Neurotransmitters are not just signals between nerve cells; they also act directly on immune cells. Dopamine, for example, can switch inflammation off in microglia — the brain's resident immune cells — by suppressing the Toll-like receptor 4 (TLR4) to Nuclear Factor kappa-light-chain-enhancer of activated B cells (NF-κB) cascade: in one study, 10 μM dopamine took activated amoeboid microglia from 5.8% down to 0.64% of cells and restored phagocytosis from 9.6% back to 45.7%, near the 47.2% healthy baseline, with pro-inflammatory transcripts (TNF-α, IL-1β, CD86, iNOS) all falling (Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway). But that study, like the rest of this literature, measured only total transcript levels and protein bands. Meanwhile, work on alternative splicing — the process by which a single gene's RNA is cut and rejoined into different isoforms — has shown that splicing is a regulatory layer orthogonal to total expression: a cell can keep its mRNA counts identical while producing entirely different functional isoforms, including truncated proteins that act as built-in brakes on the full-length enzyme (Alternative splicing across the C. elegans nervous system, Upstream and Competitive Regulation of Alternative Splicing).
This proposal asks a question nobody has asked: does the dopamine immune brake in microglia work partly by rewriting which splice isoforms the cells produce, and can we watch that rewrite happen live? The evidence making this plausible comes from splicing biology itself: isoform choice is governed by reversible phosphorylation of splicing factors — protein phosphatase 1 (PP1) activation flips Bcl-x and caspase 9 isoforms (Upstream and Competitive Regulation of Alternative Splicing) and CDC-like kinase 1 (CLK1) inhibition switches ERCC1 isoforms (Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching) — precisely the kind of kinase and phosphatase cascades that G-protein-coupled dopamine receptors engage. Yet the dominant splicing predictors condition only on sequence, RNA structure, and baseline expression of splicing factors, and are therefore structurally blind to this route. The proposal builds a new measurement instrument, uses it for a genome-scale perturbation screen, and extends splicing prediction with a signaling-state input, closing a genuine three-way gap: no study in the surveyed literature connects neurotransmitter signaling to splicing programs in immune cells, nor uses optical reporter assays to measure splice choice.
Core Claim
The core claim is two-part. First, instrumentally: a dual-luciferase minigene cassette, in which exon inclusion puts firefly luciferase in frame and exon skipping puts Renilla luciferase in frame, converts splice-isoform selection into a ratiometric optical signal that can be read continuously in living cells — merging the noise-cancelling logic of dual-luciferase transcription assays (Measuring Transcriptional Activity with Dual Luciferase Assays) with the exon-inclusion logic of minigene reporters (SR proteins regulate V6 exon splicing of CD44 pre-mRNA). Second, biologically: dopamine's suppression of the TLR4–NF-κB cascade in microglia is executed in part through rapid, kinase-driven splicing reprogramming — for instance dominant-negative truncated isoforms of iNOS or of TLR-pathway adapters such as MyD88, the same adapter shown to drive neuron-to-microglia inflammatory signaling (TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation) — a mechanism invisible to the total-mRNA assays used in Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway.
The optical readout is a calibrated inclusion estimate. With $L_F$ the firefly light output (proportional to inclusion) and $L_R$ the Renilla output (proportional to skipping), after a one-time calibration against gel-densitometric ground truth we compute
$$\Psi_{\text{luc}} = \frac{L_F}{L_F + \alpha L_R},$$
where $\alpha$ is the per-construct calibration constant that equalizes the two enzymes' expression and detection efficiencies; $\Psi_{\text{luc}}$ is then directly comparable to the percent-spliced-in ($\Psi$) values used in transcriptomics. Because it is a within-well ratio, it inherits the property that made dual-luciferase assays reliable — cancellation of cell death, transfection efficiency, and pipetting variance (Measuring Transcriptional Activity with Dual Luciferase Assays, Measuring Transcriptional Activity with Dual Luciferase Assays).
Research Program
The program has three linked components: the reporter instrument ("PSI-luc"), the biological study and screen in microglia, and a signaling-conditioned extension of splicing prediction.
Building PSI-luc starts from the validated CD44 variable-exon-6 cassette of the pFlare minigene system (SR proteins regulate V6 exon splicing of CD44 pre-mRNA) as the calibration standard, then generalizes to microglial inflammatory genes. Two reciprocal reading-frame constructs are made per target exon: one in which inclusion yields in-frame firefly luciferase, and one in which skipping yields in-frame Renilla. A matched construct set for candidate microglial switches (iNOS, MyD88, Traf6, Tlr4 itself) follows. Calibration exploits the quantitative anchors already in the literature: overexpression of SRSF1, SRSF6, or SRSF9 raised CD44 V6 skipping by roughly 36–46% while SRSF3 and SRSF4 had no effect (SR proteins regulate V6 exon splicing of CD44 pre-mRNA) — the reporter must reproduce both the direction and approximate magnitude of these shifts before it is trusted on new constructs.
The biological study runs in BV-2 microglia under the exact stimulation matrix of Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway: 200 ng/ml LPS, with 2 μM and 10 μM dopamine, plus dopamine-receptor subtype antagonists (D1-like and D2-like) — resolving which subtype couples to the brake, an explicitly open problem in that paper. Live $\Psi_{\text{luc}}$ kinetics are recorded over 48 hours, the same tracking window that made the NF-κB–GFP signalome screen informative (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy). In parallel, bulk RNA-seq of the same conditions (0, 2, 6, 24 hours), quantified with junction-level $\Psi$ estimation in the MAJIQ style used in the C. elegans neuronal atlas (Alternative splicing across the C. elegans nervous system), maps the endogenous dopamine-responsive splicing program and validates every reporter hit against the native locus — a mandatory step, since the literature shows episomal reporter readouts can diverge from endogenous tissue measurements when compensatory networks operate (A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population).
The screen mirrors the arrayed CRISPR design of the signalome study — which targeted 706/708 kinases, phosphatases, and adaptors with a live NF-κB readout and recovered canonical regulators plus novel ones invisible at the mRNA level (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy) — but points it at the splicing machinery: an arrayed knockdown library covering the SR and hnRNP factor families and splicing-relevant kinases and phosphatases (CLK1, PP1 subunits), with dual readouts in each well, the $\Psi_{\text{luc}}$ ratio and a conventional NF-κB luciferase reporter. The screen asks which splicing regulators are required for dopamine to shift isoform choice and, separately, which are required for the NF-κB brake itself. Factors that satisfy both are candidate executors of the splicing arm of the brake.
The computational component extends splicing prediction with a signaling-state modality. Current models condition on static inputs only: CellSpliceNet integrates sequence patches, ViennaRNA structure graphs, junction regions, and baseline expression of 243 splicing factors, reaching Spearman correlation $\rho = 0.88$ on held-out C. elegans data, with removing any single modality costing at least 0.06 in $\rho$ (CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons); the Discrete Compositional Energy Network (DCEN) conditions on a 3,971-dimensional tissue-specific regulatory factor vector, reaching Spearman 0.623 and Pearson 0.651 versus 0.579 and 0.574 for the best SpliceAI-based baseline (RNA Alternative Splicing Prediction with Discrete Compositional Energy Network). None of them receive receptor-signaling flux as input, yet the Bcl-x, caspase 9, and ERCC1 results show isoform switches are driven by phosphorylation changes with no change in splicing-factor transcript abundance (Upstream and Competitive Regulation of Alternative Splicing, Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching). We therefore retrain a mouse model — using the modular OpenSpliceAI implementation, whose transfer-learning recipe converges in one epoch versus ten from scratch and whose species-specific retraining produced gains of up to 57% in Arabidopsis and 54% in honeybee over human-trained SpliceAI (OpenSpliceAI provides an efficient modular implementation of SpliceAI enabling easy retraining across nonhuman species) — augmented with a small conditioning vector encoding agonist identity, dose, receptor antagonist, and time. The decisive test is generalization: train on dopamine conditions and held-out doses, predict the $\Delta\Psi$ map for the calcitonin gene-related peptide (CGRP)/RAMP1 axis, the other established neuropeptide immune brake (TRPV1+ neurons alter Staphylococcus aureus skin infection outcomes by affecting macrophage polarization and neutrophil recruitment, RAMP1 signaling attenuates acute lung injury by inhibiting cytokine production and neutrophil recruitment).
Baselines
The "do nothing" reference point is the field's current default: endpoint total-mRNA measurement, exactly as Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway measured TNF-α, IL-1β, CD86, and iNOS by reverse-transcription quantitative PCR (RT-qPCR) with no isoform resolution. It is the strongest competitor in the trivial sense that it is what any lab would run first, and it already reproduces the dopamine brake at the total-transcript level. We run it in parallel on every condition, on the same wells, so the comparison is exact. Beating it means demonstrating dopamine-induced isoform switches at genes whose total mRNA is statistically unchanged — the orthogonality scenario established across 55 neuron types, where 1,940 genes underwent differential splicing independently of expression (Alternative splicing across the C. elegans nervous system) — and showing that truncated dominant-negative isoforms of the NOS type (Upstream and Competitive Regulation of Alternative Splicing) are undetectable by those primers.
The strongest existing measurement method is the pFlare minigene with endpoint RT-PCR and gel densitometry, the technique that produced the 36–46% V6-skipping quantification (SR proteins regulate V6 exon splicing of CD44 pre-mRNA). It is the competitor because it is validated, cheap, and isoform-resolving, which the RT-qPCR baseline is not. We run it on every construct and every condition in the study — same wells, same transfection batches, split for optical and gel readout — so throughput and time-resolution are the only variables. Beating it means concordance of $\Psi_{\text{luc}}$ with gel-derived $\Psi$ across at least 40 condition-construct pairs (target rank correlation ≥ 0.85, matching the median Pearson 0.857 that OpenSpliceAI achieved against SpliceAI in in-silico mutagenesis concordance (OpenSpliceAI provides an efficient modular implementation of SpliceAI enabling easy retraining across nonhuman species)) while delivering kinetic resolution that gels cannot: transient isoform shifts within the first two hours of dopamine exposure, at a well count (~243 factors × multiple conditions) that gel densitometry cannot practicably cover.
On the prediction side the strongest competitor is DCEN, the regulator-conditioned energy model reporting Spearman 0.623 and Pearson 0.651 on its benchmark against 0.579 and 0.574 for the SpliceAI-based baseline (RNA Alternative Splicing Prediction with Discrete Compositional Energy Network) — it is the incumbent method that, like ours, conditions on more than sequence, and the file reports no method on our exact mouse data. We run it on our BV-2 RNA-seq with its regulatory-factor vector set to measured splicing-factor expression, and our signaling-augmented model under the identical training data, split, and tuning budget; since CellSpliceNet's ablations show each non-sequence modality is worth ≥ 0.06 Spearman (CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons), the fair bar we set for ourselves is a gain of comparable or larger magnitude from the signaling modality, specifically on held-out stimuli where DCEN and sequence-only models cannot distinguish conditions. The key ablation is our own model with the signaling conditioning vector zeroed: if it matches the full model, the new ingredient is worthless and the proposal's computational claim fails.
Why not just…?
Why not just do bulk RNA-seq and RT-qPCR on dopamine-treated microglia? This is the most serious objection, because bulk RNA-seq is in fact part of this proposal — we need it to map the endogenous splicing program. But it cannot do the two things the proposal turns on. First, orthogonality: splicing changes can be invisible to expression-level analysis (Alternative splicing across the C. elegans nervous system), and functional truncations of the dominant-negative NOS type escape constitutive-exon RT-qPCR entirely (Upstream and Competitive Regulation of Alternative Splicing). Second, bulk sequencing is a destructive snapshot: it cannot resolve whether the dopamine-induced isoform switch precedes, coincides with, or follows the NF-κB shutoff, which is exactly the causal ordering the reporter's kinetics are designed to establish. The honest position is that RNA-seq and PSI-luc answer different questions, and the proposal uses each for what it can do; the decisive experiment on causality is the splice-factor knockdown screen, not the measurement modality.
Why not just use the existing pFlare minigene with RT-PCR gels? Because its own deployment shows the cost: the CD44 V6 study required lysis, RNA extraction, RT-PCR, and gel densitometry for every condition, producing endpoint numbers only (SR proteins regulate V6 exon splicing of CD44 pre-mRNA). An arrayed screen over 243 splicing regulators with kinetic sampling is out of reach of that workflow, and endpoint measurement cannot distinguish a transient two-hour isoform flip from a stable one — a distinction that matters, since kinase-driven splicing switches are expected to be fast and reversible (Upstream and Competitive Regulation of Alternative Splicing). The 48-hour live tracking that the NF-κB–GFP screen achieved (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy) is the standard we import, and no gel-based method can meet it.
Why not just retrain a sequence-only model (OpenSpliceAI) and skip the fancy conditioning? OpenSpliceAI's own results argue for species retraining — gains up to 57% in Arabidopsis over human-trained SpliceAI (OpenSpliceAI provides an efficient modular implementation of SpliceAI enabling easy retraining across nonhuman species) — and we will run exactly that retrained model as a baseline, so the alternative gets its fair shot. But two results say it should fall short on our question. DCEN, which conditions on regulatory factors, already beats the best SpliceAI-based baseline (Spearman 0.623 versus 0.579) (RNA Alternative Splicing Prediction with Discrete Compositional Energy Network), and CellSpliceNet loses at least 0.06 Spearman whenever any non-sequence modality is removed (CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons). More fundamentally, a sequence-plus-expression model represents two conditions with identical splicing-factor transcripts but different kinase flux as identical inputs, while the Bcl-x and ERCC1 results show such flux alone flips isoforms (Upstream and Competitive Regulation of Alternative Splicing, Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching). Still, this is a genuine threat: if the retrained OpenSpliceAI predicts the held-out CGRP $\Delta\Psi$ map as accurately as our signaling-conditioned model, the modality adds nothing, and the ablation is designed to detect exactly that outcome rather than argue around it.
Why not just measure phosphoproteins by mass spectrometry to trace the dopamine pathway? The signalome study confronted this directly: physiological signaling fluxes between interacting cells are often too weak and localized for phosphoprotein detection, which is why it used a functional optical reporter instead and recovered regulators — including novel negative ones — that showed zero mRNA-level change (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy). A phosphoproteomic panel would tell us kinases are active; it would not tell us which isoforms the cell consequently builds, which is the question here. Targeted phospho-Westerns for SR proteins remain a useful secondary readout and we include them, but as confirmation, not as the primary instrument.
Proposed Experiments and Decisive Checks
The program runs in five steps, each with a pre-stated decisive outcome. First, construct and calibrate PSI-luc against the CD44 V6 cassette: it must reproduce the SRSF1/SRSF6/SRSF9 skipping increase of roughly 36–46% and the null effect of SRSF3/SRSF4 reported in SR proteins regulate V6 exon splicing of CD44 pre-mRNA; failure to do so invalidates the instrument. Second, record live $\Psi_{\text{luc}}$ kinetics for iNOS, MyD88, Traf6, and Tlr4 constructs in BV-2 under LPS, LPS plus 2 μM and 10 μM dopamine, and D1-like/D2-like antagonist pre-blockade, for 48 hours — the same stimulation scheme as Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway — with bead phagocytosis and amoeboid-fraction assays run in parallel, so any isoform switch can be ordered against the known functional landmarks (phagocytosis restored from 9.6% toward 45.7%, activated cells reduced from 5.8% to 0.64%). Third, RNA-seq time courses across all conditions, junction-quantified in the MAJIQ style, to map the endogenous dopamine-responsive splicing program and to validate each reporter construct against its native locus — required because reporter cassettes have been shown to diverge from endogenous regulation when compensatory networks are in play (A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population). The primary end-point here is the count of significant $\Delta\Psi$ events at unchanged total expression, benchmarked against the atlas finding that roughly a quarter of expressed neuronal genes are splicing-regulated (Alternative splicing across the C. elegans nervous system). Fourth, the arrayed splicing-regulator screen with dual $\Psi_{\text{luc}}$ and NF-κB luciferase readouts, following the 706/708-gene arrayed CRISPR design (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy); the decisive comparison is between knockdowns that block dopamine's isoform switch but spare its NF-κB suppression (splicing is then dispensable for the brake) versus knockdowns that block both (splicing is on the causal path). Fifth, causal validation of the top hits by targeted splice-site modulation with catalytically dead Cas13d, the perturbation tool validated for engineered splicing outcomes in Generative modeling for RNA splicing prediction and design: forcing the dopamine-associated isoform without dopamine should reproduce part of the brake (reduced NF-κB reporter output, restored phagocytosis), and blocking it should attenuate dopamine's effect. Alongside, the signaling-conditioned model is trained on the dopamine conditions and evaluated on held-out doses and on the CGRP/RAMP1 axis (RAMP1 signaling attenuates acute lung injury by inhibiting cytokine production and neutrophil recruitment), against the retrained OpenSpliceAI sequence baseline, DCEN, and the zeroed-signaling ablation, on Spearman correlation of $\Psi$ and direction accuracy of $\Delta\Psi$.
The falsifiable core, stated plainly: if no dopamine-induced isoform switches occur at unchanged total mRNA, or if the screen finds no splicing regulator required for the brake, or if the signaling-conditioned model fails to beat its ablation on held-out stimuli, then the dopamine immune brake is not splicing-mediated in this system, and the proposal's biological claim is wrong even if the reporter instrument works. Either outcome is informative: the instrument and the signaling-state model remain reusable, and the negative result would itself rule out a mechanism that the total-mRNA literature cannot see.
References
- Role of dopamine in regulating microglia inflammatory responses through TLR4-NFb pathway
- Alternative splicing across the C. elegans nervous system
- Regulation of multimers via truncated isoforms: a novel mechanism to control nitric-oxide signaling
- Homoharringtonine regulates the alternative splicing of Bcl-x and caspase 9 through a protein phosphatase 1-dependent mechanism
- Genome-wide CRISPR-Cas9 screening identifies CLK1 inhibition as a strategy to restore PARP inhibitor sensitivity via ERCC1 isoform switching
- U4 at the 3′ UTR of PB1 Segment of H5N1 Influenza Virus Promotes RNA Polymerase Activity and Contributes to Viral Pathogenicity
- SR proteins regulate V6exon splicing of CD44 pre-mRNA
- TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation
- Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy
- A Functional Variant at miR-520a Binding Site in PIK3CA Alters Susceptibility to Colorectal Cancer in a Chinese Han Population
- CellSpliceNet: Explainable Multimodal Transformers for Splicing in C. elegans Neurons
- RNA alternative splicing prediction with discrete compositional energy network
- OpenSpliceAI provides an efficient modular implementation of SpliceAI enabling easy retraining across nonhuman species
- TRPV1+ neurons alter Staphylococcus aureus skin infection outcomes by affecting macrophage polarization and neutrophil recruitment
- RAMP1 signaling attenuates acute lung injury by inhibiting cytokine production and neutrophil recruitment
- Generative modeling for RNA splicing predictions and design