Cells in tissue talk to their neighbors by secreting diffusible molecules — growth factors, cytokines, enzyme-laden vesicles — that drift across the extracellular space and reprogram nearby recipient cells. Computational tools now infer these conversations at scale from single-cell RNA sequencing, tracing paths from a secreted ligand to a receptor to a transcription factor (the proteins that switch genes on and off) to downstream target genes. But these inferred paths rest entirely on mRNA readouts, which are imperfect proxies for what actually happens physically inside the recipient cell. Two measurements can settle it: chromatin immunoprecipitation (ChIP) — chemically "freezing" cells and fishing out the DNA fragments a transcription factor is physically bound to — and co-immunoprecipitation (Co-IP) — capturing a bait protein from gently lysed cells to see which partner proteins are stably attached to it. A recent review of over 140 cell-cell communication inference tools calls the absence of such physical validation a field-wide "ground truth crisis" (Cell-cell communication inference and analysis: biological mechanisms, computational approaches, and future opportunities).
The same missing layers haunt the in vitro-to-in vivo translation gap — the recurring failure of findings from cells grown on plastic to hold up in living organisms. The best machine-learning systems that transfer drug response from lab cell-line screens to patient tumors (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction; Cell-Level Virtual Screening) represent a cell only by its static gene-expression vector, ignoring both the paracrine signals a cell receives from its microenvironment and the physical chromatin and protein-complex states those signals induce. This proposal builds a single resource that attacks both problems at once: a paired atlas in which defined paracrine perturbations are applied to recipient cells and the response is measured simultaneously by ChIP-seq, Co-IP mass spectrometry, and RNA-seq. The atlas then serves two consumers — as physical ground truth to score mRNA-based communication inference, and as feature inputs to test whether chromatin-occupancy and protein-assembly data close the cell-line-to-patient transfer gap.
Core Claim
The core claim is twofold. First, a defined paracrine signal rewrites the recipient cell primarily at two physical layers — where transcription factors occupy DNA, and which multi-protein complexes assemble — and these changes are systematically mispredicted by mRNA-only inference, so a paired ChIP/Co-IP perturbation atlas provides the missing physical ground truth for cell-cell communication tools. Second, when these physically measured response layers, collected under paracrine conditioning rather than monoculture, are added as features to transfer-learning models, they improve prediction of in vivo (patient) drug response beyond what static expression features achieve. One prior study proves the atlas is feasible in a single niche — but only there: Endothelial Dickkopf-1 Promotes Smooth Muscle Cell-derived Foam Cell Formation via USP53-mediated Deubiquitination of SR-A During Atherosclerosis showed that shear-stressed endothelial cells secrete DKK1, which drives CREB promoter binding in smooth muscle cells (targeted ChIP) and assembly of a USP53–SR-A protein complex (targeted Co-IP), validated in endothelial-specific knockout and overexpression mice plus human plaque biopsies. That paper spans all four directions but stays with one ligand, one axis, and locus-level assays. Our proposal generalizes this into a multi-axis, genome-wide, framework-grade resource with computational consumers built in.
Research Program
The program has three stages. Stage one is a pilot on the DKK1→smooth-muscle axis, chosen because its answers are already known and can serve as positive controls: the transcriptomic data of Zhang et al. is deposited as GEO PRJNA851945 (RNA-seq of smooth muscle cells treated with recombinant DKK1 or DKK1 siRNA under shear), and the paper's own targeted CREB ChIP and USP53–SR-A Co-IP results define expected hits. We will repeat those perturbations under three arms — static conditioned medium, transwell non-contact co-culture, and a parallel-plate flow chamber that adds physiological shear — and measure, per arm and time point, ChIP-seq for a transcription-factor panel, Co-IP followed by mass spectrometry using the receptor and its adaptors as baits, and matched RNA-seq. Success in the pilot means the genome-wide assays recover the known CREB occupancy and USP53–SR-A complex and reveal how the flow arm differs from the static arms.
Stage two expands to two additional donor–recipient axes with quite different biology, so the atlas is not a vascular curiosity: the macrophage–cancer-cell GM-CSF/HB-EGF paracrine loop identified in transwell co-cultures and human liver-metastasis biopsies (Macrophages may promote cancer growth via a GM-CSF/HB-EGF paracrine loop that is enhanced by CXCL12), and mesenchymal-stromal-cell-driven macrophage reprogramming toward a pro-healing phenotype (Adipose-Derived Mesenchymal Stem Cells Accelerate Diabetic Foot Ulcer Healing by Promoting Macrophage M2 Polarization Through Downregulation of EREG and CSTA; Extracellular Vesicles Derived from Ovarian Cancer Cells Promote Tumor Progression through M2 Macrophage Polarization and Enhanced Angiogenesis). Both are cases where a downstream effect is known but the physical intermediates — which transcription factors actually bind which promoters, which complexes actually assemble in the recipient — are explicitly unmapped, making them ideal atlas targets. Across axes we compute the key quantity nobody has: edge-level concordance between inferred communication paths and physical measurements. For each ligand→transcription-factor→gene and receptor→adaptor edge predicted from the matched RNA-seq by a tool in the style of Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data (SigXTalk, a hypergraph model connecting receptors, intracellular signaling, transcription factors, and target genes), we check whether the transcription factor shows ChIP-seq occupancy at that gene's regulatory region and whether the adaptor pair co-immunoprecipitates. No reported result found on Lacuna for any cell-cell communication tool scored against ChIP or Co-IP ground truth, which is precisely the hole the atlas fills.
Stage three turns the atlas into features for translation-gap models. Following the architecture of DeepSADR, which learns bipartite interactions between drug chemical fragments and genomic pathways on DepMap cell-line screens (966 lines, 225,781 samples) and transfers them to 555 TCGA patient tumors with a set-transformer readout, we add a third feature family: recipient-cell differential ChIP occupancy and Co-IP complex membership measured under paracrine conditioning versus monoculture. The comparison target is also Cell-Level Virtual Screening (CellVS-Net), which showed that gene-connectivity rewiring, not raw expression, carries therapeutic signal. Our hypothesis is that paracrine-conditioned physical features capture the stromal reprogramming that DeepSADR and CellVS-Net both omit, and that this is where part of the cell-line-to-patient performance loss lives — a hypothesis directly motivated by the finding that gene-set collections built from cell lines behave discordantly from clinically curated ones (Importance of collection in gene set enrichment analysis of drug response in cancer cell lines).
Baselines
The do-nothing reference point for the retrieval side is random and molecular-fingerprint matching on the Drug-Disease Retrieval benchmark (DDR-Bench), where CellVS-Net reports a random baseline of 0.0179 Hits@1 and a molecular-fingerprint baseline of 0.0357 Hits@1 (Cell-Level Virtual Screening). We will run these unchanged; anything below the fingerprint number means the physical features added noise rather than signal.
The strongest existing competitor on the translation question is a tie between two systems that must both be beaten on their home turf. DeepSADR reports an AUC of 0.957 on clinical Sorafenib response in TCGA (against a 0.811 baseline), the best clinical drug-response transfer number in the file; CellVS-Net reports Hits@1 of 0.1250 on DDR-Bench and a 12–47% reduction in network-prediction MSE across perturbation types on the Drug-Target Retrieval benchmark (DTR-Bench). To make the comparison fair, we keep their architectures, data splits (DepMap pre-training, TCGA evaluation, OpenTargets ground truth), and hyperparameter-tuning budgets identical, and change only the input features; our model is their model plus atlas features. Any gain is then attributable to the features, not to extra tuning or compute. Beating the baseline means, concretely, exceeding 0.957 AUC on held-out TCGA drug-response prediction with consistent gains across drugs (not just Sorafenib), and exceeding 0.1250 Hits@1 on DDR-Bench.
For the ground-truth-scoring side, the strongest baseline is SigXTalk (Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data), which reports an 18% increase in AUROC and a 15% increase in AUPRC over 12 gene-regulatory-network methods on the SERGIO synthetic benchmark. SERGIO, however, is simulated data with synthetic ground truth; our evaluation replaces it with measured ChIP/Co-IP edges from the atlas, which is exactly the comparison that does not exist today.
The key ablation isolates the proposal's central ingredient: the same transfer model with atlas features measured from monocultures versus atlas features measured under paracrine conditioning. If paracrine conditioning is the active ingredient, the conditioned features should win; if monoculture ChIP/Co-IP features do just as well, then the value is generic epigenomic data, not paracrine context, and the claim must be weakened accordingly.
Why not just…?
Why not just run existing single-cell RNA-seq communication tools instead of building wet-lab ChIP/Co-IP panels? The review of 140+ such tools (Cell-cell communication inference and analysis: biological mechanisms, computational approaches, and future opportunities) documents the core failure: mRNA abundance is an imperfect proxy for protein abundance and says nothing about post-translational assembly, and the field's own authors flag the ground-truth crisis. SigXTalk (Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data) validated itself only against simulated SERGIO networks, never against physical occupancy or complex assembly. Our experiment two directly measures how often inferred edges survive physical validation — and honestly, if a large fraction do, the computational tools gain credibility at our proposal's expense; that outcome would itself be a publishable calibration.
Why not just align gene-expression distributions between cell lines and tumors (standard domain adaptation)? The file contains two pieces of evidence that global expression alignment is the wrong currency. Removing DeepSADR's modular interaction readout dropped its average AUC from 0.856 to 0.662 (DeepSADR), showing pooled expression loses the signal; and CellVS-Net's connectivity-rewiring features outperformed expression-level baselines on indication retrieval, 0.1250 versus 0.0357 Hits@1 (Cell-Level Virtual Screening). Still, batch-corrected expression alignment is a real threat because it is cheap; our monoculture-versus-paracrine ablation is designed to settle it, since both arms can include expression alignment and differ only in physical conditioning.
Why not just predict transcription-factor binding computationally instead of running ChIP? EpiBinder (EpiBinder: a multimodal deep learning model at base-resolution to analyze in vivo Transcription Factor-DNA Binding) shows that adding base-resolution DNA methylation and chromatin accessibility to sequence yields up to a 15-point auPRC gain over sequence-only models — meaning sequence alone is known to be insufficient, and structural dynamics such as DNA breathing further modulate binding (Contribution of DNA breathing to physical interactions with transcription factors). But the deeper problem is causal: no existing model predicts how an extracellular paracrine ligand shifts occupancy, because the inputs to such a shift are precisely the measurements we propose to make. EpiBinder needs the recipient cell's methylation and accessibility states as inputs, and those states are what paracrine conditioning changes. We will nonetheless include an EpiBinder-based in-silico arm as a cheap competitor: feed it the monoculture epigenome and test whether it predicts the paracrine-conditioned ChIP landscape. If it does, ChIP assays beyond the calibration set could be scaled back — a genuinely useful outcome.
Proposed Experiments and Decisive Checks
Experiment one (pilot, positive controls): repeat the recombinant-DKK1 and DKK1-siRNA shear perturbations of Zhang et al. across the three culture arms, with genome-wide ChIP-seq and Co-IP-MS. Decisive check: recovery of CREB occupancy at the known targets and the USP53–SR-A complex, plus a quantified flow-versus-static divergence — the first measurement of whether shear-dependent secretion, which that paper showed matters in one axis, systematically reshapes the physical response layers.
Experiment two (concordance): score mRNA-inferred communication edges from the matched RNA-seq (SigXTalk-style inference) against ChIP-validated promoter occupancy and Co-IP-validated complex assembly. Metric: edge-level precision against physical ground truth, reported per axis and per culture arm. No reported result found on Lacuna for any prior concordance number of this kind, so even a null result calibrates the field.
Experiment three (transfer): augment the DeepSADR-style bipartite model with atlas features and evaluate on TCGA drug response and DDR-Bench, against the reported 0.957 AUC and 0.1250 Hits@1. Experiment four (decisive ablation): monoculture-derived versus paracrine-conditioned physical features in the identical architecture. The proposal stands or falls here: if paracrine conditioning adds nothing over monoculture physical features, the honest conclusion is that epigenomic data helps transfer while paracrine context does not, and we would report that as the finding — with the concordance results from experiment two carrying the paper's value instead.
References
- Cell-cell communication inference and analysis: biological mechanisms, computational approaches, and future opportunities
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cell-Level Virtual Screening
- Endothelial Dickkopf-1 Promotes Smooth Muscle Cell-derived Foam Cell Formation via USP53-mediated Deubiquitination of SR-A During Atherosclerosis
- Macrophages may promote cancer growth via a GM-CSF/HB-EGF paracrine loop that is enhanced by CXCL12
- Adipose-Derived Mesenchymal Stem Cells Accelerate Diabetic Foot Ulcer Healing by Promoting Macrophage M2 Polarization Through Downregulation of EREG and CSTA
- Extracellular Vesicles Derived from Ovarian Cancer Cells Promote Tumor Progression through M2 Macrophage Polarization and Enhanced Angiogenesis
- Dissecting crosstalk induced by cell-cell communication using single-cell transcriptomic data
- Importance of collection in gene set enrichment analysis of drug response in cancer cell lines
- EpiBinder: a multimodal deep learning model at base-resolution to analyze in vivo Transcription Factor-DNA Binding
- Contribution of DNA breathing to physical interactions with transcription factors