Below is our recent interview with Dr. Ying Zhou, Product Director at CUSABIO:
Q: How is AI-driven antibody discovery changing the traditional antibody discovery workflow?
A: Traditional antibody discovery has always been a long, linear, immunization-dependent process. A researcher identifies a target, produces antigen, immunizes an animal, and then screens hybridomas or phage libraries for binders. The whole campaign can take six to eighteen months, and the quality of what comes out is constrained by whatever the animal’s immune repertoire happens to favor — which is often not a therapeutically relevant epitope.
AI changes this in two fundamental ways. First, it can generate antibody sequences in silico against a defined target structure, skipping immunization entirely. A well-trained model can output thousands of candidate CDR sequences in hours, each designed to engage a specific epitope. Second, it shifts the discovery logic from screening to engineering. Instead of panning a massive library and accepting whatever binds, one designs what should bind and then tests it. This also means developability filters — aggregation risk, immunogenicity, predicted expression yield — can be applied before any wet-lab work begins, which dramatically reduces downstream attrition.
What emerges is a parallel, iterative design-make-test-analyze loop rather than a single sequential campaign. CUSABIO’s move to multiple parallel antibody discovery platforms — hybridoma, single B cell, and phage display — reflects this shift, because no single wet-lab approach can absorb the burst of candidates that computational tools now produce.
Q: As AI tools generate increasing numbers of antibody candidates, where do you see the main wet-lab bottlenecks emerging?
A: The bottleneck has shifted from finding candidates to processing them. When a model generates a thousand sequences overnight, the limiting factor becomes how fast teams can express, purify, and characterize each one. Expression is the first rate-limiting step — every sequence needs cloning, expression, and purification, and even with high-yield HEK293 systems delivering 100 mg/L, doing this in parallel for hundreds of candidates demands substantial infrastructure. The problem gets considerably worse for membrane protein targets, which are the focus of a large fraction of therapeutic antibody programs. Multi-pass transmembrane proteins are notoriously difficult to express as soluble, correctly folded antigens, and CUSABIO’s investment in three separate platforms — VLPs, detergent solubilization, and Nanodiscs — reflects the reality that there is no universal solution.
Beyond expression, the validation matrix is where things truly slow down. Each candidate needs specificity checks, affinity measurement, cell-surface binding by FACS, and ideally some functional readout. Running that across hundreds of antibodies, each requiring different cell lines, assay formats, and timelines, becomes the real bottleneck.
Q: Why do candidate number, expression scale and validation requirements often change from one discovery round to another?
A: This comes down to the stage-gate logic of drug discovery. In early rounds, the goal is exploratory breadth: hundreds of AI-designed sequences may enter small-scale expression, with high-throughput, low-cost assays like ELISA used merely to flag any binding. Here, the candidate pool is large, but characterization per candidate is deliberately shallow.
As data accumulate, the funnel narrows and the focus shifts to specificity. Candidate numbers drop sharply — from perhaps five hundred to fifty — while expression scale increases to milligram quantities for purified material. Validation deepens to include kinetic profiling (SPR or BLI), cellular recognition (FACS), and cross-reactivity checks. By the lead stage, only a handful of candidates remain, demanding resource-intensive functional assays such as inhibition studies, ADCC, and developability panels.
Candidate number and validation depth move in opposite directions, which is why flexible expression capacity matters more than raw throughput — and why CUSABIO has built its service model around it.
Q: CUSABIO emphasizes “flexible” antibody expression rather than simply “high-throughput” expression. What is the reasoning behind this approach?
A: High-throughput pipelines imply a one-size-fits-all approach optimized for volume, yet antibody discovery is inherently non-uniform. One round may yield five candidates needing careful optimization, while the next delivers three hundred sequences requiring rapid triage. CUSABIO prioritizes flexible expression to accommodate this variability without forcing every project into the same mold.
Our infrastructure supports scales from 1 mg to 1 g with turnaround as fast as two weeks, enabling researchers to align expression strategy with project stage. Early screening demands microgram quantities in 96-well formats, whereas lead optimization requires gram-scale material for animal studies. Flexibility also extends beyond scale to format and expression host: a standard IgG expresses efficiently in HEK293, a VHH nanobody is far more economical in E. coli, and a glycoengineered Fc requires the complete post-translational modification capabilities of a mammalian host. CUSABIO has the right platform for each case.
The deeper rationale is that modern discovery is iterative and non-linear: a computational campaign might generate a handful of sequences in one iteration and hundreds in the next. Expression infrastructure must adapt to the project, not the other way around.
Q: How can researchers decide when they only need antibody expression and when additional validation such as ELISA, SPR/LSPR or BLI becomes useful?
A: The decision comes down to what is at stake when skipping validation. If a sequence already has strong computational support and only purified protein is needed for downstream assays, expression alone suffices. In this case, basic QC such as SDS-PAGE is sufficient to confirm identity and purity.
ELISA is worth adding when antigen recognition needs independent confirmation or cross-species reactivity needs screening. A cross-species ELISA panel can reveal dramatic affinity differences — between candidates, and across species homologs — that would never be apparent from sequence alone. But ELISA only provides an endpoint signal; it cannot reveal how an antibody binds or how fast it comes off.
That is where SPR, LSPR, or BLI come in. These biophysical methods provide kinetic parameters — on-rates and off-rates — that matter a great deal for therapeutic development, because a slow off-rate often predicts better in vivo efficacy than equilibrium affinity alone. For ranking leads, that resolution is essential. FACS on stable cell lines becomes critical for membrane protein targets where native-conformation recognition is essential.
Q: What do you expect the interaction between computational antibody design and wet-lab validation to look like over the next few years?
A: We see this integration accelerating rapidly. With our own proprietary model now in place, the boundary between computational antibody design and wet-lab validation is already dissolving. Rather than sequential handoffs, our model actively requests specific validation data to refine its predictions, and our wet-lab is structured to deliver those assays on demand.
We anticipate design tools prioritizing testability — selecting sequences that express efficiently, avoid aggregation confounding SPR or BLI, and target accessible epitopes. For difficult membrane proteins, AI will predict which extracellular domains or loops enable binding, reducing reliance on full-length GPCR production.
Most importantly, we foresee integrated platforms where researchers submit a target and receive designed, expressed, and validated candidates through a single workflow. This demands flexible expression and risk-adaptive validation pipelines. At CUSABIO, our end-to-end infrastructure — spanning antigen preparation to lead discovery and customized assays — is already evolving to make this AI–wet-lab partnership seamless.
