiments.
For each variant, an unnormalized fitness estimate is computed as the log2 ratio of its frequency after selection to its frequency before selection, log2[(count_post/total_post)/(count_pre/total_pre)], averaged across replicates and (for growth assays) divided by elapsed generation number to express fitness per generation. This raw score is then linearly rescaled using two internal reference populations present in every saturation library: nonsense (stop-codon) variants, which truncate the protein and are expected to be uniformly non-functional, anchor the null/loss-of-function end of the scale (rescaled to approximately −3 to −4 in most conventions); synonymous or wild-type-matched codon variants, expected to behave identically to true wild type, anchor the neutral point at exactly 0. Missense variants then fall continuously between these anchors — mildly destabilizing substitutions cluster near −0.5 to −1.5, strongly deleterious substitutions approach the nonsense-control floor, and a minority of substitutions at some positions score above 0, indicating a gain of function or stability relative to wild type.
Assembling every position's 19 (or 20, including stop) substitution scores into a single matrix — positions along one axis, substituted amino acid identity along the other, color-mapped from deep red/loss-of-function through neutral gray to gold/gain-of-function — produces the canonical DMS heatmap. Visual inspection of this matrix is itself diagnostic: vertical stripes of intolerance mark positions where almost every substitution is deleterious (typically buried hydrophobic core residues or catalytic residues), while horizontal or patchy tolerance marks surface loops accommodating nearly any substitution. Because generating a full experimental DMS dataset costs an estimated $10,000–50,000 in library synthesis, selection, and sequencing per protein, these matrices are also the primary ground truth used to benchmark zero-shot computational variant-effect predictors — ESM-1v achieves Spearman ρ ≈ 0.4–0.5 against experimental DMS scores across the ~250 curated assays in the ProteinGym benchmark, EVE (an evolutionary variational autoencoder) reaches ρ ≈ 0.5–0.55 on proteins with sufficient homologs, and ensemble methods combining evolutionary and structural features are approaching ρ ≈ 0.6, though no purely computational method yet fully substitutes for direct experimental measurement.
The completed fitness landscape becomes maximally useful once it is overlaid onto the protein's three-dimensional structure and connected to downstream decisions: which variants are clinically pathogenic, which residues to target for stabilizing directed evolution, and which viral mutations are likely to emerge next under immune or drug pressure. This final interpretive step is where deep mutational scanning transitions from a wet-lab dataset into a decision-making tool used across clinical genetics, protein engineering, and infectious-disease surveillance.
Overlaying per-position mean fitness scores onto a solved or predicted structure (via PyMOL or ChimeraX, colored on a per-residue basis) reveals a strong, physically interpretable pattern: relative solvent accessibility (RSA) correlates with substitution tolerance at r ≈ 0.6–0.75 in most well-folded domains, because buried hydrophobic-core positions are constrained by tight packing geometry while surface-exposed loops tolerate nearly any side chain. Positions that break this expected pattern — buried yet tolerant, or exposed yet highly intolerant — are frequently functionally special: catalytic residues, allosteric hotspots, or protein-protein interface residues, and are flagged for closer mechanistic follow-up.
In clinical genetics, DMS-derived functional scores now feed directly into variant classification under ACMG/AMP guidelines. Findlay et al. (Nature 2018) used saturation genome editing (SGE) — a variant of DMS performed by CRISPR-mediated variant knock-in directly at the endogenous BRCA1 locus in haploid cells, followed by a cell-fitness readout — to functionally score essentially all possible single-nucleotide variants across BRCA1's critical exons, generating function scores for ~4,000 variants including hundreds previously listed as "variants of uncertain significance" (VUS) in ClinVar. Because a strongly loss-of-function score is concordant evidence for pathogenicity under the ACMG PS3 criterion, this single dataset enabled confident reclassification of a substantial fraction of previously ambiguous BRCA1 VUS, directly changing clinical risk counseling for carriers. The same logic has since been applied to TP53, PTEN (via VAMP-seq abundance scanning), and dozens of other disease-relevant genes, and large DMS compendia (MaveDB, ProteinGym) are now used as the primary ground truth for training and validating genome-scale pathogenicity predictors such as AlphaMissense.
In infectious disease, DMS performed directly on viral proteins forecasts evolutionary trajectories before they are observed in circulating strains. Starr et al. (Science 2020, Bloom lab) mapped the fitness effect of every possible single substitution in the SARS-CoV-2 spike receptor-binding domain (RBD) on both ACE2 binding affinity and antibody escape, using yeast surface display DMS as described in Stage 2. Positions and substitutions found to preserve or enhance ACE2 affinity while escaping neutralizing antibody classes constituted a predictive map of where future variants of concern were likely to arise. Finally, in protein engineering, DMS-identified stabilizing or activity-enhancing single substitutions are used directly as validated starting points for combinatorial directed evolution — rather than randomly recombining uncharacterized mutations, engineers combine only substitutions independently shown by the scan to be neutral-or-beneficial, sharply increasing the hit rate of multi-mutant libraries built in subsequent rounds.
The Bloom lab's 2020 RBD deep mutational scan measured ACE2-binding and antibody-escape fitness for all 3,804 possible single amino-acid substitutions across the receptor-binding domain's 201 positions — before most of these substitutions had ever been observed in a sequenced viral genome. Several substitutions later found in Omicron and other variants of concern, including changes at position 501 and in the 484 loop, scored among the top ACE2-affinity-preserving, antibody-escaping mutations in the original 2020 dataset, months to years before natural selection surfaced them in circulating lineages — a striking demonstration that saturation fitness landscapes can anticipate evolution rather than merely explain it retrospectively.