Rescuing a stalled IgG campaign: 10 candidates meeting potency and developability criteria in 3 rounds

Context

Therapeutic antibody development requires balancing efficacy (potency) against multiple developability properties that determine whether a candidate can be manufactured, formulated, and administered safely. Aggregation propensity, polyreactivity, immunogenicity, and cell binding represent critical failure modes that can halt development during late-stage testing or manufacturing scale-up.

Developability optimization is particularly challenging for IgGs because many of these properties involve complex biophysical phenomena that are difficult to predict from sequence alone and often exhibit antagonistic relationships with potency.

Challenge

The commercial partner had a single functional parental sequence but this template exhibited unfavorable manufacturability characteristics, including two existing sequence liabilities that ideally should be engineered out. Potency was specified to be tuned within a narrow range of one order of magnitude—tight enough to maintain efficacy but with some flexibility for optimization.

The partner's prior screening campaign failed to obtain potent variants, despite successful binding being confirmed via ELISA, highlighting the gap between simple target recognition and functional efficacy. An in-house lead optimization effort similarly failed to obtain satisfactory variants, leaving the program at a dead end. The challenge was compounded by the need to simultaneously address six developability criteria alongside potency—a seven-property optimization problem with multiple antagonistic trade-offs.

Approach

Scientists performed three rounds of Cradle optimization with 96 candidates each, consuming data from the failed screening and optimization campaigns as training inputs. All seven properties were specified as constraints or objectives:

• Potency: target range (one order of magnitude window)

• SMAC and AC-SINS: aggregation assays (minimize)

• PAIA: polyreactivity/nonspecificity assay (minimize)

• Cell binding: off-target binding assay (minimize)

• Immunogenicity: in silico prediction and/or ex vivo assay (minimize)

• Expression: yield in standard expression system (maintain or improve)

The optimization required eliminating the two existing sequence liabilities while maintaining potency within specification and improving all six developability metrics.

Results

Cradle delivered 10 candidates that met both potency and developability criteria. These candidates were selected for advancement to the next stage of the partner's pipeline—an outcome that had eluded both the prior screening campaign and in-house optimization efforts.

All 10 advanced candidates achieved:

• Potency within the specified one-order-of-magnitude target range

• Improved aggregation propensity (SMAC and AC-SINS metrics)

• Reduced polyreactivity (PAIA assay)

• Minimal off-target cell binding

• Acceptable immunogenicity profiles

• Sufficient expression for manufacturing considerations

The successful elimination of both sequence liabilities while maintaining potency demonstrates that Cradle can enforce hard constraints (removing specific residues or motifs) alongside multi-property optimization.

Critically, where prior efforts produced candidates that either showed binding but no potency (screening campaign) or failed to meet developability criteria (in-house optimization), Cradle's multi-property approach ensured that generated variants satisfied all requirements simultaneously. This reflects the system's ability to navigate high-dimensional constraint spaces rather than optimizing properties in isolation.

What this means

This result demonstrates Cradle's applicability to the most challenging stage of antibody development: late-stage lead optimization where candidates must satisfy numerous developability criteria simultaneously. The ability to rescue a stalled program that had exhausted conventional approaches validates the system's value for high-stakes pharmaceutical development.

The 10 successful candidates represent genuine program progression—moving molecules from a dead end into downstream development stages. For the commercial partner, this avoided program termination and preserved substantial prior investment in target validation and early-stage research.

The ability to consume data from failed campaigns (screening that yielded no potent binders, optimization that yielded no developable candidates) and synthesize it into successful designs illustrates a key advantage of machine learning approaches: extracting signal from negative results that would otherwise represent sunk costs.

Methods note

Cradle ingested all available sequence-function data from the partner's screening and optimization campaigns. Supervised predictors were trained for each of the seven properties, with specialized loss functions to handle the different batch effects and measurement characteristics of each assay (additive for some, multiplicative for others). Potency was assessed via the partner's functional assay, developability metrics via standard industry assays (SMAC, AC-SINS, PAIA, cell binding panels, computational immunogenicity prediction). Full details remain confidential to the commercial partner.