Narrowing Antibody-Drug Conjugate Design Choices With Machine Learning

Antibody-drug conjugates combine a monoclonal antibody, a linker, and a cytotoxic payload into one therapeutic molecule. The design space is enormous: the choice of antibody, linker chemistry, conjugation site, and drug-to-antibody ratio (DAR) all influence stability, safety, and efficacy. This write-up outlines a generalized pattern synthesized from publicly discussed research and development approaches rather than a single verified programme. Individual company pipelines, data quality, and outcomes vary significantly.

The problem

Traditional ADC development proceeds through iterative wet-lab screening. Teams synthesize and test many antibody-linker-payload combinations, measuring properties such as aggregation, plasma stability, payload release, and off-target binding. Each cycle is expensive and time-consuming, and the combinatorial space means that many plausible designs are never explored [1].

Computational tools have long supported parts of this process, for example through structure-based antibody modelling. More recently, machine learning is being applied to predict developability problems and prioritize designs before they are sent to the bench.

How teams approach it

To narrow the design space, discovery teams use models across four areas: antibody developability, conjugation-site prediction, linker and payload selection, and DAR/stability trade-offs.

Developability models predict biophysical liabilities such as aggregation, viscosity, and self-interaction from sequence and structure features [2]. Conjugation-site models suggest positions on the antibody where a linker is less likely to disrupt binding or stability [3]. Linker and payload models draw on historical data to prioritize chemistries with favourable release kinetics and bystander activity profiles. DAR models estimate how payload load affects pharmacokinetics and toxicity, helping teams avoid either under-loading or over-loading the conjugate [4].

What the workflow looks like

A typical computational ADC design workflow runs in parallel with experimental validation:

• Candidate antibodies are scored for developability and target binding using sequence- and structure-based models.
• Promising candidates are passed to conjugation-site prediction to identify positions that preserve stability.
• Linker-payload combinations are ranked by predicted stability, release profile, and manufacturability.
• A shortlist is advanced to in vitro characterization; results feed back into the model training set.

The model functions strictly as a prioritization aid. Wet-lab data, not model scores, decide which candidates progress [5].

How it works in practice

• What data goes in. Antibody sequence and structure-derived features (length, surface patches, charge distribution, hydrophobicity), linker chemistry descriptors, payload physicochemical properties, and historical measurements of aggregation, viscosity, plasma stability, payload release, and off-target binding [2][4].
• What the model predicts. Developability liabilities (aggregation propensity, viscosity, self-interaction), likely stable conjugation sites, linker-payload rankings by release profile, and the effect of DAR on stability and pharmacokinetics.
• Technique or model family. Sequence- and structure-based predictors, gradient-boosted trees or neural networks on tabular molecular descriptors, and physics-based or statistical models for conjugation-site accessibility.
• How it is validated. Models are validated against held-out historical ADC data and, crucially, against new in vitro measurements on candidate designs. A high developability score is treated as a hypothesis until confirmed by aggregation and stability assays.
• What the team does with the output. Models prune the obvious poor candidates and rank the remainder so medicinal chemists and protein engineers spend wet-lab capacity on the most informative designs. Any candidate that advances is reviewed by toxicologists and process scientists before synthesis.
• A concrete example. Input: antibody sequence with a predicted hydrophobic patch near CDR3, linker with a cleavable valine-citrulline chemistry, payload with moderate hydrophobicity, proposed DAR 4. Developability model output: aggregation propensity score 0.23 (low). Conjugation-site model output: site HC_A114 is predicted to preserve antigen binding and stability. The team advances this design to in vitro expression and conjugation; aggregation and plasma-stability assays either confirm the prediction or send the design back for revision.

Where it breaks

Deployments frequently stumble when teams optimize model scores that do not correlate with the real therapeutic objective. A conjugate with a high developability score may still fail in vivo because of immunogenicity or unexpected payload metabolism. Another common failure point is training-data bias: if historical datasets over-represent certain antibody formats or linker chemistries, the model will under-predict risks for novel designs [6].

Stability predictions are also context-dependent. A conjugation site that looks safe in one antibody scaffold may be destabilizing in another, and small changes in linker length or chemistry can shift aggregation behaviour in ways that are hard to capture without experimental data.

What good looks like

High-performing teams use models to prune the obvious poor candidates and focus wet-lab capacity on the most informative experiments. They maintain tight feedback loops: every in vitro and in vivo result is captured, and models are retrained or recalibrated as new data arrive. They also keep human medicinal chemists, protein engineers, and toxicologists in the decision loop for any candidate that is advanced [7].

What to check before you start

• Ensure the training data cover the antibody formats, linkers, and payloads you intend to explore.
• Define which properties the model is allowed to prioritize and which require direct experimental measurement.
• Plan for iterative validation: model predictions should be treated as hypotheses until confirmed by data.
• Document model versions and performance metrics so that improvements (or regressions) can be traced over time.

Sources

1. Fu et al., Antibody drug conjugate: the "biological missile" for targeted cancer therapy, Signal Transduction and Targeted Therapy, 2022.
2. Jain et al., Biophysical properties of the clinical-stage antibody landscape, PNAS, 2017.
3. Behrens & Liu, Methods for site-specific drug conjugation to antibodies, mAbs, 2014.
4. Chari, Miller & Widdison, Antibody-drug conjugates: an emerging concept in cancer therapy, Angewandte Chemie International Edition, 2014.
5. US FDA, Clinical Pharmacology Considerations for Antibody-Drug Conjugates, guidance for industry.
6. Wang et al., Artificial intelligence in antibody-drug conjugate development, Trends in Pharmacological Sciences, 2025.
7. ICH, S6(R1) Preclinical safety evaluation of biotechnology-derived pharmaceuticals.