Antibody Discovery

The pipeline

Antibody discovery starts with a large pool of candidates and progressively narrows it. Initial candidates come from hybridoma campaigns, phage or yeast display, single B-cell workflows, immune repertoire mining, or computational generation. From there, the pipeline moves through binding assays, functional screens, recombinant expression, specificity testing, stability profiling, developability assessment, and eventually manufacturability review.

At each stage, the pool shrinks. A display campaign may yield hundreds of hits. Only a fraction get expressed recombinantly. Fewer still make it through specificity and stability panels. By the time a candidate reaches developability assessment, it has consumed significant resources across multiple assay types, expression systems, and characterization rounds.

The decisions about which candidates advance, which assays to run next, and which tradeoffs to accept are made manually at each gate. They draw on program experience, threshold rules, and team judgment. The data from earlier stages is available but rarely integrated into a model that systematically informs the next decision.

ProviGen connects all of this into a decision loop. The platform ingests sequence features, binding data, expression levels, stability measurements, specificity profiles, and execution metadata. After each characterization round, it updates its model and recommends which candidates to advance, which to re-engineer, and which assays are most informative to run next.

Generate and predict

Modern antibody campaigns increasingly start with computational candidate generation. Generative models can propose variants of a lead sequence by exploring mutations in complementarity-determining regions (CDRs) while respecting structural constraints. Multi-task property predictors then score each variant across binding, expression likelihood, stability, and developability before anything is synthesized.

This filtering step matters because synthesis and screening capacity is limited. A generative model may propose thousands of plausible variants. Property predictors narrow the field to those worth testing. Recent lab-in-the-loop work^[1] has shown that this approach can produce variants where 97 to 100% express and up to 70% show functional binding, dramatically reducing wasted synthesis and screening effort.

ProviGen supports this workflow by connecting generative proposals with experimental results. As real assay data comes back, the property predictors improve. Unexpected failures and surprising results on stability or specificity all refine the scoring. Each round of synthesis and testing makes the next round of generation more targeted.

Select and test

Given a scored pool of candidates, the question is which ones to test next. Active learning formalizes this: instead of testing the top-ranked candidates by a single metric, the system selects variants that are most informative given what the model already knows. Some are selected because they are predicted to perform well. Others because they sit in regions of the design space where the model is uncertain and testing them would improve future predictions.

In practice, this means each round of synthesis and screening produces data that the model uses to update its predictions and propose the next set of candidates. A typical strategy enforces sequence constraints that relax across rounds: early rounds stay close to the lead (within a few mutations), later rounds explore further. This graduated exploration balances confidence in the starting point against the need to discover better regions of sequence space.

ProviGen manages this iterative selection across the full characterization pipeline. Binding data from an early round informs which candidates are worth the more expensive stability profiling. Expression failures from one batch inform sequence design choices in the next. Specificity patterns learned from one target can transfer to related campaigns.

Multi-property optimization

Therapeutic antibody candidates must perform across multiple properties simultaneously: binding affinity and kinetics, expression yield, specificity (low cross-reactivity, low polyreactivity), thermal and colloidal stability, functional potency, and developability (viscosity, self-interaction, charge variants). Tighter binding often comes with reduced stability or increased polyreactivity. High expression variants may have developability liabilities.

The model maps the Pareto frontier (the set of candidates where no single property can be improved without compromising another). Teams define which properties matter most for their specific program, and the model proposes candidates and experiments that move toward that frontier. This replaces the typical approach of optimizing affinity first and hoping the other properties follow.

Lab-in-the-loop

The strongest public demonstration of this approach comes from Genentech, Roche, and Prescient Design.^[1] Their lab-in-the-loop system combined generative models, multi-task property predictors, active-learning selection, synthesis, and in vitro testing into a semi-autonomous optimization loop. Applied to four clinically relevant targets (EGFR, IL-6, HER2, OSM), the system designed and tested over 1,800 unique antibody variants across iterative rounds, achieving 3 to 100x binding improvements with best binders in the therapeutically relevant sub-100 pM range.

Arc Institute's MULTI-evolve framework^[2] points in the same direction: protein language models guiding rapid directed evolution of complex multi-mutant proteins, with computational prediction and experimental design coupled from the start. The Coefficient Bio team that built much of the Genentech system was acquired by Anthropic for $400M in April 2026,^[3] signaling that foundational AI companies see lab-in-the-loop biological design as a core application.

---

ProviGen brings this decision loop to antibody discovery teams working with their own targets, assays, and automation. Every candidate tested produces a result, updates the model, and improves the next selection. Learnings compound across rounds and transfer across programs. Failed candidates are informative. Partial characterization data is usable immediately, without waiting for full panels to complete.

If you are interested in exploring what a lab-in-the-loop campaign could look like for your antibody program, reach out to contact@provigen.ai.