Preference-based Antibody Expression Ranking

The core idea

Turn heterogeneous evidence into one preference signal.

Quantitative yields provide strong pairwise supervision; immunization-derived antibodies provide weak positive supervision at industrial scale. Both become a unified preference dataset.

Construction of strong and weak antibody preference pairs and their union into a single preference dataset — Preference-pair construction. Strong supervision uses measured yield differences; weak supervision pairs immunization-derived sequences with zero-yield antibodies.

Overview

Ranking antibodies when labels are scarce.

Antibody expression ranking is critical to antibody design, but quantitative experimental labels are scarce. This work introduces a unified preference-learning framework that combines a diverse internal dataset of 1,254 labeled sequences with 4.2 million unlabeled, immunization-derived camelid antibodies. A union-masked likelihood approximation and IMGT alignment make Direct Preference Optimization efficient for variable-length protein sequences.

01 / UNIFY

One objective, two signals

Quantitative yield differences and weak biological priors are expressed as relative preferences.

02 / ADAPT

DPO for masked pLMs

Union masking approximates sequence likelihood differences with a single forward pass per pair.

03 / ALIGN

Variable-length sequences

IMGT numbering provides a consistent coordinate system for comparing diverse VHH sequences.

04 / SCALE

More weak data helps

The two-stage pipeline continues to benefit as weak supervision grows from 1K to 4.2M samples.

Method

Domain alignment first. Preference optimization second.

The pipeline adapts the backbone to antibody sequence space before aligning sequence likelihoods with expression preferences.

Continual pretraining

Masked language modelling on Camel-4M adapts each backbone to the antibody-specific sequence distribution.

Preference fine-tuning

DPO combines strong yield-derived pairs and weak immunization-derived pairs in a single objective.

Union-masked scoring

Only positions that differ across an IMGT-aligned pair are jointly masked and evaluated.

Why it matters: exact pseudo-log-likelihood needs one forward pass per position. The union-mask approximation targets the relative score needed by DPO without evaluating unchanged positions.

Results

Consistent gains across three backbones.

The full Stage I + II pipeline achieves the strongest overall performance across AntiBERTa2, IgBERT, and ProtBERT on the Exp-1K test set.

Antibody LM · Stage I + II

AntiBERTa2

82.4AUC

43.7MCC

46.3Kendall’s Tau

50.8SCC

Composite Avg: 45.0 · Recall: 46.2

Antibody LM · Stage I + II

IgBERT

74.8AUC

37.5MCC

42.9Kendall’s Tau

54.7SCC

Composite Avg: 40.1 · Recall: 61.5

Protein LM · Stage I + II

ProtBERT

78.0AUC

40.1MCC

40.5Kendall’s Tau

44.5SCC

Composite Avg: 40.3 · Recall: 53.8

Average performance as the number of Camel-4M samples increases, with and without continual pretraining — Scaling with weak supervision. Continual pretraining enables more stable and sustained gains as data scale increases.

Validation performance trajectories comparing full-data DPO, 1K-data DPO, and supervised training — Optimization stability. Full-data DPO converges more smoothly than low-data DPO and supervised regression.

Data & analysis

Scarce measurements, abundant biological priors.

Exp-1K provides quantitative expression labels and realistic failure cases. Camel-4M contributes a broad weakly positive signal from immunization-derived VHH sequences.

Exp-1K

Labeled expression benchmark

1,254 unique VHH sequences, including 335 non-expressible candidates and substantial length variability.

Camel-4M

Weak supervision corpus

4.2 million non-redundant, immunization-derived VHH sequences used for domain alignment and preference construction.

Histogram of expression yields in the Exp-1K dataset — Exp-1K yield distribution, including a distinct zero-yield peak.

Sequence length distributions for Exp-1K, Camel-4M, and public antibody expression datasets — Sequence-length diversity across internal and public datasets.

Pairwise IMGT-aligned Hamming similarity distribution for Exp-1K with thresholds at 0.31 and 0.59 — Exp-1K pairwise similarity is tri-modal. Using the full distribution outperforms similarity-threshold filtering.

Citation

Cite this work.

Accepted at the 43rd International Conference on Machine Learning (ICML 2026).

@inproceedings{sun2026preference,
  title     = {Preference-based Antibody Expression Ranking:
               Scaling with Large-scale Weak Supervision},
  author    = {Sun, Josh Qixuan and Babaie, Morteza and Hou, Wenyang
               and Crowley, Mark and Young, David},
  booktitle = {Proceedings of the 43rd International Conference
               on Machine Learning},
  year      = {2026}
}