When a computer judge grades text answers, longer answers win even when they say nothing more — automated evaluators, and even human annotators, use length as a shortcut for quality. I propose to test whether the same shortcut corrupts the computer systems now being built to design antibody drugs against tumor-specific glycoepitopes — the sugar-and-protein composite structures that appear on cancer cells but not healthy ones. In molecular design, the suspicious "length" is not word count but interface size: the length of the antibody binding loop, the number of predicted contacts, and especially the number of contacts the binder makes to the attached carbohydrate chain. My core claim is that learned evaluators for glycoepitope binders — affinity scorers, reward models over antibody–antigen pairs, epitope predictors — systematically conflate interface quantity with binding quality, where quality means the one property that makes these drugs safe: high affinity to the tumor glycoform and no affinity to the healthy glycoform of the same protein. And I propose to port the full diagnostic-and-mitigation toolkit built for text verbosity bias — attribute perturbation, counterfactual data augmentation, length-controlled scoring — to this molecular setting, where it has never been applied.
The biological stakes make the conflation dangerous rather than merely inconvenient. Tumor-specific glycoepitope targeting works because an antibody can be made to require the simultaneous presence of a specific peptide segment and a specific attached glycan, so it binds malignant cells and spares healthy cells carrying the identical protein. Wang et al. demonstrated exactly this with MW11-h317, an antibody whose binding strictly requires PD-1's glycosylation at Asn58 while retaining full checkpoint-blocking activity Identification of a monoclonal antibody that targets PD-1 in a manner requiring PD-1 Asn58 glycosylation; Canals Hernaez et al. did it with PODO447 on podocalyxin, sparing kidney and vasculature PODO447: a novel antibody to a tumor-restricted epitope on the cancer antigen podocalyxin and then, as an armed conjugate, achieving complete regressions without detectable systemic or renal toxicity Targeting a Tumor-Specific Epitope on Podocalyxin Increases Survival in Human Tumor Preclinical Models; Liu et al. did it with NXA01, which reaches CD24 only where hypoglycosylation has removed the healthy tissue's steric glycan shield 1364 Discovery of NXA01, a novel CD24 antibody with tumor selectivity and potent anti-tumor activity. The reason these took years each is that no computational pipeline can search for such binders at scale: as the open-problem surveys in this literature note, generative antibody design tools operate on naked amino-acid sequences and ignore glycans entirely. The gap is starting to close — EvoStruct adapts protein language models to design antibody binding loops EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation, and ConTact designs them through explicit interface reasoning ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning — and the moment such pipelines become glycan-aware, their internal evaluators become the safety-critical component. That is precisely where the text-alignment literature predicts a failure mode: Kim et al. showed that reward models grab length as an easy statistical shortcut Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization, and Sungdong Kim et al. showed that optimizing against such proxies raises nominal reward while true utility falls Rethinking the Role of Proxy Rewards in Language Model Alignment.
Core Claim
Learned molecular evaluators trained on ordinary affinity data will prefer binders with larger interfaces and more glycan contacts over compact binders, and will fail to distinguish tumor glycoforms from healthy glycoforms, because contact count and glycan-contact density are cheap surrogate features that correlate with affinity in training data but are orthogonal to the dual-specificity that defines a safe glycoepitope drug. This is a direct structural analog of verbosity bias, and the text literature's fixes transfer: matched counterfactual pairs that isolate the confounded attribute, augmentation with deliberately "padded" decoys, and length-controlled score correction. The claim is falsifiable: if glycoform-swapped evaluation pairs show no preference gap beyond what affinity labels explain, the bias does not exist and the mitigation is moot.
Research Program
The program has three parts. First, build a molecular counterfactual benchmark. Start from the three structurally characterized dual-specificity systems above (PD-1 Asn58, podocalyxin PODO447, CD24 NXA01) plus the MUC1 tandem-repeat glycopeptide epitopes whose crystallography is reviewed by Zhou et al. Epitopes of MUC1 Tandem Repeats in Cancer as Revealed by Antibody Crystallography: Toward Glycopeptide Signature-Guided Therapy. For each, construct two families of matched pairs: binder-padding pairs, where a binder is rewritten with a longer CDR-H3 loop or extra glycan-contact residues while the paratope core is held fixed — the molecular analog of the rewrites Bharadwaj et al. use to diagnose length and structure biases in preference models Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models — and glycoform-swap pairs, where the identical peptide surface carries the tumor glycan versus the healthy glycan. CDR-H3 length is the right attribute to perturb because it is already known to be a heavily skewed, biologically loaded variable in human repertoires Heavy Chain CDR 3 and Junctional Length Biases in Human Antibody Repertoires Associated with Heavy and Light Chain Germline Utilization and because protein language models for antibody design are demonstrably biased toward germline templates with characteristic loop lengths Addressing the antibody germline bias and its effect on language models for improved antibody design, so an evaluator that reads loop length as quality has a ready-made shortcut. The counterfactual designs are grounded in the fact that canonical CDR conformations span loop lengths independently Length-independent structural similarities enrich the antibody CDR canonical class model, so we can vary length while preserving structural class.
Second, diagnose. Evaluate existing and trainable scorers — a docking/physics scorer, a protein-language-model affinity scorer in the EvoStruct style, and a trained reward model over antibody–glycoepitope pairs — on both pair families using an accuracy-parity metric adapted from Saito et al., who measured how often an LLM judge prefers a longer response when the ground-truth preference goes the other way (GPT-4 scored 0.328 on HH-RLHF; GPT-3.5 scored 0.428) Verbosity Bias in Preference Labeling by Large Language Models. The molecular parity score measures whether the evaluator prefers the padded binder and whether it keeps scoring a binder highly after the tumor glycan is swapped for the healthy glycan — the exact situation that produces on-target, off-tumor toxicity, which Kuroda et al. flag as the central validation burden for any glycoform-based target Prostaglandin F2 receptor negative regulator as a potential target for chimeric antigen receptor-T cell therapy for glioblastoma. A secondary diagnostic leg asks whether epitope predictors share the bias: SEPPA 3.0 established that N-glycosylation sites are statistically enriched in epitope regions and reduced false-positive rates from 0.34 to 0.25 through spatial calibration SEPPA 3.0 - enhanced spatial epitope prediction enabling glycoprotein antigens, and I will test whether its residue scores — and any learned successor — inflate on glycan-dense surfaces irrespective of glycoform tumor-relevance.
Third, mitigate. Port three text mitigations and compare them head-to-head. Counterfactual data augmentation in the style of Bharadwaj et al., who cut reward-model miscalibration from 39.4% to 32.5% and skew difference from 20.5% to 10.0% on text while preserving RewardBench performance Flattery, Fluff, and Fog: add synthetic rejected binders with injected padding and with healthy-glycoform targets. Response-conditioned modeling in the style of Rc-BT, which fixes the response and varies the prompt's length constraint to decouple length compliance from quality Disentangling Length Bias in Preference Learning via Response-Conditioned Modeling: fix the binder and vary the glycoform, forcing the evaluator to score the same paratope against tumor and healthy glycans. And a length-controlled score in the style of Dubois et al., who debiased AlpacaEval with a generalized linear model over a tanh-squashed length difference Length-Controlled AlpacaEval: A Simple Debiasing of Automatic Evaluators. For a binder $b$ against epitope $s$, the corrected score is
$$r_{\text{LC}}(s, b) = r(s, b) - \beta \cdot \tanh\!\big(L(b) - \bar{L}_s\big),$$
where $r$ is the raw evaluator score, $L(b)$ is the binder's CDR-H3 loop length (or predicted glycan-contact count), $\bar{L}_s$ is the median for the epitope's structural class, and $\beta$ is fit on the diagnostic pairs. This correction must be adaptive, not a flat penalty, because bulky legitimate targets — oncofetal chondroitin sulfate chains of the kind detected pan-cancer by rVAR2 binding Establishment and validation of a plasma oncofetal chondroitin sulfated proteoglycan for pan-cancer detection, or MUC1's extended tandem repeats — genuinely require long loops; this mirrors Huo et al.'s argument that length penalties must adapt to tasks whose correct answers are naturally long Beyond Excess and Deficiency: Adaptive Length Bias Mitigation in Reward Models for RLHF, and the over-correction failure mode Rc-BT documents (models collapsing into uninformative brevity) maps onto binders collapsing into minimal loops that cannot bridge peptide and glycan simultaneously.
Finally, close the loop: run a ConTact/EvoStruct-style generative CDR design pipeline against glycan-aware epitope inputs, with raw versus debiased evaluators, and measure whether debiasing shifts the generated distribution toward tumor-selective binders.
Baselines
The naive reference point is an unmodified affinity-trained reward model or protein-language-model scorer — "do nothing" about interface size. It is the strongest competitor precisely because it is the current default: every existing antibody-design pipeline scores candidates by predicted affinity or interface quality alone, and if affinity-based scoring already yields tumor-selective binders at the top of the ranking, the proposed debiasing machinery is unnecessary complexity. I will train and run it on exactly the same paired-affinity and structural data, with the same model capacity, tuning budget, and compute as every debiased variant, so the only difference is the correction. Beating it means: on glycoform-swap pairs, the corrected evaluator's preference accuracy for the tumor glycoform exceeds the raw evaluator's by a margin outside the bootstrap confidence interval, with no loss of raw affinity-prediction performance on held-out affinity data.
The strongest existing mitigation baseline is counterfactual data augmentation applied naively — adding long-loop decoys as rejected candidates without the glycoform-swap counterfactuals. This is the direct transplant of the method that worked best in text, where it reduced miscalibration from 39.4% to 32.5% Flattery, Fluff, and Fog, and if it transfers unchanged it should be preferred on simplicity grounds. It will get the same data budget and tuning. Beating it requires showing that CDA alone suppresses padding preference but leaves glycoform-blindness — predicted, since padded decoys never teach the evaluator what a healthy glycan looks like — and that adding glycoform-swap counterfactuals closes that specific gap.
For the epitope-prediction leg, the strongest baseline is SEPPA 3.0 itself, which on its external glycoprotein antigen set (106 antigens with verified N-glycosylation sites) reports AUC 0.749 and balanced accuracy 0.665, with calibration cutting false positives from 0.34 to 0.25 SEPPA 3.0. It is the incumbent because it is the only published glycan-aware epitope predictor; the proposal's claim on this leg is not that I will beat its residue-level AUC but that I will show whether its scores, and learned successors trained on the same 767-antigen corpus, carry a glycan-density bias analogous to verbosity bias — a property SEPPA's spatial calibration was never designed to test. The key ablation isolates the proposal's central ingredient: an evaluator trained with binder-padding pairs only versus one trained with both padding pairs and glycoform-swap pairs, holding everything else fixed. If the glycoform-swap pairs are what repair selectivity scoring, the core claim stands; if padding pairs alone suffice, the molecular analog of verbosity bias is real but the glycoform component is the wrong story.
Why not just…?
Why not just use a physics-based docking score instead of a learned evaluator? Physics scorers have no trained preferences, so they cannot "learn" length as a shortcut — but they encode an equally crude structural prior: more predicted contacts and larger buried interface area generally score higher, and they are entirely glycoform-agnostic unless glycans are explicitly modeled, which standard pipelines do not do. The diagnostic experiments directly settle this: the binder-padding and glycoform-swap pairs will be scored by a physics scorer alongside the learned models, and if the physics scorer shows no parity gap on either pair family, it wins on robustness grounds and the proposal's mitigation effort should redirect to making physics scoring glycan-aware. I consider this the most serious threat to the proposal and have made it a first-class arm of every experiment rather than a footnote.
Why not just normalize the score by interface size? This is the molecular SimPO — dividing a summed sequence reward by length fixed reward hacking in DPO-style training SimPO: Simple Preference Optimization with a Reference-Free Reward — and it is the first thing I will try. The text literature already documents its failure modes, which is why I do not expect it to be sufficient: flat normalization over-corrects into uninformative brevity (the collapse Rc-BT observed Disentangling Length Bias in Preference Learning via Response-Conditioned Modeling), and some settings genuinely need length, as Huo et al. argue for tasks with naturally long answers Beyond Excess and Deficiency. The molecular analog is sharp: MUC1 tandem-repeat and oncofetal chondroitin sulfate epitopes need long loops, so a flat penalty discards exactly the binders those targets require. Per-contact normalization is included as a baseline; beating it means achieving equal padding suppression without degrading predicted affinity on bulky-glycan targets, and the $\tanh$-form length-controlled score is designed for that.
Why not just screen the final binders experimentally for selectivity? Wet-lab selectivity screening is how MW11-h317, PODO447, and NXA01 were actually found, and nothing here replaces it. But screening operates at the end of a funnel whose top is now generative: if the evaluator ranks a thousand hacked, glycan-greedy-but-unselective binders above the hundred genuinely dual-specific ones, the screen sees the wrong molecules. The reward-hacking literature gives the quantitative warning — nominal proxy reward rises while true utility falls Rethinking the Role of Proxy Rewards in Language Model Alignment — and the proposal's decisive downstream metric, enrichment of tumor-selective binders within the top-scoring decile, measures funnel efficiency that end-stage screening cannot recover. This proposal is a claim about where in the funnel selectivity signal must enter, not about skipping validation.
Why not just make the generative models glycan-aware and skip the bias study? Making the representation glycan-aware — feeding glycans into the generator — is the obvious combination and, per the literature survey, not yet done; EvoStruct and ConTact EvoStruct, ConTact still design on naked sequences. But a glycan-aware generator with a biased evaluator is the worst case: the evaluator can then exploit glycan-contact count as a shortcut feature, exactly as text reward models exploit token count. The bias diagnosis is what makes the glycan-aware upgrade safe rather than merely faster at producing plausible-looking failures.
Proposed Experiments and Decisive Checks
Experiment one is the diagnostic. Trained on paired-affinity data from public antibody–antigen structures, evaluate physics, PLM, and reward-model scorers on the matched counterfactual pairs, reporting the molecular accuracy-parity score alongside Saito et al.'s text reference points of 0.328 (GPT-4) and 0.428 (GPT-3.5) on HH-RLHF Verbosity Bias in Preference Labeling by Large Language Models. The decisive number is the glycoform-swap parity gap: if no scorer exceeds chance on tumor-versus-healthy glycan discrimination, the core claim's second half is confirmed and the mitigation is justified; if all scorers are already near-perfect, the proposal's premise fails and should be abandoned. Experiment two is the mitigation bake-off: raw evaluator, per-contact normalization, padding-only CDA, padding-plus-glycoform CDA, the response-conditioned variant, and the adaptive length-controlled score, all at matched data and compute, judged on three axes — padding-parity reduction, glycoform-swap accuracy, and preserved affinity prediction, with the bulky-glycan targets as the over-correction guard, in the spirit of Huo et al.'s excess-and-deficiency framing Beyond Excess and Deficiency. Experiment three closes the loop: generate CDR designs for the PD-1 Asn58, podocalyxin, and MUC1 systems with raw versus best-debiased evaluators and report the fraction of tumor-selective binders in the top-scoring decile, the analog of the enrichment at which the RLHF overoptimization literature predicts collapse Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization. Experiment four is the epitope-prediction leg: retrain a SEPPA-style residue scorer on its own corpus and test for glycan-density inflation on glycoform-annotated external sets, benchmarked against SEPPA 3.0's reported external glycoprotein AUC of 0.749 and balanced accuracy of 0.665 SEPPA 3.0, with the HIV-1 gp120 case — where SEPPA correctly identified 29 of 30 known epitope residues — as the sanity check that calibration survives. No reported result exists on Lacuna for EvoStruct's or ConTact's designed-binder affinities, so the generative baselines will be run from scratch under matched budgets rather than compared to published figures. The single most decisive artifact is a public benchmark of matched glycoform-swap and binder-padding pairs with ground-truth labels, which lets any group score their evaluator's contact-verbosity bias in an afternoon — the molecular counterpart of the parity metric that made text verbosity bias measurable, and the prerequisite for anyone building the glycan-aware design pipelines this field currently lacks.
References
- Identification of a monoclonal antibody that targets PD-1 in a manner requiring PD-1 Asn58 glycosylation
- PODO447: a novel antibody to a tumor-restricted epitope on the cancer antigen podocalyxin
- Targeting a Tumor-Specific Epitope on Podocalyxin Increases Survival in Human Tumor Preclinical Models
- 1364 Discovery of NXA01, a novel CD24 antibody with tumor selectivity and potent anti-tumor activity
- EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation
- ConTact: Contact-First Antibody CDR Design via Explicit Interface Reasoning
- Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization
- Rethinking the Role of Proxy Rewards in Language Model Alignment
- Epitopes of MUC1 Tandem Repeats in Cancer as Revealed by Antibody Crystallography: Toward Glycopeptide Signature-Guided Therapy
- Flattery, Fluff, and Fog: Diagnosing and Mitigating Idiosyncratic Biases in Preference Models
- Heavy Chain CDR 3 and Junctional Length Biases in Human Antibody Repertoires Associated with Heavy and Light Chain Germline Utilization
- Addressing the antibody germline bias and its effect on language models for improved antibody design
- Length-independent structural similarities enrich the antibody CDR canonical class model
- Verbosity Bias in Preference Labeling by Large Language Models
- Prostaglandin F2 receptor negative regulator as a potential target for chimeric antigen receptor-T cell therapy for glioblastoma
- SEPPA 3.0 - enhanced spatial epitope prediction enabling glycoprotein antigens
- Disentangling Length Bias in Preference Learning via Response-Conditioned Modeling
- Length-Controlled AlpacaEval: A Simple Debiasing of Automatic Evaluators
- Establishment and validation of a plasma oncofetal chondroitin sulfated proteoglycan for pan-cancer detection
- Beyond Excess and Deficiency: Adaptive Length Bias Mitigation in Reward Models for RLHF
- SimPO: Simple Preference Optimization with a Reference-Free Reward