Antibody-Conditioned Likelihood Shifts for Escape-Resistant Epitope Targeting

A language model's most basic output is not text but probability: a number for every possible next token. Researchers who measure social and cultural bias in language models have learned to exploit this directly. Instead of reading what a model generates, they compare the probability the model assigns to a piece of text with and without some context — an identity term, a cultural prompt, a question — and treat the difference as a clean measurement of what that context contributes. The Normalized Probability Shift metric, for instance, subtracts the log-probability of an answer choice given only the other choices from its log-probability given the question too, and this subtraction is what makes the measurement robust: under adversarially rewritten choices, standard accuracy on HellaSwag dropped by 54.23% while the shifted metric moved by less than 0.17% (Choices Speak Louder than Questions).

This proposal brings the same subtraction trick to a problem where it has never been applied: predicting which viral mutations will let a pathogen escape a given antibody, and designing vaccines that focus immune responses onto conserved, hard-to-mutate viral sites. The motivating failure is well documented. When pretrained protein language models — sequence models trained on millions of amino-acid strings — were used zero-shot to score how 3,819 single mutations in the SARS-CoV-2 receptor-binding domain escape any of 247 monoclonal antibodies, the likelihood-based score had "almost no predictive power," because escape is not a property of the virus sequence alone but of the virus–antibody pair (Predicting Immune Escape with Pretrained Protein Language Model Embeddings). The fix suggested by the NLP bias literature is exactly to condition on the missing partner and subtract. We propose to build that conditioned metric, validate it against the strongest baselines, and use it both as an antibody-specific escape predictor and as a "focus" score that tells immunogen designers whether an engineered vaccine candidate presents its transplanted conserved epitope or its own unwanted scaffold surface.

Core Claim

The claim is that a single quantity — the difference between a mutation's likelihood cost in two contexts, with and without the antibody — isolates antibody-specific escape, works without 3D structures, and doubles as a design-time diagnostic for conserved-epitope focusing.

Concretely, let $x_i$ be the wild-type amino acid at antigen position $i$, $a$ a mutant amino acid, and let $\log P(\cdot \mid \mathrm{Ag})$ and $\log P(\cdot \mid \mathrm{Ag}, \mathrm{Ab})$ be per-residue log-probabilities from a model that reads the antigen sequence alone or the antigen and antibody sequences together. Define the escape score as

$$ E(a,i) \;=\; \Delta_{\mathrm{free}}(a,i) - \Delta_{\mathrm{bound}}(a,i), \qquad \Delta_{\mathrm{ctx}}(a,i) = \log P(x_i \mid \mathrm{ctx}) - \log P(a \mid \mathrm{ctx}). $$

A mutation that is nearly free for the virus on its own ($\Delta_{\mathrm{free}} \approx 0$) but costly only when the antibody is present ($\Delta_{\mathrm{bound}}$ large) gets a high escape score: the virus can tolerate it, and it breaks the antibody. This is deliberately the same algebra that makes the NPSQ metric robust in Choices Speak Louder than Questions — subtract the unconditioned baseline so that intrinsic, partner-independent effects cancel — and the same thermodynamic-cycle logic that Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions showed is necessary for likelihoods to track binding-energy changes at all. That paper's BA-Cycle achieved an unsupervised Spearman correlation of 0.3201 on SKEMPI v2 precisely by contrasting bound-complex and unbound-chain likelihoods — but it requires rigid 3D backbones and targets general binding affinity, not antibody-specific escape. Our claim is that paired sequence conditioning delivers the same contrast without structures and resolves escape per antibody.

The same machinery yields the second deliverable, an epitope focus score. For an engineered immunogen — a conserved viral loop grafted onto an unrelated protein scaffold, as in the iterative optimization of the influenza HA 220-loop (Iterative immunogen optimization to focus immune responses on a conserved, subdominant viral epitope) — we mutate every solvent-exposed residue to a panel of alternatives and record how much antibody conditioning penalizes each one. Residues whose mutation strongly reduces the antibody-conditioned likelihood are "presentation hotspots." The focus score is the mean hotspot strength on the grafted epitope minus the mean on the scaffold body. A well-focused design should concentrate antibody sensitivity on the epitope; a distracted design spreads it over the scaffold, which is the documented failure mode that forced multiple redesign generations in the 220-loop work.

Research Program

The program has three parts. First, we train a paired-sequence protein language model. The architecture follows A paired sequence language model for protein-protein interaction modeling, whose hybrid intra-chain and cross-chain attention improved binding-affinity prediction by up to 18% in Pearson and Spearman correlation over a structure-based method and raised interface-contact precision from 16.7% to 81.3% on a hard case — evidence that cross-chain attention captures interaction-specific signal that single-chain models miss. Unlike that paper, which stops at classification and contact prediction, we use the model's conditional sequence likelihoods as measurements, never as generators. Training data comes from BindingGYM: A Large-Scale Mutational Dataset Toward Deciphering Protein-Protein Interactions, which provides 0.5 million curated mutational scanning measurements paired with partner sequences and complex structures across 25 assays — enough paired (mutation, partner) supervision to learn the difference between the two contexts in $E(a,i)$. The file reports no baseline numbers on BindingGYM, so we will establish our own split-level reference numbers on it as an intermediate sanity check before touching the escape benchmarks.

Second, we evaluate $E(a,i)$ zero-shot and frozen on the Cao et al. dataset of 247 antibodies against all 3,819 single point mutations of the SARS-CoV-2 receptor-binding domain, the same benchmark used in Predicting Immune Escape with Pretrained Protein Language Model Embeddings. We additionally stratify predictions by structural region to test the biological expectation that conditioned escape potential is depleted in conserved regions (the S2 subunit, the HA stalk) and enriched in variable ones (the receptor-binding domain, the HA head) — the pattern that Learning the language of viral evolution and escape found with escape-agnostic AUCs of 0.692 to 0.856 across five viral deep mutational scanning datasets. Where that paper could only score escape against an anonymous immune backdrop, our per-antibody conditioning directly addresses its stated limitation.

Third, we compute focus scores on the published lineage of HA 220-loop immunogens (HA220-1 through M4T+glyc) from Iterative immunogen optimization, testing retrospectively whether the score rises across design generations and tracks the experimentally observed improvement toward cross-reactive antibodies with nanomolar affinities. As a stretch extension, we use the score's gradient, in the spirit of the Contrastive Affinity Guidance that lifted binding energy by 23% in EpiCoCo, to steer scaffold redesign away from escape-permissive and off-target surfaces during sampling.

Baselines

The do-nothing reference is zero-shot single-sequence likelihood: the log-odds shift that a standard protein language model assigns to each mutation on the antigen alone, exactly the setup that was reported to have "almost no predictive power" on the Cao dataset in Predicting Immune Escape with Pretrained Protein Language Model Embeddings. Any positive mean per-antibody correlation from our conditioned score beats this reference, and re-running it on our splits confirms the negative result reproduces.

The strongest simple competitor is the Site Model from the same paper, which predicts escape from the antigen residue index alone, ignoring the amino acid and the antibody, and which the authors report was surprisingly strong and competitive with embedding models in per-antibody settings; no reported result found on Lacuna for its exact AUC values, so we will run it ourselves on identical splits with an identical tuning budget (a handful of hyperparameters, matched search space). It is the strongest competitor on the within-antibody prediction question because position-level mutability captures most of the static conserved-versus-variable structure of the antigen. We beat it only if our score achieves a higher mean per-antibody Spearman correlation with measured escape, and decisively so if we also predict escape for antibodies held out entirely from any fitting, which a per-site categorical model cannot do without per-antibody labels.

The strongest existing likelihood-based method is BA-Cycle from Boltzmann-Aligned Inverse Folding, which reported Spearman 0.3201 unsupervised (0.5134 for its supervised BA-DDG variant) on SKEMPI v2. It shares our subtraction logic but needs structures, so we run it on the Cao data using AlphaFold-predicted antibody–antigen complexes at matched compute, and we count ourselves as beating it only if the sequence-only score matches or exceeds its correlation while also applying where backbone rigidity fails. The supervised references — the residue-level recurrent network trained from scratch (RNN Res) and the ESM2 embedding heads from the escape-prediction paper — are run on the same training split as an upper bound; we expect them to be strong and we do not claim to beat them with labels available. Finally, one ablation isolates the key ingredient: the same paired model scored without the $\Delta_{\mathrm{free}}$ subtraction, and the subtraction computed with a single-chain model instead of a paired one. If the full combination does not beat both partial versions, the proposal's central mechanism is not what is doing the work.

Why not just…?

Why not just use conservation scores or the Site Model? This is the most serious objection, and the file's evidence supports it: the Site Model was competitive with far heavier models, and a from-scratch RNN matched multi-billion-parameter ESM2 representations (Predicting Immune Escape with Pretrained Protein Language Model Embeddings). Conservation-based scores, however, are intrinsically antibody-agnostic — they cannot distinguish two antibodies targeting the same antigen at different epitopes, which the Cao dataset's 247 antibodies directly test. The decisive experiment is the cross-antibody split: fit the Site Model on a subset of antibodies, evaluate on antibodies it has never seen labels for. If the conditioned score does not beat it there, the proposal fails on its own terms, and we will report that plainly. Within-antibody, the Site Model may well win, and that is an acceptable outcome for a zero-shot metric.

Why not just train a supervised head on the DMS data? Because that is exactly the regime the existing paper already covers, with the modest conclusion that small supervised models are enough. Our contribution is the regime where labels do not exist: a new antibody, an uncharacterized epitope, or an engineered immunogen that has never been assayed. EscapeMap makes the same point from the other side — its antibody-resilience estimates require linear models fitted to extensive binding data for each antibody, which blocks transfer to novel binders. We do include supervised heads as the labeled upper reference, so the comparison is explicit rather than evaded.

Why not just use physics-based scoring (Rosetta, FoldX, or structure-based cycles)? These are grounded and effective when structures are reliable — the 220-loop immunogens themselves were built with RosettaScripts and ProteinMPNN and reached nanomolar affinity (Iterative immunogen optimization). But Boltzmann-Aligned Inverse Folding documents that static-backbone methods break under mutation-induced conformational change, and EpiCoCo found that rigid-target templating cost roughly 26% in binding energy relative to dynamic co-modeling. The retrospective 220-loop experiment settles this empirically: if Rosetta-style design scores track the generation-by-generation focusing improvements as well as our focus score does, the sequence-only metric is not needed for that application, and we will say so.

Proposed Experiments and Decisive Checks

Experiment one evaluates the escape score on the full Cao et al. grid of 247 antibodies and 3,819 receptor-binding-domain mutations, reporting mean per-antibody Spearman correlation with measured escape and the area under the curve for identifying the top-escaping mutations, with Hie et al.'s 0.692–0.856 AUC range on escape-agnostic DMS as the contextual reference point. Experiment two repeats this with antibodies (and epitope classes) fully held out, which is the decisive check against the Site Model. Experiment three maps per-site mean conditioned escape onto structure and tests statistically for depletion in conserved regions and enrichment in variable ones. Experiment four computes focus scores across the published 220-loop immunogen lineage and correlates them with the experimental focusing outcomes; no reported result found on Lacuna for the individual binding affinities of those variants, so the correlation is assessed against the generation ordering and the qualitative cross-reactivity results reported in the paper. A final diagnostic guards against a subtle confound raised by Distributional Readout: A Memorization Reg regime in Autoregressive Generative Models and EpiBench, both of which show models reciting memorized high-frequency strings rather than reasoning: we compare performance on mutations at widely documented positions against mutations at matched but obscure positions. If the conditioned score collapses on the obscure set, it is reading the training corpus, not the interaction, and the metric should not be trusted for design. Compute is modest by current standards — EpiCoCo trained its full co-modeling system in 7 hours on a single RTX 3090, and its released codebase (EpiCoCo Codebase) offers a practical template for the paired co-modeling component.