Computational Prediction of Antibody Developability

24 concepts18 papers1 proposal

Imagine designing a key that perfectly fits a lock in a digital simulation, only to find that when you manufacture it out of real metal, it is too sticky to turn, melts at room temperature, or is too brittle to use. This scenario mirrors the core challenge of modern therapeutic antibody design.

For decades, computational and experimental pipelines prioritized binding affinity—how tightly an antibody binds to its target antigen. However, as highlighted by Khetan et al., many promising candidates fail late in clinical development because they lack essential biophysical properties required for manufacturing, formulation, and delivery. This challenge is known as the developability bottleneck (Golinski et al.). To solve this, researchers are shifting developability assessments from the end of the drug discovery pipeline to the very beginning using machine learning.

The Biophysical Mechanics of Developability

Why do some high-affinity antibodies aggregate or become highly viscous while others remain stable? The answer lies on the molecule's three-dimensional surface.

Primary sequence data alone often obscures these liabilities. As Kovaltsuk et al. argue, sequence-only analysis leaves researchers "functionally blind" to spatial arrangements. Critical developability liabilities—such as surface-exposed hydrophobic patches or unfavorable charge distributions—only manifest when considering the molecule's 3D surface properties. When antibodies are concentrated for clinical administration, these surface patches interact non-specifically with each other, leading to high viscosity or irreversible aggregation.

Furthermore, these properties are not determined by a single protein chain in isolation. In a therapeutic antibody, the physical behavior is heavily dictated by how the heavy and light chains interact (Turnbull et al.). Models that fail to account for paired-chain structural dynamics struggle to accurately predict real-world biophysical performance.

Traditionally, properties like solubility, viscosity, thermal stability, and low aggregation were evaluated only at the final stages of discovery due to the high cost and material requirements of wet-lab assays (Sormanni et al.). If a candidate failed these tests, the substantial time and financial investments were wasted, or researchers had to engage in extensive post-discovery optimization that risked compromising the antibody's primary binding function (Latent Labs Team et al.).

Shifting Assessment to the Early Pipeline

To circumvent late-stage failures, recent research focuses on moving developability assessments to the very beginning of the design pipeline using computational models. Several strategies have emerged to predict these complex physical traits from sequence or structural representations:

  • Surrogate and Transfer Learning: High-throughput proxy assays can generate large, albeit noisy, datasets. Golinski et al. demonstrated that transfer learning can bridge this gap; by pre-training a model (DevRep) on proxy data and fine-tuning it on actual recombinant yields, they significantly improved the prediction of manufacturing viability.
  • Fast Sequence-Based Approximations: While physics-based simulations are computationally expensive, frameworks like TherAbDesign bypass them entirely, predicting downstream traits like viscosity directly from sequence features to enable rapid, early-stage screening (Wang et al.).
  • Surface-Level Modeling: To capture the true physical drivers of aggregation, tools like SurfProp map the antibody's 3D structure to its surface properties, accelerating the calculation of electrostatics and hydrophobicity to predict experimental developability metrics (Rao et al.).
  • Integrative Frameworks: Open-source tools like PROPERMAB combine sequence data with predicted 3D structural features to screen thousands of candidates for viscosity and hydrophobicity before they ever enter a wet lab (Bian et al.).

Generative Co-Design and Active Filtering

Filtering out bad candidates after generation is useful, but a more powerful approach is to bake developability constraints directly into the generative process.

Instead of treating developability as a post-hoc filter, modern architectures treat antibody design as a multi-objective optimization problem. For instance, Villegas-Morcillo et al. utilized gradient-free guidance techniques to steer diffusion models, co-designing antibody sequences and structures that simultaneously optimize for antigen binding, solubility, and thermal stability without requiring model retraining. Similarly, language models can be fine-tuned or paired with high-precision screening models to generate sequences that inherently minimize polyreactivity and aggregation (He et al.; Turnbull et al.).

[Generative Model] ───> [Predicted 3D Surface] ───> [Guidance / Gradients]
       ▲                                                    │
       └─────────────────── Optimize Both ──────────────────┘
                    (Affinity + Developability)

This holistic approach changes how candidates are selected for experimental validation. As Sayeed et al. argue, experimental success is often limited more by laboratory assay compatibility and expression yield than by predicted binding energy. By prioritizing developability metrics over pure structural or affinity scores within multi-agent filtering frameworks, researchers can drastically reduce wet-lab false-positive rates.

Ultimately, integrating these biophysical assessments early into discovery engines ensures that highly active therapeutic candidates are fundamentally compatible with large-scale manufacturing and clinical use (DeVorkin et al.). While a gap still exists between digital fold stability and physical marketability (Ferruz et al.), the transition toward developability-aware computational design represents a paradigm shift in biotherapeutic engineering.

Key Foundational Work

Five computational developability guidelines for therapeutic antibody profiling (2019) — contributed 1 concept elements to this direction.
How B-Cell Receptor Repertoire Sequencing Can Be Enriched with Structural Antibody Data (2017) — contributed 1 concept elements to this direction.
Rapid and accurate in silico solubility screening of a monoclonal antibody library (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_biophysical_manufacturability_f1f25826,
  title = {biophysical manufacturability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/biophysical-manufacturability/rdr_f1f2582692224f869c4769c69b02680c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_biophysical_manufacturability_f1f25826_v1,
  title = {biophysical manufacturability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/biophysical-manufacturability/rdr_f1f2582692224f869c4769c69b02680c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake