Phage display is a workhorse of antibody and peptide discovery: a library of billions of protein variants is physically selected over repeated rounds — binders to a target are kept, the rest washed away, and survivors amplified in bacteria — and modern sequencing now reads millions of variants per round. The problem is that the selection is not clean. Clones can survive because they stick to the plastic plate, the blocking protein, or the beads rather than the target, or because mutations in the viral genome make them replicate faster in bacteria. These parasites, called target-unrelated peptides (TUPs), are well known, and tools exist to flag the known ones after the fact. What nobody has measured is the second, quieter failure: when we train neural networks on this contaminated count data to predict binding affinity, do the networks themselves amplify the parasitic signals beyond what was in the data?
This is exactly the question the fairness literature has learned to answer precisely. In vision and language, leakage-based metrics compare how predictable a spurious attribute is from a model's outputs against how predictable it was from the training labels, and the difference is called bias amplification. We propose transplanting that machinery, with a biophysical reinterpretation of each term, into next-generation phage display (NGPD) pipelines. The payoff is concrete: an audit that tells a wet-lab group whether their affinity model is quietly ranking fast-growing cheater clones at the top, and a training fix that suppresses the amplification rather than merely filtering the data.
Core Claim
We claim two things. First, that affinity predictors trained on multi-round phage display counts exhibit measurable parasitic leakage amplification: a probe predicting known parasitic traits from the model's affinity scores will succeed more than the same probe applied to the raw enrichment data, exactly analogous to the gender-leakage amplification demonstrated in image classification by Tianlu Wang et al.. Second, that pre-filtering with existing TUP detectors will not eliminate this amplification, because the network re-learns the parasitic signal from unflagged physicochemical surrogates (net charge, hydrophobic patches, composition biases) — the same mechanism by which balanced vision datasets still produced strongly biased models in that same paper, and by which data-only de-biasing left residual discrimination at 0.09–0.10 in the tabular setting of Lu Zhang et al..
We formalize the measurement as follows. Let $\lambda_D$ be dataset leakage — the area under the ROC curve (AUC) of a probe trained to classify a parasitic trait (plastic-binder, streptavidin-binder, or fast-propagator status, as labeled by the tools in SAROTUP 3.1) from the raw round-to-round count vector of a clone. Let $\lambda_M$ be model leakage — the AUC of the same probe trained on the affinity scores (or the top-layer representation) of an affinity predictor trained on those counts. Then
$$\text{PLA} = \lambda_M - \lambda_D,$$
the parasitic leakage amplification. This is a direct re-instantiation of the $\lambda_M - \lambda_D$ and $LIC_M - LIC_D$ constructions of Tianlu Wang et al. and Yusuke Hirota et al., where the "protected attribute" is no longer gender but non-target parasitic fitness, and the "task label" is true thermodynamic affinity. A positive PLA means the model is a worse filter than the experiment itself; a negative PLA means it actively suppresses parasites. To our knowledge, no reported result exists on Lacuna for any leakage measurement applied to biopanning data.
Research Program
The program has three parts: an audit, a decomposition, and a fix.
The audit applies the PLA measurement to existing affinity-prediction architectures on real NGPD data. We will train a Bayes-PD-style Bayesian neural network on ESM-2 sequence embeddings with a Poisson likelihood over counts — reusing the forward-simulation machinery (upsampling, selection, downsampling) that Bayes-PD built to model experimental noise — plus a simpler baseline affinity regressor. We then measure PLA with respect to each parasitic trait separately (plastic, streptavidin, propagation), because the traits have different physics and different fingerprint sequences, and a single aggregate number would hide which failure mode dominates. The trait labels come from the SAROTUP tool suite's flags on the same libraries, which is exactly the setting where SAROTUP 3.1 says its classifiers are calibrated (Ph.D.-7); we restrict the headline audit to Ph.D.-7 data for label validity and treat generalization to other libraries as an explicit limitation to quantify.
The decomposition separates how much of the residual prediction error is irreducible assay noise versus model-induced. Hao Wang et al. built the FairFront — a theoretical ceiling on accuracy at a given fairness level — but their Blackwell-experiment machinery is tractable only for discrete tabular features, and they explicitly call for harder data. We cannot port the machinery wholesale to sequence space. Instead we build an empirical frontier by replaying the Bayes-PD simulator at varying subsampling depths and bottleneck sizes: for each simulated experimental regime we train the same architecture and record the achievable $K_d$ correlation, so that the observed plateau can be attributed to read depth (aleatoric, per the taxonomy of that paper) versus architecture and embedding choices (epistemic). This gives the community a defensible answer to whether the $-0.46$ correlation ceiling reported by Bayes-PD is a property of the assay or of the model — a question that paper explicitly leaves open. As a complementary lens, we will fit a Fair-IRT-style model in the spirit of Ziqi Xu et al., treating each clone as a "test question" whose difficulty parameter should concentrate on clones with anomalous round-to-round count trajectories — the assay-artifact signature — separating them from clones the model simply fails to score.
The fix is representation-level, not data-level: attach an adversary with a gradient-reversal connection that tries to predict parasitic trait labels from the affinity model's intermediate representation, following the design that reduced bias amplification by 53% to 67% at an F1 cost of 1.2 to 2.2 points in Tianlu Wang et al.. We deliberately train the adversary only on known TUP labels; the hypothesis is that suppressing the parasitic direction in representation space also suppresses the unflagged surrogates, which filtering cannot reach. We will also check for the failure mode Yusuke Hirota et al. exposed in captioning — the Equalizer Paradox, where suppressing bias in one place caused extreme compensatory bias in unconstrained outputs. Here the analog is whether TUP-suppression in the binding head pushes the model to reward developability-hostile traits elsewhere (aggregation-prone or sticky segments outside the optimized region), which we assess by tracking developability-associated sequence statistics in the top-ranked candidates.
Proposed Experiments
E1: Do affinity models amplify parasites? Train the Bayesian neural affinity model and a plain regression baseline on the multi-round panning counts of the three-target dataset used by Bayes-PD, with parasitic labels from the SAROTUP suite. Measure PLA per trait, with bootstrap confidence intervals over clones. A decisive result is PLA significantly above zero for at least one trait. No reported result found on Lacuna for any prior leakage measurement on phage display data, so any nonzero, well-calibrated PLA is itself new evidence.
E2: Does pre-filtering suffice? Train the identical architectures on data from which all SAROTUP-flagged sequences have been removed, then re-measure PLA using held-out, unflagged clones whose parasitic status is inferred from cross-target enrichment in the Biopanning Data Bank (the "Four-Target Rule" signal documented in BDB 2018). The prediction, drawn from the balanced-dataset result of Tianlu Wang et al. and the residual-discrimination result of Lu Zhang et al., is that PLA remains substantially positive after filtering. If PLA collapses to zero after filtering, our second claim is falsified and the simple pipeline wins — we would report that honestly.
E3: Does the adversary fix it without breaking accuracy? Train the gradient-reversal variant under the same compute and tuning budget. Success criteria: PLA reduced by at least the ~50% range achieved in vision, while Spearman correlation with experimental $K_d$ on the validation clones degrades by no more than the 1.2–2.2 point (relative) cost band seen in Tianlu Wang et al.. We also count TUP contamination among the top-100 ranked candidates before and after.
E4: Aleatoric versus epistemic attribution of the correlation plateau. Using the simulation-replay frontier, plot achievable $K_d$ correlation against sequencing depth. If the empirical frontier at the real data's depth sits at roughly $-0.46$ — the value Bayes-PD reports for BFN-designed sequences — the plateau is aleatoric and the field should invest in deeper sequencing, not better models; if the frontier sits far above it, the gap is epistemic and architecture work is warranted. As an architecture sanity check on the scoring side, we will verify our affinity scorer against SKEMPI v2, where supervised Boltzmann Alignment reports a Spearman correlation of 0.5134 (with prior best Surface-VQMAE at 0.4324) per Boltzmann Alignment; no reported result found on Lacuna for a FairFront-style decomposition applied to any biological screening assay.
Baselines
The naive reference point is raw round-to-round enrichment: rank clones by their count fold-change between consecutive rounds. This is what practitioners actually use, and it is the comparator that Bayes-PD already showed to be beaten — its noise-aware model reached a correlation of $-0.46$ with experimental $K_d$ on BFN-generated sequences while raw selectivities performed worse, precisely because a fold-change conflates target affinity with propagation speed. We will run enrichment ranking on the same datasets and report both its $K_d$ correlation and its PLA, so the reference point is characterized on both axes. The strongest existing competitor is the Bayes-PD Bayesian neural network itself: ESM-2 embeddings, Poisson likelihood, and the simulated upsampling-selection-downsampling loop. We will reimplement it faithfully (the file gives no reported result beyond the $-0.46$ figure), run it on identical data splits with the same compute and tuning budget as our variants, and use its reported $-0.46$ as the external anchor. Beating it means improving $K_d$ correlation while reducing PLA — improving one at the expense of the other will be reported as a trade-off, not a win. The third baseline is the SAROTUP pre-filter pipeline: the same Bayes-PD architecture trained on data cleaned by TUPScan and its companion classifiers, which flagged 36,197 unique TUPs out of 3.05 million Ph.D.-7 sequences in the throughput benchmark of SAROTUP 3.1. This baseline isolates data-level cleaning and is the direct test of our second claim. Finally, the key ablation is our own architecture with the gradient-reversal adversary removed — identical embeddings, likelihood, and budget — which isolates whether any PLA reduction comes from the representation-level suppression rather than from architecture or tuning differences.
Why not just…?
Why not just filter out known TUPs with SAROTUP before training? This is the pipeline the field would actually run, and E2 is designed to answer it empirically rather than by argument. The prior evidence is against it: Tianlu Wang et al. showed that models trained on datasets balanced to remove attribute-label correlation still amplified the attribute through unannotated background features, and Lu Zhang et al. found that data-only de-biasing left final discrimination at 0.09–0.10 versus 0.013 with model-side intervention. SAROTUP's own scope limits compound the problem: its propagation classifier is calibrated only for Ph.D.-7, and it has no predictors for common contaminants like biotin or protein A/G, per SAROTUP 3.1. Still, this is a real threat — if filtering happens to suffice on these particular datasets, our mitigation is unnecessary, and E2 is the experiment that would say so.
Why not just use raw enrichment ratios? Practitioners do, and it costs nothing. But the evidence in the file is that enrichment conflates everything: Bayes-PD directly compared raw selectivity against noise-aware modeling and only the latter tracked experimental $K_d$. More importantly for this proposal, enrichment ranking is the mechanism by which propagation cheaters win — a fast-replicating clone shows a large fold-change by construction — so it is the worst baseline on the PLA axis, which our audit will quantify. Its role in our study is as the naive reference point, not a rival.
Why not just use physics-based scoring (FoldX, Rosetta) to filter non-binders? Physics tools evaluate structural models of binding and are blind to the selection process entirely — they cannot see propagation bias because it lives in the viral genome and the assay, not in the binder's structure. They are also too slow for the regime in question: the throughput arguments documented across EvoAb and Light-DDG (an 89.7x inference speedup over Prompt-DDG being the illustrative figure) show why learned scorers replaced them for high-volume screening. Physics-based scores could serve as a third-party affinity signal for validation clones where structures exist, and we will use them there, but they are not a competitor for the attribution question this proposal asks.
Why not just use a generic Bayesian neural network, as Bayes-PD already does? Bayes-PD is the closest prior work and the strongest baseline, and we adopt its simulator. But its authors state the open problem themselves: the model cannot disentangle positive target-binding signals from non-target and propagation artifacts, because all round-to-round variance is lumped into one Poisson noise term. A bigger or better-tuned Bayesian network with the same likelihood does not create the missing decomposition — that requires an explicit notion of which nuisance direction is parasitic (our trait-conditioned leakage probes) and a training-time intervention against it. Our attribution experiment E4 also answers a question the generic approach cannot: whether its $-0.46$ plateau is an assay property or a model property.
Decisive Checks
The proposal lives or dies on three numbers. First, PLA on unmodified affinity models must be measurably positive (E1); if trained models are no worse than the raw data at leaking parasitic traits, the amplification story does not transfer to this domain and only the attribution part survives. Second, PLA must remain positive after SAROTUP filtering (E2); if it drops to zero, the field gets a cheaper answer and we report it. Third, the gradient-reversal variant must reduce PLA by a substantial fraction at a small accuracy cost (E3), mirroring the 53–67% reduction band from Tianlu Wang et al.; if accuracy collapses instead, the Equalizer-Paradox-style compensatory pathology of Yusuke Hirota et al. has a biophysical analog, which would itself be a finding worth publishing. Each check is falsifiable on data that already exists — the multi-round NGPD datasets with $K_d$ validation, the Biopanning Data Bank's cross-target enrichment records, and the 3.05-million-sequence naïve Ph.D.-7 library — so the entire program requires no new wet-lab experiments to reach its first conclusions.
References
- Balanced Datasets Are Not Enough: Estimating and Mitigating Gender Bias in Deep Image Representations
- Achieving Non-Discrimination in Prediction
- SAROTUP: a suite of tools for finding potential target-unrelated peptides from phage display data
- Quantifying Societal Bias Amplification in Image Captioning
- Bayes-PD: Exploring a Sequence to Binding Bayesian Neural Network model trained on Phage Display data
- Aleatoric and Epistemic Discrimination: Fundamental Limits of Fairness Interventions
- Fairness Evaluation with Item Response Theory
- Biopanning data bank 2018: hugging next generation phage display
- Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions
- Structure-Guided Reinforcement Learning for High-Affinity Antibody Design
- A Simple yet Effective $\Delta\Delta G$ Predictor is An Unsupervised Antibody Optimizer and Explainer