Retention, not flux: endpoint confounding caps computational prediction of peptide skin penetration, with a delivery-aware reframing

Technical Summary: Retention, not flux: endpoint confounding caps computational prediction of peptide skin penetration

Problem Statement
The field of computational dermatology currently faces a stagnation in predicting whether bioactive peptide sequences will successfully reach their intended site of action within the skin. The authors argue that this lack of progress is not due to insufficient modeling power or algorithmic complexity, but rather stems from an ill-posed problem formulation. The dominant paradigm attempts to predict a single binary "skin permeability" label directly from amino acid sequence. The paper contends that this label is fundamentally flawed because it aggregates distinct physical outcomes and ignores critical contextual variables, rendering the learning task impossible to solve with high fidelity using sequence data alone.

Methodology and Analysis
The authors employ a first-principles analysis combined with an empirical study of public-source datasets to deconstruct the sources of predictive failure.

1. Endpoint Deconstruction: The study identifies that the most common experimental endpoint, transdermal flux into a diffusion-cell receptor compartment (aligned with OECD Test Guideline 428), is a confounded metric. It conflates two opposing outcomes: genuine deep delivery (a success for therapeutic/cosmetic actives) and undesired systemic transport (a safety failure). For cosmetic applications, high receptor flux often indicates a failure of retention rather than a success of delivery.
2. Triad Analysis: The authors posit that "permeability" is not an intrinsic property of the peptide sequence ($x$) but a function of a triad: the sequence, the delivery vehicle, and the measurement compartment. Public datasets frequently lack records for the vehicle and specific compartment conditions, introducing uncontrolled noise.
3. Empirical Benchmarking: The authors evaluate predictive performance using public-source data against a "clean" single-endpoint membrane dataset. They compare:
   • Physicochemical intrinsic-permeability estimates (Potts-Guy).
   • Sequence-only classifiers trained on scraped, heterogeneous labels.
   • The same descriptor pipeline applied to a curated, single-endpoint dataset.

Key Results
The empirical analysis reveals that the noise introduced by endpoint confounding severely limits machine learning performance:

• Lack of Signal in Intrinsic Models: Physicochemical estimates (Potts-Guy) show no positive predictive signal for the scraped penetration labels, yielding a grouped Area Under the Curve (AUC) of 0.45 (95% CI 0.40–0.51).
• Plateau in Sequence Models: Sequence-only classifiers trained on the noisy, aggregated public data plateau in the mid-0.70s (AUC 0.70–0.77), showing diminishing returns even as the volume of labeled data increases.
• Performance Recovery in Clean Data: When the same descriptor pipeline is applied to a clean, single-endpoint membrane dataset, performance improves materially to an AUC of 0.83, with a non-overlapping confidence interval compared to the noisy data results.

Key Contributions

• Diagnosis of Stalled Performance: The paper provides evidence that the ceiling on predictive accuracy is caused by "endpoint confounding" in the training labels, not a lack of computational capability.
• Reframing the Problem: The authors propose separating the prediction task into two distinct components:
  1. Barrier-crossing: A data-driven, sequence-level function.
  2. Depth-and-retention: A physics-driven, delivery-aware function.
• Regulatory Perspective: The study reframes intrinsic transdermal flux not as a primary efficacy metric for cosmetics, but as a regulatory risk axis (systemic exposure).
• Proposed Standards: The authors propose a "triad-annotated reporting schema" to ensure future datasets record the sequence, vehicle, and measurement compartment, alongside a seed benchmark to facilitate reproducible research.

Significance and Claims
The paper's primary significance lies in its modest but critical claim: the field's predictive performance is capped by the quality and definition of the target variable, not the sophistication of the models. By demonstrating that removing endpoint confounding (via clean, single-endpoint data) significantly restores predictive power (AUC 0.77 to 0.83), the authors argue that the solution to the "unsolved" problem of peptide skin penetration lies in data curation and problem reframing. The work asserts that barrier-crossing propensity is a legitimate, learnable function of molecular structure, whereas the current vehicle- and endpoint-agnostic binary labels found in literature are not.