Merck & Co., Inc., Rahway, NJ, USA (United States)

Noncanonical amino acids push peptide design spaces to the 10⁵⁰ scale, where exhaustive screening is infeasible. PepOpt encodes each residue as a position-specific molecular fingerprint, predicts potency with linear regression, and uses dynamic programming under an additivity assumption to find the globally optimal sequence deterministically. Across 11 internal peptide series it matches or beats nonlinear models, and in 10²⁰ and 10⁵⁰ spaces it returns higher predicted scores than a genetic algorithm.

Noncanonical amino acids push peptide design spaces to the 10⁵⁰ scale, where exhaustive screening is infeasible. PepOpt encodes each residue as a position-specific molecular fingerprint, predicts potency with linear regression, and uses dynamic programming under an additivity assumption to find the globally optimal sequence deterministically. Across 11 internal peptide series it matches or beats nonlinear models, and in 10²⁰ and 10⁵⁰ spaces it returns higher predicted scores than a genetic algorithm.

In a multi-agent hypothesis-generation workflow, the authors held opening hypotheses fixed and replayed downstream stages, using an LLM-as-a-judge to decide whether paired hypotheses describe the same mechanism: moving from 0 to 1 self-critique round produced 34.5 percentage points of extra mechanism-level divergence while 1 to 5 rounds added 1.3 pp; under controlled perturbations all three equivalence rules were invariant to meaning-preserving edits, but when the initiating event was replaced the LLM-as-a-judge read 83% of pairs as different mechanisms versus 0% for TF-IDF and 8% for embeddings, and changing only the rubric moved this rate from 38% to 96%.

In a multi-agent hypothesis-generation workflow, the authors held opening hypotheses fixed and replayed downstream stages, using an LLM-as-a-judge to decide whether paired hypotheses describe the same mechanism: moving from 0 to 1 self-critique round produced 34.5 percentage points of extra mechanism-level divergence while 1 to 5 rounds added 1.3 pp; under controlled perturbations all three equivalence rules were invariant to meaning-preserving edits, but when the initiating event was replaced the LLM-as-a-judge read 83% of pairs as different mechanisms versus 0% for TF-IDF and 8% for embeddings, and changing only the rubric moved this rate from 38% to 96%.