Antibody Deep Mutational Scanning

A DMS library systematically introduces every possible single amino acid substitution at every position in the antibody variable domain. For a 200-residue scFv, this generates a library of ~3,800 unique variants (19 substitutions × 200 positions). This exhaustive coverage contrasts sharply with traditional alanine scanning or sparse mutagenesis, ensuring no epistatic interactions are assumed and every residue's contribution to function is measured independently.

The variant library is expressed on the surface of yeast, phage, or mammalian cells and subjected to a functional pressure—most commonly binding to a fluorescently labeled antigen at a defined concentration. Fluorescence-activated cell sorting (FACS) partitions variants into binding and non-binding populations. The enrichment or depletion of each variant is quantified by deep sequencing the pre- and post-selection populations, yielding a precise enrichment ratio that serves as a proxy for binding affinity.

The output of a DMS experiment is a heatmap of mutational tolerance across the entire sequence. Each cell in the matrix represents the functional score of a single substitution. These landscapes reveal:

• Hotspots: Positions where most mutations are deleterious, indicating functional or structural constraints.
• Tolerant regions: Positions permissive to mutation, often in framework regions or non-paratope loops.
• Beneficial mutations: Rare substitutions that enhance binding, providing direct candidates for affinity maturation.

DMS datasets are the gold-standard ground truth for training machine learning models that predict the effect of unseen mutations. Because the data is dense (nearly all single mutants measured) and quantitative, it enables the training of:

• Antibody-specific language models fine-tuned on DMS fitness scores.
• Structure-based graph neural networks that learn the relationship between local structural environments and mutational tolerance.
• Generative models that propose novel sequences with high predicted fitness, validated against the DMS landscape.

Advanced DMS platforms simultaneously assay multiple antibody properties in a single experiment. By using orthogonal fluorescent labels or distinct selection conditions, a single library can be probed for:

• Binding affinity to the target antigen.
• Specificity against off-target homologs.
• Thermal stability via protease susceptibility or heat challenge.
• Expression yield via epitope tag quantification. This multiplexing generates a multi-dimensional fitness landscape that is essential for multi-objective optimization.

A critical application of antibody DMS is viral escape mapping. By expressing all single amino acid mutations of a viral glycoprotein (e.g., SARS-CoV-2 RBD) and selecting against a therapeutic antibody, researchers identify every mutation that reduces neutralization. This pre-emptive map of escape mutations informs:

• Antibody cocktail design to minimize overlapping escape profiles.
• Pandemic surveillance by flagging circulating mutations that match escape profiles.
• Proactive engineering of antibodies that target conserved, mutationally constrained epitopes.

The primary machine learning task enabled by deep mutational scanning data. Supervised models are trained on DMS-derived fitness landscapes to predict the functional consequence of any single amino acid substitution from sequence alone. Key architectures include:

• Antibody language models fine-tuned on DMS labels
• Geometric neural networks incorporating 3D structural context
• Ensemble methods combining evolutionary and experimental scores These predictors are used to guide in silico affinity maturation and escape mutation forecasting.

Deep mutational scanning constructs a quantitative genotype-phenotype map for an antibody, revealing the local fitness landscape around a starting sequence. Each position in the antibody is assigned a substitution tolerance profile, identifying:

• Invariant residues critical for structural integrity or binding
• Mutable hotspots where mutations enhance affinity
• Epistatic interactions where the effect of one mutation depends on another This landscape serves as the ground-truth training signal for generative antibody design models.

A specialized DMS application where an antibody's neutralization capacity is measured against a comprehensive library of antigen variants. This identifies viral escape mutations that abrogate binding. The resulting data trains escape predictors used to:

• Assess the evolutionary durability of therapeutic antibody candidates
• Design cocktails of antibodies targeting non-overlapping escape profiles
• Anticipate pandemic virus evolution for proactive vaccine design This is a critical component of pandemic preparedness pipelines.

The broader experimental framework encompassing antibody DMS. MAVEs couple high-throughput mutagenesis with a functional selection or screen and deep sequencing to quantify variant abundance pre- and post-selection. The enrichment ratio for each variant serves as a proxy for its functional score. Critical quality control steps include:

• Establishing replicate concordance
• Normalizing for sequencing depth
• Correcting for library composition biases Rigorous MAVE design ensures the resulting datasets are fit for training high-stakes predictive models.

Transformer-based neural networks pre-trained on hundreds of millions of natural antibody sequences. These models learn the statistical grammar of immune repertoires, capturing constraints from evolution and biophysics. When fine-tuned on DMS data, they become powerful zero-shot variant effect predictors, generalizing to unseen sequences. Leading examples include AntiBERTy, IgLM, and AbLang. They represent the convergence of unsupervised representation learning and high-throughput experimental data for antibody engineering.

While single-mutant DMS scans are comprehensive, they miss epistatic interactions where mutations have non-additive effects. Advanced DMS designs incorporate combinatorial libraries to probe pairwise and higher-order interactions. The resulting data trains models that capture:

• Positive epistasis where two mutations synergistically enhance binding
• Negative epistasis where individually benign mutations combine to destabilize the antibody Understanding epistasis is essential for navigating the combinatorial explosion of multi-mutation designs in affinity maturation campaigns.