Artificial intelligence is transforming antibody discovery. Machine learning models can predict novel binders, optimize sequences for improved properties, and navigate sequence spaces impossible to explore experimentally. Yet even the most sophisticated AI cannot fully capture the complexity of protein biology.
Physical validation remains essential, and High-Throughput (HTP) Expression Platforms serve as the critical bridge between computational design and experimental reality—enabling rapid iteration while ensuring candidates are evaluated in a manufacturing-relevant context.
AI in Antibody Discovery
Modern AI tools address two fundamental challenges:
1. De novo binder prediction
Machine learning models generate novel antibody sequences predicted to bind specified targets, leveraging large datasets to learn relationships between sequence, structure, and binding.1, 2
2. Sequence optimization
AI can optimize existing antibodies for improved affinity, stability, reduced aggregation, or higher expression yield3. Multi-objective approaches simultaneously balance multiple parameters, proposing variants representing optimal trade-offs.
Rather than sampling sequence space randomly as in phage display, AI-guided design focuses experimental effort on high-confidence sequences—down-selecting from millions of possibilities to a manageable set of 100–200 candidates, dramatically reducing experimental burden while increasing the probability of success.
The Necessity of Physical Validation
AI models remain imperfect because training data may not fully represent the vast diversity of antibody-antigen interactions. Subtle biological effects can influence binding and stability in ways that current models may not accurately capture. Properties such as formulation stability and high-concentration viscosity remain difficult to predict from sequence alone.
Physical validation is therefore a strategic requirement rather than an optional step. The primary challenge for Biotech companies is executing this validation efficiently to maintain discovery momentum. Standardized High-Throughput Antibody Production ensures that these physical reality checks keep pace with digital design cycles.
Lab-in-the-Loop: Iterative Model Refinement
The most powerful paradigm for AI-driven discovery is the “lab-in-the-loop” approach, where prediction and validation operate in tight iterative cycles:
- The AI model generates candidate sequences
- Candidates are expressed, purified, and characterized
- Experimental results finetune the model
- The refined model generates improved candidates
- The cycle repeats until requirements are met
Each round of experimental data improves model accuracy for the specific target and properties of interest. Over successive iterations, predictions become increasingly reliable, and final candidates emerge from computational-experimental co-optimization.
The Role of High-Throughput Production
The lab-in-the-loop paradigm places specific demands on experimental capabilities:
- Speed: Turnaround time matters more than production scale. Faster results accelerate iteration cycles and maintain discovery momentum
- Scalability: Parallel processing of 100–200 variants per round requires platforms capable of simultaneous small-scale expression of large variant sets
- Minimal material: Intermediate validation rounds—binding confirmation, basic stability—require only microgram quantities. evitria’s HTP platform delivers sufficient material with rapid turnaround
- Data consistency: Standardized protocols and robotic handling minimize technical variability to prevent “Garbage In, Garbage Out” scenarios. This ensures that experimental noise does not corrupt model refinement by providing the clean, comparable data necessary for high-fidelity predictive modeling.
As iterations converge on a small set of leads—typically fewer than ten—comprehensive characterization becomes necessary. evitria can directly upscale using the same DNA plasmids and CHO expression platform, ensuring full continuity throughout the workflow.
Why CHO Matters in AI-Driven Workflows
Expression system choice significantly impacts outcomes even in AI-accelerated discovery. If models are trained on cell-free or HEK293-produced material, predictions will be optimized for that context and candidates that perform well in HEK293 may not translate to CHO. Starting physical validation in CHO from the earliest iterations offers clear advantages:
- Model refinement based on manufacturing-relevant data improves predictive accuracy for properties that matter in development
- Glycosylation, charge variants, and aggregation behavior measured in CHO reflect later development stages
- Candidates optimized through CHO-based iterations require no re-evaluation after a host switch
- Final candidates advance directly toward cell line development without host-dependent uncertainty
evitria’s CHO platform supports the speed and throughput requirements of AI-driven discovery, with transient expression timelines compatible with iterative workflows and miniaturized formats enabling parallel production of large variant sets.
Conclusion
AI-driven antibody discovery navigates sequence space with unprecedented efficiency, proposing candidates that traditional screening would never find. Yet physical validation remains essential to confirm predictions and ensure candidates possess the properties required for development.
HTP expression in CHO, from the earliest iterations, ensures the entire process operates in a manufacturing-relevant context, informs model training with meaningful data, and allows selected candidates to progress without host-transition risk. The future of antibody discovery lies in the synergistic integration of AI prediction and experimental validation, supported by HTP CHO expression capabilities.
Frequently Asked Questions
It is the disconnect between the thousands of sequences designed in-silico and the limited capacity to validate their actual biological behavior in a wet-lab environment.
By providing physical ground-truth data the model can be tuned to avoid future design errors when a prediction is compared against actual expression titers or binding kinetics.
If an AI is trained on data from non-standard hosts like HEK-293 cells it may predict designs that are only stable in those surrogate cells. Using CHO ensures the AI learns manufacturing-relevant patterns and avoids glycosylation issues.
High-quality data has a higher signal-to-noise ratio which is vastly more predictive for AI models than massive datasets clouded by process-noise and inconsistent laboratory protocols.
Projects are typically initiated within 24 hours of receipt in Zurich and deliver assay-ready material within a standard four-week window to support rapid iterative cycles.
References
- Shanehsazzadeh, A., McPartlon, M., Kasun, G., Steiger, A. K., Sutton, J. M., Yassine, E., McCloskey, C., Haile, R., Shuai, R., Alverio, J., Rakocevic, G., Levine, S., Cejovic, J., Gutierrez, J. M., Morehead, A., Dubrovskyi, O., Chung, C., Luton, B. K., Diaz, N., . . . Bachas, S. (2023). Unlocking de novo antibody design with generative artificial intelligence. bioRxiv (Cold Spring Harbor Laboratory). https://doi.org/10.1101/2023.01.08.523187 ↩︎
- Bennett, N.R., Watson, J.L., Ragotte, R.J. et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature 649, 183–193 (2026). https://doi.org/10.1038/s41586-025-09721-5 ↩︎
- Hie, B. L., Shanker, V. R., Xu, D., Bruun, T. U. J., Weidenbacher, P. A., Tang, S., Wu, W., Pak, J. E., & Kim, P. S. (2024). Efficient evolution of human antibodies from general protein language models. Nature biotechnology, 42(2), 275–283. https://doi.org/10.1038/s41587-023-01763-2 ↩︎
