When the Wet Lab Becomes the Bottleneck: Self-Driving Platforms for Protein Engineering

The lab is the bottleneck: Generative antibody models now produce sequences faster than wet labs can build, test, and characterize them.
Closed loops are the new default: Self-driving Design–Build–Test–Learn platforms compress weeks of human cycles into unattended overnight runs.
ML-ready by design: Modern automation often emits standardized numerical feature vectors that flow straight into model training pipelines.
Frictions remain: Well-to-well variability, instrument API integration, and immunogenicity assays still constrain end-to-end autonomous antibody discovery.

A high-throughput platform for biophysical antibody developability assessment to enable AI/ML model training: Arsiwala et al., mAbs, 2025
Self-driving laboratories to autonomously navigate the protein fitness landscape: Rapp et al., Nature Chemical Engineering, 2024
Engineering multispecific antibodies with complete killing selectivity through the closed-loop integration of machine learning and high-throughput experimentation: Mullin et al., mAbs, 2025
Autonomous ‘self-driving’ laboratories: a review of technology and policy implications: Tobias & Wahab, Royal Society Open Science, 2025
An Automated End-to-End Workflow for Production of Secreted Proteins in Transfected Mammalian Cells:Vasnarungruengkul et al., bioRxiv, 2025

How closed-loop automation, ML-ready datasets, and autonomous Design–Build–Test–Learn cycles are reshaping therapeutic antibody discovery

For most of the past decade, generative models have been outrunning the laboratory. Protein language models, diffusion-based structure generators, and Bayesian flow networks now produce de novo antibody sequences faster than any team can clone, express, and characterize them. The constraint has shifted: cloning vectors, transient transfections, microscale purification, biophysical panels, and developability assays now sit on the critical path. Generating the experimental data needed to train better models has become the rate-limiting step in antibody engineering, and the response has been a new class of infrastructure explicitly built to address that gap. Autonomous “self-driving” laboratories integrate robotics, scheduling software, and active-learning agents into a closed-loop Design–Build–Test–Learn (DBLT) engine that runs continuously with minimal human intervention.

The blueprint is not new, but recent demonstrations show that the DBLT principle can work. SAMPLE, described in a 2024 Nature Chemical Engineering paper, paired an intelligent agent with a remote cloud lab to engineer thermostable enzymes without human input. A 2025 Royal Society Open Science review then cataloged how these systems span chemistry, materials science, and the biological sciences, framing closed-loop experimentation as the dominant motif of the next decade of autonomous research.

“Self-driving laboratories automate and accelerate the scientific discovery process.”— Rapp, Bremer & Romero, Nature Chemical Engineering (2024)

Three 2025 reports show where antibody-specific platforms are heading. Ginkgo Bioworks introduced PROPHET-Ab, a high-throughput developability suite that profiles 246 therapeutic IgGs across ten biophysical assays and emits ML-ready datasets at the scale required to train modern predictors. LabGenius reported EVA, a closed-loop multispecific platform that traversed 44,160 T-cell engager designs and recovered sub-10 pM HER2×CD3 candidates with 14-fold enrichment over Sobol sampling. A Caltech group published a fully automated end-to-end workflow that takes plasmid DNA through to assay-ready secreted protein on a single 96-well robotic line.

The common thread is that experimental output is no longer treated as ad hoc data. Every assay is engineered to emit standardized numerical feature vectors (kinetic constants, T_m, ΔPWL, polyspecificity indices) that flow directly into an ML feature store. Bottlenecks remain in well-to-well variability, instrument API integration, and immunogenicity assays that are difficult to miniaturize. The trajectory is nonetheless clear: when the lab itself becomes a programmable component of the model, the question shifts from how to design a million antibodies at scale to which million to measure next.