English · 简体中文

SecRank

Ranking signal peptides in the context that matters: their cargo.

An ongoing research project for cargo-aware prediction and ranking of signal-peptide secretion performance.

Important

SecRank is under active development. This repository currently documents the research direction and architecture; model code, training data, checkpoints, and runnable prediction tools have not been released.

Architecture preview

_{SecRank v0.7 research architecture. The diagram describes the current design direction, not a released performance claim.}

At a glance

• Question — which signal peptide is expected to perform better for a given cargo?
• Context — the signal peptide and cargo sequence are represented as one secretion construct.
• Learning objective — compare constructs pairwise and learn their relative ordering.
• Current stage — architecture development, validation, and release preparation are ongoing.

The architecture can be read from left to right:

1. Each construct is separated into its signal peptide and cargo sequence.
2. Frozen ESM-2 protein language model representations are combined with AAindex physicochemical descriptors for the signal peptide.
3. Feature fusion and gated cross-attention build a cargo-aware construct representation.
4. A shared pairwise comparator learns the ranking objective while continuous difference and ratio heads are explored where the experimental measurements support them.

Why cargo context?

SecRank is built around the research hypothesis that secretion performance should be evaluated for a signal-peptide–cargo construct, rather than treating the signal peptide as an isolated sequence. The model therefore shares one construct encoder across a pair and compares the resulting representations.

Research scope

• cargo-aware representation of signal-peptide secretion constructs;
• pairwise ranking of relative secretion performance;
• continuous-difference and continuous-ratio supervision when justified by the measurement semantics;
• reproducible data curation, evaluation, and release documentation.

Project status

• Public now: bilingual project overview and the v0.7 architecture.
• In progress: architecture comparison, validation, and reproducibility work.
• Not released yet: model code, datasets, checkpoints, and prediction interfaces.

Roadmap

• Launch the public, bilingual SecRank project page
• Publish the v0.7 architecture preview
• Add architecture notes and validation documentation
• Release reproducible model interfaces and worked examples
• Publish model cards, qualified checkpoints, and distributable data where licensing and provenance permit

Availability

SecRank is not yet a production-ready model, and this repository does not make a general performance claim. Public research artifacts will be considered for gradual release after the corresponding validation, reproducibility, licensing, and data-provenance requirements are met.

_{Research in progress · Public documentation first · Reproducible artifacts when ready}