A screen of 15,360 fully random, sub-kilodalton cyclic peptides yielded a starting point against the intracellular Keap1–Nrf2 protein–protein interaction, and iterative design, synthesis, and testing converted it into a membrane-permeable inhibitor active in living cells, according to a 2026 study published in Nature Chemical Biology. Karsten Eastman, PhD, CEO and co-founder of Sethera Therapeutics, cites the work as evidence that binding and permeability can be treated as connected design problems rather than sequential hurdles.

Eastman's argument, laid out in a GEN commentary, targets a structural weakness in peptide discovery portfolios: programs organized around a single objective — identify the strongest binder — defer the questions that determine whether a molecule can become a drug. Selectivity, solubility, synthetic tractability, stability, and compartment access all surface later, when changing molecular direction costs more.

The library is already a hypothesis

No peptide library is truly unbiased, Eastman writes. Length distribution, residue alphabet, cyclization chemistry, topology, display format, and synthesis method determine which molecules can be produced and which conformations the screen can explore. Those choices also shape stability, solubility, permeability, and presentation of binding groups.

Sequence alone does not encode function. Two molecules with similar residue composition can behave differently when one is linear and the other cyclic, when stereochemistry changes, or when a backbone amide is modified. For constrained peptides, linkage position and ring topology can reorganize the conformational ensemble — meaning a sequence-only analysis may group together molecules that are chemically and pharmacologically distinct.

Screen in biological space

Purified-protein screens remain useful when material is limited or throughput matters, but the assay format can become an unintended selection pressure. Peptides may recognize a purification tag, a surface, an exposed hydrophobic patch, or a conformation poorly represented in the native setting. High apparent affinity can also reflect nonspecific interactions that vanish in another assay.

The remedy, Eastman argues, is not to force every primary screen into cells but to design a staged assay hierarchy before screening begins. Early counterselections can remove binders to tags, matrices, related proteins, or abundant off-targets. Competition experiments test epitope dependence; homolog counterscreens reveal selectivity within a target family. Orthogonal confirmation through kinetic binding, solution-phase competition, biochemical function, or cell-based activity distinguishes reproducible target engagement from assay-specific behavior.

For intracellular programs, permeability and functional activity should enter the workflow as soon as candidate numbers permit. For extracellular targets, serum stability, target turnover, tissue context, and the consequences of sustained versus transient engagement may matter more.

Time as a variable

Most screening readouts are snapshots. An endpoint measurement can obscure differences in association rate, dissociation rate, target rebinding, internalization, degradation, or intracellular retention. A peptide with modest equilibrium affinity but a slow off-rate may produce stronger functional activity than a tighter binder that dissociates rapidly. Conversely, prolonged engagement may be undesirable when an off-target interaction creates risk.

Eastman proposes practical routes to incorporating time: extend wash periods or add soluble competitor to raise selection pressure; test candidates after defined exposure to serum, proteases, reducing conditions, or relevant tissue fluids; separate immediate pathway modulation from activity that persists after washout; and measure internalization and cytosolic access at multiple time points rather than inferring from one image. Round-by-round enrichment data also carries temporal signal — a sequence that rises steadily may be more credible than one that appears abruptly after a bottleneck.

Failure modes are data

The final hit list, Eastman contends, is often the least informative version of a screening dataset. It contains winners but discards the evidence needed to understand why they won. "Inactive" is not one label: a candidate may have failed production, display, cyclization, or modification; aggregated; degraded; bound nonspecifically; failed to enter cells; or reached the target without changing function. Models trained on collapsed negatives would be learning biology from a mixture of technical and pharmacological failures.

Useful datasets require provenance: starting-library abundance, enrichment by round, control behavior, counterscreen results, synthesis yield, purity, modification efficiency, assay conditions, batch identity, and detection limits should travel with each sequence. Missing values need interpretation too — "not detected" can mean below an assay threshold, absent from the starting population, lost during processing, or simply not tested.

Assign physics and ML complementary roles

Physics-based modeling suits conformational ensembles, intramolecular hydrogen bonding, solvent exposure, target contacts, and substitution effects — especially when experimental data are sparse — but carries computational cost, sampling challenges, and model sensitivity. Machine learning can recognize relationships across larger datasets, rank candidates, and support multi-parameter optimization. Recent work published in Nature Machine Intelligence coupled generative models with Bayesian optimization and prospective synthesis and testing to improve peptide scaffolds under experimentally defined constraints, though limited training data remain a central challenge for deep-learning cyclic-peptide structure prediction.

The productive workflow, Eastman writes, uses physical insight to define plausible chemical space, experimental data to anchor predictions, and machine learning to identify the next informative measurements. Teams should examine trade-offs directly — a modest loss in affinity may be acceptable if it buys a major gain in selectivity, stability, permeability, or manufacturability — rather than collapsing every objective into one opaque score.

Five decisions before the first hit

Eastman frames an integrated campaign around five decisions: define the intended product profile early, including compartment, route, dosing, selectivity, and minimum functional effect; treat molecular architecture — constraint, stereochemistry, backbone composition, topology — as an explicit variable; establish the assay hierarchy and controls in advance so each measurement answers a distinct question; retain complete, interpretable data including technical failures; and test models prospectively, since a model that explains an existing dataset may still fail on new architecture classes.

The output of a screen, he concludes, should be a calibrated map showing which sequences and architectures succeed under which conditions — not merely a ranked list. Programs that move most efficiently will treat library design, screening, functional biology, physics, and machine learning as parts of one experimental system, built to understand how to make the next molecule better.