Macrophages are the immune system's cleanup crews: in healthy tissue they first switch to an inflammatory mode (conventionally called M1) to fight damage, then switch to a repair mode (M2) to close the wound. Many chronic diseases — osteoarthritis, fibrosis, colitis, valve calcification — are increasingly understood as a stuck switch, with tissues dominated by macrophages that never make the M1-to-M2 transition (TREM2 promotes macrophage polarization from M1 to M2 and suppresses osteoarthritis through the NF-κB/CXCL3 axis; Shared gene characteristics and molecular mechanisms of macrophages M1 polarization in calcified aortic valve disease). Two technical failures keep this switch hard to flip therapeutically. First, researchers pick delivery targets by reading messenger RNA levels, but mRNA abundance predicts protein abundance only 40–60% of the time (A high-stringency blueprint of the human proteome), so a receptor that looks abundant on an M1 cell at the transcript level may be nearly absent as a physical docking site. Second, lipid nanoparticles (LNPs) — the fatty droplets that carry RNA payloads into cells — are screened against immortalized cell lines, not against the distinct inflammatory and repair states a therapy must discriminate between (AGILE platform: a deep learning powered approach to accelerate LNP development for mRNA delivery).
This proposal closes both failures in one loop. The core claim: a single-cell model that predicts protein abundance from transcriptomes can nominate which surface receptors are physically present on inflammatory macrophages; gating LNP ligand design on those predictions, and screening lipids for inflammatory-state selectivity in primary cells rather than cell lines, should yield RNA carriers that selectively reach the stuck macrophages — and the whole pipeline should be judged on protein-level polarization rescue, not transcript readouts.
Core Claim
The existing pieces each stop one step short. Models like scTEL (A joint analysis of single cell transcriptomics and proteomics using transformer) and scChord (scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation) predict surface-protein levels from single-cell RNA data but have never been used to nominate delivery targets. AGILE (AGILE platform) found lipid R6, which transfects RAW 264.7 macrophage cells five-fold better than standard formulations, but treated macrophages as one homogeneous line and measured bulk reporter luminescence. The TREM2 osteoarthritis work (TREM2 promotes macrophage polarization...) proved the therapeutic concept — restoring TREM2 pushes macrophages from M1 to M2 and protects cartilage in mice — but delivered naked siRNA intravenously and systemic recombinant protein, with no cell-selective vehicle and no model of the transcript-protein divergence it empirically observed. The LA5-LNP study (LPA1 antagonist-derived LNPs deliver A20 mRNA and promote anti-fibrotic activities) showed receptor-guided LNP delivery of a pathway-modulating mRNA works, but targeted fibroblasts chosen from bulk literature, not macrophages, and never went in vivo. Nobody has connected the protein-prediction models to the delivery engineering, and nobody has screened delivery vehicles for polarization-state selectivity. That connection, evaluated end-to-end at the protein level in a mouse model of stuck polarization, is what we propose to build.
Research Program
The program has three coupled components, each reusing a technique the file shows works, aimed at one disease anchor (the TREM2/osteoarthritis axis) so the loop terminates in a real efficacy readout.
Target nomination by predicted protein, not transcript. We will train an RNA-to-protein translation model of the scTEL/scChord class on paired CITE-seq data — the GSE164378 PBMC atlas (161,764 cells, 228 surface proteins) and the 37,112-cell monocyte cohort used by scTEL (scTEL) — then apply it to RNA-only profiles of M1- and M2-polarized primary macrophages and inflamed joint tissue, where paired protein measurements do not exist. For each surface protein $j$ we compute a target score
$$T_j = \hat{a}_j^{\mathrm{M1}} \cdot \log_2\!\left(\frac{\hat{a}_j^{\mathrm{M1}}+\epsilon}{\hat{a}_j^{\mathrm{M2}}+\epsilon}\right),$$
where $\hat{a}_j^{\mathrm{M1}}$ and $\hat{a}_j^{\mathrm{M2}}$ are the predicted protein abundances in the two states and $\epsilon$ is a small constant. This ranks receptors that are physically abundant and M1-specific — the opposite of ranking by transcript fold-change, which the 40–60% correlation ceiling (A high-stringency blueprint of the human proteome) and single-cell divergence studies (High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma...) both warn against. Before scoring, droplet artifacts are removed: macrophages adhere tightly to lymphocytes and parenchymal cells, and in the PBMC_16k benchmark 19.3% of droplets were artificial doublets with chimeric surface profiles (Artificial-Cell-Type Aware Cell Type Classification in CITE-seq) — unfiltered, these would nominate receptors that exist on no real cell.
Polarization-selective LNP screening. We will screen a lipid library against primary mouse and human macrophages polarized to M1 and M2 in parallel, not against RAW 264.7. The predictive model will follow LANTERN's finding rather than fight it: on the curated 1,100-lipid benchmark, a plain multi-layer perceptron on molecular fingerprints and physical descriptors reached $R^2 = 0.8161$ while AGILE's graph network managed $R^2 = 0.2655$ (LANTERN: A Machine Learning Framework for Lipid Nanoparticle Transfection Efficiency Prediction) — simple models win in this data regime. The screen optimizes a selectivity index
$$S = \log_2\!\left(\frac{f_{\mathrm{M1}}+\epsilon}{f_{\mathrm{M2}}+\epsilon}\right),$$
where $f$ is the fraction of each state transfected, jointly with absolute M1 transfection, so the carrier reaches the stuck cells and spares the cells already doing repair work — the failure mode the TREM2 paper itself flags when unselective delivery would hit M2 cells too (TREM2 paper). The top-scoring lipid gets its surface decorated with the ligand for the top-ranked receptor, following the conjugation chemistry that gave LA5-LNPs over 100-fold higher delivery than standard MC3 formulations in receptor-expressing cells (LA5-LNP).
Payload and protein-level endpoint. The payload is TREM2 mRNA, the intervention already validated in mice (TREM2 paper); the disease model is the same destabilization-of-the-medial-meniscus (DMM) osteoarthritis model, in both wild-type and TREM2-knockout mice. In a secondary arm we redesign the codons of the payload using the translation-equivariant encoding of Equi-mRNA (Equi-mRNA: Protein Translation Equivariant Encoding for mRNA Language Models): disease states alter tRNA pools in a codon-specific way without any transcriptional shift (Altered tRNA expression profile associated with codon-specific proteomic changes in the suicide brain), so an mRNA optimized for a generic codon table may translate poorly precisely in the cells we target. Every readout is protein-level — flow cytometry for iNOS (M1) and CD206 (M2), the markers validated in the TREM2 study, plus a paired single-cell RNA/ADT measurement of joint macrophages — because the central lesson of the discordance literature is that transcript endpoints can report success while the proteome has not moved.
Proposed Experiments
First, the translation model must reproduce known results before it is trusted on macrophages: train and evaluate on GSE164378 (161,764 cells, 228 proteins), with held-out protein prediction benchmarked against scTEL's reported 86.6% classification accuracy on the same atlas (versus 77.0% for sciPENN and 76.1% for Seurat) and scChord's reported lowest-RMSE/highest-Pearson performance, which it achieved while training on 100,000 cells in roughly 27 minutes with a 229 MB memory footprint. Second, on primary polarized macrophages and DMM-joint single-cell data, predict the surface proteome; verify the top-nominated receptors by flow cytometry, and report the hit rate of the protein-gated list versus a transcript-fold-change list — this comparison is the ablation that isolates the proposal's key ingredient. Third, run the polarization-selective lipid screen (a few hundred lipids spanning the AGILE design space, purified, with AGILE's R6 included as a candidate), measure $S$ in primary M1/M2 co-cultures, and confirm top hits in DMM mice with a paired joint-macrophage CITE-seq readout showing which states were actually transfected. Fourth, the efficacy comparison in DMM mice: protein-level M2 shift (CD206 up, iNOS down), MMP13 and aggrecan levels, and cartilage histology. The TREM2 paper reports these rescue effects qualitatively (recombinant sTREM2 "markedly promotes macrophage polarization from M1 to M2" and improves joint pathology); no reported result found on Lacuna for quantitative DMM cartilage or polarization metrics in that study, so we will establish our own numeric thresholds prospectively, with intravenous recombinant sTREM2 run head-to-head as the efficacy bar. Finally, the codon arm: deliver standard versus Equi-mRNA-redesigned TREM2 sequences at matched RNA dose and measure TREM2 protein per transfected cell, testing whether codon choice moves protein yield in M1 macrophages specifically.
Baselines
The do-nothing reference is vehicle-treated DMM mice and untreated disease progression, which anchors the effect size; the naive delivery reference is an untargeted MC3-standard LNP carrying the same TREM2 mRNA, since MC3 is the industry-standard formulation that AGILE, LANTERN, and the LA5-LNP study all use as the floor. The strongest existing delivery competitor is AGILE's R6 lipid, which delivered a five-fold improvement in macrophage transfection efficiency over MC3 and the muscle-tropic H9 in RAW 264.7 cells (AGILE platform); R6 will be synthesized and included in our primary-cell screen under identical conditions, dose, and mRNA payload, because if R6 already discriminates M1 from M2 — a property its RAW 264.7-only screen never tested — the lipid-discovery component of this proposal is unnecessary and we would report that. The strongest therapeutic competitor is intravenous recombinant sTREM2, the intervention the TREM2 paper showed reverses cartilage breakdown in DMM mice; we will run it dose-matched in the same cohorts, and beating it means achieving equal or better protein-level polarization shift and cartilage protection at a lower systemic dose, with transfected cells verifiably restricted to M1-dominated tissue. On the computational side, the strongest competitor for target nomination is scTEL with its 86.6% accuracy on GSE164378, plus scChord as the distribution-reconstruction alternative; our model must match these on held-out proteins while adding macrophage-state outputs they lack, and all models get the same training data and tuning budget. The key ablation is a single change: keep the lipid, the ligand chemistry, the payload, and the mouse model, but select the targeting receptor by transcript fold-change instead of predicted protein abundance. If the protein-gated and transcript-gated carriers perform equivalently in vivo, the discordance-gating claim is falsified and the delivery engineering stands on its own.
Why not just…?
Why not just pick targets from scRNA-seq? This is the cheapest alternative and the one the proposal is built against. The correlation ceiling of 40–60% between transcript and protein abundance (A high-stringency blueprint of the human proteome) means up to half the variance in surface-receptor abundance is invisible at the transcript level, and scTEL demonstrated concretely that RNA data alone cannot resolve immune subsets that protein markers (CD45RA, CD38) distinguish (scTEL). The transcript-versus-protein ablation above is designed to settle this empirically rather than by argument: if the transcript-gated carrier performs as well, the simpler route wins and we will say so.
Why not just use the measured antibody data directly, with no prediction model? This is a genuine threat. Where paired CITE-seq exists, direct antibody-derived tag measurement is ground truth and no model is needed. But antibody panels rarely exceed 200–300 surface markers, and the macrophage tissues we target — arthritic joints, fibrotic lung — lack paired proteomics entirely, which is exactly the regime the transfer model is for; it extends protein inference to unmeasured proteins and RNA-only tissue datasets. The decisive check is the held-out-protein benchmark on GSE164378: if the model adds nothing over measured ADT plus literature, we drop it and use direct measurement.
Why not just deliver the RNA with AGILE's R6 lipid and skip the screen? R6 is macrophage-tropic, and it will be in our library. But its five-fold advantage was measured in a single immortalized line with no M2 arm, and the TREM2 study shows the two polarization states differ in the markers and cytokines that govern uptake; non-selective transfection risks pushing M2 cells around too. R6's $S$ index in primary M1/M2 co-culture is the first number we will report — if it is already high, we concede the point and the program simplifies to ligand gating on top of R6.
Why not just give systemic recombinant sTREM2 or steroids? Systemic sTREM2 already works in the DMM mouse (TREM2 paper), so it is the bar, not a strawman. But broad immunosuppression fails precisely in the imbalanced-inflammation setting: elevated IL-6 creates a Th17/Treg imbalance that is fundamentally refractory to corticosteroids (Risk factors for steroid-refractory in immune checkpoint inhibitor-induced colitis), and unguided systemic agents sequester in clearance organs while diluting at diseased sites (Diabetic cardiomyopathy: The role of microRNAs and long non-coding RNAs). Recombinant protein also cannot be codon-tuned, and its production and half-life costs are the reason LA5-type mRNA delivery exists. The dose-matched head-to-head in DMM mice settles which margin survives in practice.
Why not just use a large graph transformer for lipid design? Because the evidence already refutes it: LANTERN showed complex graph models over-smooth in the ~1,000-lipid regime, with the simple fingerprint-plus-descriptors model at $R^2 = 0.8161$ versus $0.2655$ for the graph network, and correctly identified top-performing lipids 73.7% of the time (LANTERN). We adopt the simple model; there is no argument to make around it.
Decisive Checks
The proposal stands or falls on three numbers. The protein-gated versus transcript-gated carrier comparison in DMM mice (the ablation) tests the core claim that discordance-aware target selection matters. R6's selectivity index in primary M1/M2 cells tests whether polarization-selective screening is a real need or an artifact of cell-line screens. And the matched-dose comparison against intravenous recombinant sTREM2 tests whether targeted RNA delivery earns its complexity over an already-working systemic protein — no reported result found on Lacuna for a quantitative sTREM2-versus-RNA head-to-head, which is exactly the hole this program fills.
References
- TREM2 promotes macrophage polarization from M1 to M2 and suppresses osteoarthritis through the NF-κB/CXCL3 axis
- Shared gene characteristics and molecular mechanisms of macrophages M1 polarization in calcified aortic valve disease
- A high-stringency blueprint of the human proteome
- AGILE platform: a deep learning powered approach to accelerate LNP development for mRNA delivery
- A joint analysis of single cell transcriptomics and proteomics using transformer
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- LPA1 antagonist-derived LNPs deliver A20 mRNA and promote anti-fibrotic activities
- High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy
- Artificial-Cell-Type Aware Cell Type Classification in CITE-seq
- LANTERN: A Machine Learning Framework for Lipid Nanoparticle Transfection Efficiency Prediction
- Equi-mRNA: Protein Translation Equivariant Encoding for mRNA Language Models
- Altered tRNA expression profile associated with codon-specific proteomic changes in the suicide brain
- Risk factors for steroid-refractory in immune checkpoint inhibitor-induced colitis: a retrospective cohort study
- Diabetic cardiomyopathy: The role of microRNAs and long non-coding RNAs