💻 bioinformatics

An interpretable open platform for sequence-based antibody developability prediction

The paper introduces DELPHI, an open-source platform that enables laboratories to train, evaluate, and interpret sequence-based antibody developability predictors using rigorous cross-validation, achieving high performance on both proprietary and public datasets without requiring access to proprietary training data.

Original authors: Nguyen, H. N., Kothiwal, D., Su, Y., Kieu, M. A., Cao, R., Zhu, H., Teixeira, A. A. R.

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nguyen, H. N., Kothiwal, D., Su, Y., Kieu, M. A., Cao, R., Zhu, H., Teixeira, A. A. R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Antibodies are the immune system's specialized soldiers, Y-shaped proteins designed to recognize and neutralize specific invaders like viruses or bacteria. In modern medicine, scientists engineer these molecules to fight diseases ranging from cancer to autoimmune disorders. However, creating a therapeutic antibody is a high-stakes gamble. Just because a molecule can grab onto a target does not mean it will survive the journey to a patient. Many candidates fail because they are "sticky," clumping together or binding to the wrong proteins in the body, which can cause them to be cleared too quickly or trigger dangerous immune reactions. These failures, known as developability liabilities, are often discovered late in the development process, wasting years of research and millions of dollars. For decades, the only way to know if an antibody was safe and stable was to synthesize it in a lab and run physical tests, a slow and expensive bottleneck that could not keep up with the millions of potential candidates generated by modern genetic technologies.

A team of researchers at the Institute for Protein Innovation has built a new tool called DELPHI to solve this bottleneck. Think of it as a highly trained digital scout that can look at the genetic blueprint of an antibody and predict whether it will be stable and safe, long before a single molecule is ever made in a lab. The researchers did not just create a single prediction model; they built an open platform that allows any laboratory to train its own custom predictors using their specific data. By testing twenty-five different combinations of artificial intelligence techniques against real-world experimental data, they found that the key to accurate prediction lies not in the complexity of the computer algorithm, but in how the antibody's sequence is represented to the machine. Their system successfully learned to spot the subtle chemical signatures that lead to failure, such as an imbalance of electrical charges in the antibody's binding region, and it can now screen candidates with a level of accuracy that rivals the best human experts.

The core of this work involves teaching computers to read the language of proteins. Antibodies are made of chains of amino acids, and the specific order of these building blocks determines how the molecule behaves. The researchers gathered a massive collection of antibody sequences that had already been tested in the lab, labeling them as either "pass" (stable and non-sticky) or "fail" (prone to clumping or sticking to the wrong things). They fed this data into DELPHI, which tried different ways of translating these amino acid chains into a format a computer could understand. Some methods treated the sequence like a simple list of parts, while others used advanced "language models" that understand the context and relationships between different parts of the protein, much like how a human understands that a word's meaning changes depending on the sentence it is in.

The results were striking. The system learned that the choice of how to represent the antibody sequence was far more important than the type of computer model used to make the prediction. Specifically, models that looked at the heavy and light chains of the antibody together, rather than in isolation, were much better at spotting the subtle patterns that lead to failure. When tested on a library of over 246,000 antibodies, the best version of DELPHI achieved an AUC of 0.95 in distinguishing stable antibodies from unstable ones. This performance held true even when the system was asked to predict the behavior of antibodies it had never seen before, including those from clinical trials that were not part of the training data. In a blind test against a major industry competition, the tool performed as well as the top human-submitted entries, despite having no access to the competition's specific data.

Beyond simply predicting pass or fail, DELPHI offers a level of detail that was previously unavailable to most researchers. It does not just give a score; it explains why it gave that score by pointing to the specific amino acids responsible for the risk. The analysis revealed a consistent pattern: antibodies that failed tended to have an excess of positively charged building blocks, particularly arginine and lysine, in their binding loops, while lacking negatively charged ones like aspartate. This electrical imbalance makes the antibody "sticky," causing it to grab onto unrelated proteins or clump together. The tool identified this same dangerous signature in two different types of failure: antibodies that stuck to the wrong things and those that failed to remain as single, stable units. This suggests that fixing the electrical charge in these specific regions could prevent multiple types of failure at once.

The researchers also mapped out how much data is needed to make these predictions reliable. They found that the system's accuracy improved rapidly as more data was added, but it began to level off after about 5,000 diverse antibody sequences. This provides a clear guide for laboratories: they do not need millions of samples to build a useful predictor; a few thousand well-characterized examples are often enough to reach a high level of confidence. Furthermore, the tool is designed to be flexible. Because it is open-source software, any lab can upload their own experimental results, retrain the model on their specific type of antibody, and immediately start screening new candidates. This democratizes the ability to predict developability, moving the field away from expensive, late-stage failures and toward a future where the most promising antibodies are identified early, saving time and resources for the patients who need them.

Technical Summary: DELPHI – An Interpretable Open Platform for Sequence-Based Antibody Developability Prediction

Problem Statement

Antibody developability liabilities, such as polyreactivity (non-specific binding) and aggregation (non-monomeric failure), are major causes of candidate failure in therapeutic development, often discovered late in the pipeline at significant cost. While modern display technologies and next-generation sequencing (NGS) generate libraries of $10^6$ – $10^{10}$ diverse sequences, experimental screening cannot keep pace. Although sequence-based machine learning (ML) and protein language models (PLMs) offer a route to in silico pre-screening, most published approaches lack reusability, retrainability, and granular interpretability. Existing tools rarely provide integrated platforms that combine multi-PLM benchmarking, leakage-aware cross-validation, training-set sizing analysis, and per-residue attribution in a single, open-source framework.

Methodology

The authors present DELPHI (Deep End-to-end Learning Platform for antibody developability with High Interpretability), an open-source pipeline designed to build, benchmark, and deploy sequence-based developability classifiers.

Data Curation and Training Sets

IPI PSR Dataset: Curated from yeast display Fab libraries with diversity confined to CDR H3 across eight VH and six VL germlines. It combines 7,494 high-confidence binary-labeled antibodies from PSR-ELISA (against DNA, avidin, insulin, OVA/SMP) and 3,771 high-polyreactivity "Fail" antibodies identified via NGS-based ssDNA selection, resulting in a near-balanced set of 11,265 antibodies (53% Pass / 47% Fail).
IPI SEC Dataset: Contains 5,045 antibodies with binary annotations based on size-exclusion chromatography (SEC) profiles (~64% Pass / 36% Fail). The Pass class was downsampled using out-of-fold RF/XGBoost k-mer consensus scores to address class imbalance while retaining all Fail antibodies.
External Benchmarks: The public DS1 library (246,293 scFv antibodies) and three external clinical-stage cohorts (Jain 2017, GDPa1, GDPa3) were used for transfer and zero-shot validation.

Model Architecture and Benchmarking

DELPHI evaluates 25 combinations of classifiers and sequence representations:

Classifiers: XGBoost, Random Forest (RF), CNN, a Transformer on language-model embeddings, and a dual-branch one-hot Transformer.
Representations: Five antibody-specific PLMs (AbLang2, IgBert, AntiBERTy, AntiBERTa2, AntiBERTa2-CSSP) plus biophysical descriptors, k-mer frequencies, and one-hot encodings.
Validation Strategy: The platform employs CDR H3-cluster-stratified cross-validation (80% identity threshold) to prevent sequence-similarity leakage. It also includes automated learning-curve analysis to characterize performance scaling with dataset size.

Interpretability

The platform provides multi-resolution interpretability:

Tree-based models (RF, XGBoost): Use SHAP (SHapley Additive exPlanations) to attribute predictions to biophysical features and residue counts.
Neural Networks: The dual-branch TransformerOneHot architecture uses Integrated Gradients (IG). Unlike PLM embeddings which compress sequences into latent vectors, the one-hot representation maps inputs directly to specific amino acids at defined positions, enabling residue-level attribution across the full VH + VL + CDR H3 sequence.
In silico Mutagenesis: The platform generates CDR3 loop mutagenesis heatmaps to test engineering hypotheses by substituting residues and observing $\Delta P(\text{Pass})$ .

Key Results

Performance Benchmarks

Within-Distribution: On the IPI PSR dataset, all 20 PLM-based combinations achieved a mean AUC of 0.959 ( $\pm$ 0.005). The best combination (XGBoost + IgBert) reached 0.967. For SEC monomer purity, the mean AUC across PLM combinations was 0.933, with XGBoost + biophysical features achieving the highest AUC of 0.9596.
Cross-Library Transfer: PLM choice proved to be a stronger determinant of performance than classifier architecture. Framework-centric paired-antibody PLMs (IgBert, AbLang2) demonstrated superior generalization.
- Models trained on the smaller, highly diverse IPI library transferred to the massive DS1 public library with an AUC of 0.950 (AbLang2).
- Conversely, models trained on DS1 transferred poorly to IPI (AUC 0.741–0.883), highlighting the importance of training-set diversity over sheer volume.
Zero-Shot Clinical Transfer: IPI-trained models achieved zero-shot transfer to external clinical cohorts:
- Jain 2017: AUC of 0.73 on a binarized polyreactivity panel.
- Ginkgo PR-CHO: Spearman $|\rho| = 0.35$ , comparable to the best reported point estimate (0.337) from the 2025 Ginkgo Developability Competition, without training on the competition data.

Learning Curves

Analysis revealed that performance gains plateaued at approximately 5,000 CDR H3-diverse antibodies for the Transformer + AbLang2 model. Below this threshold, performance was volatile; beyond it, AUC stabilized near 0.95–0.96 for IPI and ~0.98 for DS1.

Biological Insights and Interpretability

Conserved CDR H3 Signature: Both polyreactivity and SEC monomer purity failure were associated with a specific electrostatic signature: enrichment of arginine (R) and lysine (K), depletion of aspartate (D), longer CDR H3 loops, and a net positive CDR H3 charge.
Residue-Level Attribution: IG and SHAP analyses confirmed that cationic residues (R, K) and tryptophan (W) drive "Fail" predictions, while anionic residues (D) drive "Pass" predictions.
Dual-Liability Hypothesis: The convergence of these signatures across two distinct biophysical liabilities suggests that reducing net positive CDR H3 charge could simultaneously mitigate polyreactivity and aggregation risks.

Significance and Claims

The paper claims DELPHI addresses critical gaps in the field by providing:

Open, Retractable Tools: Unlike many published models, DELPHI is released as runnable, retrainable software, allowing any laboratory to build predictors from their own assay data.
Quantitative Data Guidance: Learning curves provide practical design rules, indicating that ~5,000 diverse antibodies are sufficient to reach performance plateaus for these tasks.
Representation-Aware Selection: The study demonstrates that PLM choice (specifically paired-chain models like AbLang2 and IgBert) is more critical for cross-library transfer than the downstream classifier architecture.
Residue-Level Engineering Hypotheses: By resolving predictions to individual residues, DELPHI enables the generation of testable engineering hypotheses (e.g., specific CDR3 mutations to reduce risk) rather than just binary classifications.

The authors maintain modesty regarding generalization, noting that their models are trained on libraries with restricted germline diversity (8 VH, 6 VL) and that performance drops when entire VH germlines are held out during training (AUC dropping from 0.965 to 0.903). They acknowledge that liabilities governed by assembled folds rather than loop composition may still require structural information. However, for the liabilities and libraries examined, DELPHI establishes that developability is partly predictable from sequence alone, offering a scalable, interpretable framework for early-stage antibody discovery.