ApexSP

Recombinant proteins underpin biopharmaceuticals, industrial enzymes, and molecular diagnostics, but getting them out of the cell remains a production bottleneck. Secretion is initiated by the signal peptide, a short N-terminal sequence that routes a protein into the Sec or Tat export pathway. In practice, engineers reach for a handful of well-worn standards and screen empirically, because secretion efficiency is not a property of the signal peptide alone. It emerges from the compatibility among the signal peptide, the cargo protein it is fused to, the export pathway, and the host background.

ApexSP, developed at Shandong University and posted as a bioRxiv preprint in July 2026, treats this as a design problem rather than a selection problem. It is a three-stage cascade: a pathway-conditioned discrete diffusion model that generates candidate signal peptide sequences, a topology-aware multitask filter that discards candidates violating the intended pathway, N/H/C-region architecture, or taxonomic context, and a compatibility ranking model that scores each surviving candidate against the specific cargo protein it must carry. All three components are trained once and applied to new cargo proteins at inference time.

That framing separates ApexSP from established tools such as SignalP 6 and USPNet, which classify pathways and annotate cleavage sites but do not generate new sequences. ApexSP absorbs those prediction tasks into its filter while producing designs intended for direct experimental use.

#Key Features

• Evolution-constrained diffusion: The forward noising process corrupts residues according to a temperature-scaled BLOSUM62 substitution matrix rather than uniform random replacement. This lifted top-5 amino acid reconstruction accuracy from 0.9522 to 0.9744 and Tat twin-arginine motif recovery from 0.9028 to 0.9384 over a positional-embedding baseline.
• Pathway-conditioned sampling: Reverse diffusion is conditioned on a Sec or Tat label, with a repetition-blocking rule that suppresses homopolymeric runs during decoding.
• Topology-aware multitask filter: A single model jointly predicts secretion pathway, residue-level N/H/C topology via a conditional random field, and prokaryotic versus eukaryotic origin, reaching 94.3% exact-match cleavage-site accuracy against 58.6% for USPNet and 55.6% for SignalP 6 on the SPSDB benchmark.
• Cargo-aware ranking: A dual-branch encoder scores each candidate against the N-terminal 200 residues of the mature protein, which the authors identified as the informative window.
• Small experimental burden: In each validation system, 100,000 generated candidates were narrowed to 10 for wet-lab testing, and designed peptides beat the reference leader in every system.

#Technical Details

The underlying knowledge base integrates UniRef50, UniProtKB/Swiss-Prot, and the Signal6, USP_val, SPSDB, and SPE datasets. After CD-HIT clustering at 95% identity and two-stage consensus annotation by USPNet and SignalP 6, UniRef50 yielded 3.57 million Sec and roughly 74,000 Tat signal peptides, with Swiss-Prot contributing more than 200,000 high-quality signal peptide–mature protein pairs. The generator is a D3PM-style discrete diffusion model over a 20-amino-acid vocabulary at up to 70 residues, using 100 diffusion steps on a cosine schedule; its denoiser combines a convolutional embedding stack, positional, timestep, and pathway embeddings, a four-layer Transformer encoder, and a convolutional decoder.

The filter fine-tunes ESM-2 650M (esm2_t33_650M_UR50D) with DoRA adapters at rank 16, concatenating hydrophobicity, volume, charge, and positional features onto each residue representation. Pathway accuracy stayed at or above 96.2% across all signal peptide length bins on a held-out UniRef50 set, and on SPSDB N-terminal fragments it reached 79.2% against 68.9% for USPNet and 64.2% for SignalP 6. The ranking model uses two parameter-independent encoders over 128-dimensional learned residue embeddings, pretrained on pseudo-labeled natural pairs and fine-tuned on measured secretion yields from the AmyQ–Bacillus subtilis system, reaching nDCG@20 of 0.472.

#Applications

ApexSP targets secretion engineering: replacing the trial-and-error screen of signal peptide libraries with a short, ranked candidate list tailored to one cargo protein and host. The authors validated it on glucoamylase from Aureobasidium pullulans expressed in Pichia pastoris, where the best design, AE10, reached 1,448.3 U/mL against 847.5 U/mL for the α-factor leader — a 1.71-fold improvement — and on protease AprE2 in B. subtilis through both Sec and Tat routes, where four designs beat the endogenous and PhoD references respectively. The filter also stands alone as a pathway, topology, and cleavage-site predictor.

#Impact

ApexSP carries signal peptide work from attribute prediction through to experimentally validated cargo-specific design, and its one-round hit rate is the most practical claim: useful sequences from ten constructs rather than hundreds. The work remains a preprint awaiting peer review, and the evidence base is three cargo–host–pathway combinations, so generality across cargo classes is untested. Ranking quality varies by system, correlating at r = 0.529 with measured yields in a Corynebacterium glutamicum cutinase system but only r = 0.141 in a Lactiplantibacillus plantarum nuclease system. Weights and inference scripts are released on Hugging Face under CC-BY-NC-SA-4.0, a non-commercial license that excludes the industrial biotechnology users the method most directly serves; there is no separate code repository, and the released scripts carry absolute paths from the authors' original server that must be edited before they run elsewhere.