Engineering therapeutic peptides with 50% success rate under tight multi-property constraints, 5× faster than prior efforts

Context

Therapeutic peptides occupy a valuable niche between small molecules and biologics, offering target specificity approaching antibodies with improved tissue penetration and manufacturing simplicity. However, peptide drug development faces characteristic challenges: proteolytic instability, limited oral bioavailability, potential off-target binding, and manufacturability concerns.

Late-stage peptide optimization requires balancing these properties within tight specification windows. Unlike early discovery where approximate improvements suffice, late-stage programs demand that all properties simultaneously meet precisely defined thresholds—a single property falling outside specification can halt clinical progression.

Challenge

The commercial partner had three peptide programs, each at a late stage and close to a dead end following multiple prior rounds of optimization with limited remaining sequence flexibility. The key challenge was meeting tight, simultaneous constraints on potency, specificity, expression, and thermostability. Previous in-house and external optimization efforts had failed to deliver sequences satisfying the full specification across all properties.

The "close to a dead end" characterization reflected exhaustion of the rational design strategy: with limited mutations permitted (to maintain the peptide's core structure and mechanism) and tight constraints on all properties, the partner had tested most obvious sequence variants without finding candidates that satisfied all requirements simultaneously.

Approach

We performed a single round of Cradle optimization for each of the three programs, generating 96 candidates per program. The optimization operated under explicit multi-property constraints rather than trade-off-based objectives. Each candidate was evaluated as either constraint-compliant (meeting all thresholds) or non-compliant (failing one or more thresholds), with potency optimization performed only within the compliant subset.

This constraint-satisfaction framing differs from typical Pareto optimization: rather than maximizing multiple properties and accepting trade-offs, the goal was to find the subset of sequence space where all properties simultaneously exceeded their specified thresholds, and then maximize potency within that subset.

Results

Cradle generated candidate libraries with a ≥50% success rate in matching all constraints:

The ≥50% constraint compliance rate means that in each 96-candidate library, at least 48 variants met all four property thresholds (potency range, specificity threshold, expression threshold, thermostability threshold). This contrasts sharply with prior campaigns where most or all variants failed one or more constraints.

Among the constraint-compliant sequences, Cradle delivered significant potency improvements over the template. The partner advanced multiple optimized peptides from each program to downstream validation, representing tangible program progression from stalled late-stage optimization to development candidates.

The single-round completion for all three programs, compared to the partner's estimate of ~5 rounds historically, represents a 5× acceleration in development timelines. For late-stage programs where every month of delay affects patent life and commercial opportunity, this timeline compression has direct business value.

What this means

This result demonstrates Cradle's effectiveness for the most constrained optimization problems: late-stage programs with limited sequence flexibility and tight simultaneous constraints on multiple properties. The ≥50% hit rate in a single round indicates that Cradle's supervised predictors accurately modeled the multi-dimensional constraint boundaries, generating candidates concentrated in the viable region of sequence space.

The ability to rescue three separate dead-end programs validates the generalizability of Cradle's approach across different peptide sequences and therapeutic targets. Rather than requiring target-specific or modality-specific tuning, the same system successfully addressed all three programs using only their respective sequence-function datasets.

For pharmaceutical development, late-stage optimization delays represent particularly expensive bottlenecks: discovery and early development investments are sunk, clinical timelines are delayed, and patent clocks continue running. The 5× speed improvement therefore translates to both cost savings (fewer experimental rounds) and opportunity value (faster time to clinic).

Methods note

Cradle consumed all available sequence-function data from the partner's prior optimization efforts for each program. All properties were predicted jointly by a single multi-property model, utilizing a specialized multi-property Spearman rank correlation for model selection. Evotuning was performed on each peptide's evolutionary alignment. Constraint compliance was evaluated via the partner's validated assays for potency, specificity (selectivity panel), expression (yield in standard production system), and thermostability (DSF or similar). Diversity-aware selection maximized the probability of selecting candidates that were both high-performing and informative, using stochastic dropout to identify distinct "peaks" in the functional landscape. Full details of the peptide sequences, targets, and assay protocols remain confidential to the commercial partner.