DeepDelivery mines the human proteome for mRNA delivery scaffolds: 512 candidates, and 12 of 15 tested form particles that deliver
Synopsis
DeepDelivery treats the human proteome as a candidate pool for mRNA delivery scaffolds: deep learning models trained on viral capsid proteins, filtered by a dual-model screen, nominated 512 candidates from 109,375 human sequences. Of 15 candidates tested experimentally, 12 formed mRNA-encapsulating particles and mediated functional delivery; RTL8B reached 73.2% transduction in HEK293T reporter cells, above the reference proteins Arc (56.5%), PEG10 (53.1%) and PNMA2 (67.4%). The results support human proteins as a source of modularly engineerable delivery scaffolds, while immunogenicity and in vivo behavior remain untested.
Interpretation
The candidate pool for mRNA delivery scaffolds widens from a few known viral-capsid or retroelement-derived proteins to a set of human proteins not previously associated with self-assembly. Systematic discovery had been constrained by homology-based searches biased toward retroelement-derived genes and prone to miss proteins with divergent or cryptic assembly domains; only a limited number of capsid-like human proteins were known, with PEG10 a rare functional example. Across 109,375 human sequences of 100 to 1024 amino acids, 512 candidates were classified positive by both models, 0.47% of the screened proteome; the previously reported Arc and PEG10 were recovered by this dual-model screen.
Nominated candidates actually assemble into particles and deliver mRNA in cells, not merely score highly in the model. Model prediction and experimental function needed a direct test, because a prediction score alone does not show that particles form or that cargo reaches recipient cells. Of 18 candidate proteins including positive controls, 15 formed regular capsid-like structures by negative-stain TEM, and all 15 triggered EGFP expression through a Cre-loxP reporter; RTL8B transduced at 73.2%, versus Arc at 56.5%, PEG10 at 53.1% and PNMA2 at 67.4%.
Domain importance read out from the model can guide engineering that improves a middling scaffold. TRIM25 had a high prediction score but delivered at only 41.3%, a gap between in silico prediction and measured efficiency. After layer-wise relevance propagation, which propagates the network output back to each amino acid to score its contribution, pointed to the RBCC domains, the TRIM25-RCD variant lacking the B30.2/SPRY domain raised transduction in HEK293T cells from 37.8% for full-length TRIM25 to 65.4%, a 27.6 percentage-point gain.
Perspective
The demonstrated scope is two reporter cell lines, HEK293T and HepG2, where candidate scaffolds assembled through the PLCD-P6-MS2 modular framework formed particles and delivered Cre-NLS mRNA; immune assessment is limited to in vitro markers in THP-1 macrophages and human peripheral blood-derived immature dendritic cells. The source states that pharmacokinetics, biodistribution, tissue tropism, repeat dosing and immune responses require in vivo studies.
The authors frame the immune analysis as preliminary: broader primary immune-cell profiling, plasmacytoid dendritic-cell assays, repeated-dose studies and in vivo evaluation remain to be done. Scaffold performance also cannot be inferred from a single readout, since yield, cargo loading, recovery, particle dose and cell-type context all influence apparent delivery activity. Three false-positive candidates (RTL8C, PGBD1, RN186) show the model still misses some assembly determinants, and the authors suggest future models integrate local sequence, global domain organization and assembly-compatible structural features.