In particle physics, the transition from free quarks and gluons to confined hadrons cannot be derived analytically, and the intermediate mapping is unobservable — experiments only see the final, unordered cloud of hadrons. A recent line of work solved an analogous learning problem: train a generator whose local outputs are never labeled, by checking them with a global discriminator that compares whole particle sets, permutation by permutation, against reality. Breast cancer metastasis has exactly the same shape of problem. We can sequence primary tumors and we can sequence established metastases, but the regulatory rewiring that carried one into the other — driven largely by RNA-binding proteins (RBPs), the proteins that bind pre-messenger and messenger RNAs and control splicing, transcript stability, and translation — is never observed in transit in a patient. This proposal imports the physics community's solution: a two-level generative adversarial network — a generator trained against a critic — in which a latent vector, interpreted as RBP activity, is inferred purely from unpaired snapshots of primary and metastatic transcriptomes.
The core claim is that the non-perturbative QCD transition supplies a method, not a metaphor. The hadronization model of Fitting a deep generative hadronization model learns a local parton-to-hadron map from unlabeled final-state sets using a permutation-invariant Deep Sets discriminator (a network that consumes a set of particles in any order and produces one judgment), and its companion Integrating particle flavor into deep learning models for hadronization makes discrete categorical choices (particle flavor) differentiable alongside continuous kinematics via a Gumbel-Softmax relaxation (a trick that replaces a hard pick among categories with a soft, differentiable one). We port this machinery to the metastatic transition: the generator reads a primary tumor's transcriptome, emits a sparse latent vector of RBP activities, and from those activities generates the metastatic transcriptome — discrete isoform choices and continuous exon-inclusion levels together. Nothing about the intermediate RBP state is supervised; it is identified only through its effect on the observed sets.
Core Claim
The latent RBP rewiring that maps a primary breast tumor to a metastatic colony can be recovered from unpaired expression snapshots by a two-level adversarial architecture, and the recovered latent states should be measurably meaningful: they should separate patients who relapsed in brain from those who relapsed in lung, and an in-silico perturbation of the latent RBM47 activity should reproduce the documented downstream consequences of RBM47 loss. This is a claim about an inference method, evaluated against existing splicing and binding models on their own benchmarks.
Three strands of prior work constrain and motivate the design. The hadronization models supply the architecture and prove the concept on synthetic data: the discriminator in Fitting a deep generative hadronization model reached an AUC of 0.77 when distinguishing physics models, where standard summary observables sat near 0.5 — direct evidence that set-level adversarial critics extract transition information that conventional statistics miss. The biology supplies the latent variables worth inferring: Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer shows RBM47 loss destabilizes the Wnt antagonist DKK1 and tracks with brain and lung relapse in a 748-sample TCGA cohort and a 368-case untreated cohort; Arid4a Suppresses Breast Tumor Metastasis by Enhancing MTSS1 Expression via mRNA Stability shows Arid4a extends MTSS1 transcript half-life by binding stem-loop RNA structures, not linear motifs. The splicing models supply the differentiable readout: the Discrete Compositional Energy Network (DCEN) of RNA Alternative Splicing Prediction with Discrete Compositional Energy Network treats a mature transcript as a sum of junction energies, with isoform probabilities from a softmax over transcript energies — but it conditions on a fixed, observed 3,971-dimensional vector of regulatory factors. Our proposal makes that regulatory vector latent, inferred, and perturbable.
Research Program
The model has three parts. First, an encoder maps a primary tumor's expression profile to a sparse latent vector $\mathbf{r}$ of dimension equal to a curated set of metastasis-relevant RBPs (RBM47, Arid4a, ESRP1, G3BP1, hnRNPK, and on the order of a hundred others cataloged in EuRBPDB: a comprehensive resource for annotation, functional and oncological investigation of eukaryotic RNA binding proteins (RBPs)). Sparsity and partial supervision — where measured RBP expression exists in the same sample, the latent coordinate is softly anchored to it — are what make the latent space interpretable rather than arbitrary; without these constraints the adversarial setup would happily find unidentifiable factors. Second, a generative head turns $\mathbf{r}$ into an energy landscape over splice junctions, in the DCEN style, but with the energies conditioned on the latent RBP state and on structural features rather than sequence alone (this is mandatory: Arid4a binds RNA hairpins Arid4a Suppresses Breast Tumor Metastasis by Enhancing MTSS1 Expression via mRNA Stability, and PaRPI predicts RNA-Protein interactions from cross-protocol and cross-batch RNA-binding protein datasets shows icSHAPE-derived structure substantially improves binding prediction over sequence-only models). The probability of an isoform $t$ is
$$\Pr(t \mid \mathbf{r}) = \frac{\exp\!\big(-\sum_{j \in t} E_j(\text{seq}, \text{struct}, \mathbf{r})\big)}{\sum_{t'} \exp\!\big(-\sum_{j \in t'} E_j(\text{seq}, \text{struct}, \mathbf{r})\big)},$$
where $E_j$ is the energy of junction $j$ and the sum runs over the junctions composing a transcript. Continuous quantities (percent-spliced-in values, transcript abundances) are regressed from the same landscape; discrete isoform identity is sampled through a Gumbel-Softmax, exactly as Integrating particle flavor into deep learning models for hadronization did for particle flavor, so that gradients flow through the discrete choice. Third, the discriminator is a Deep Sets network over the unordered set of transcripts in a sample, trained adversarially to distinguish generated metastatic profiles from real ones — the same global-local split as Fitting a deep generative hadronization model, because gene sets, like hadron clouds, have no canonical order.
A secondary analysis addresses a known gap in the binding models. Towards in silico CLIP-seq: predicting protein-RNA interaction via sequence-to-signal learning predicts continuous crosslink profiles (Pearson correlations of 0.763 for TIA1, 0.830 for QKI, 0.872 for U2AF2) but explicitly cannot compare absolute affinity between RBPs competing for the same site. Because our energies are shared across all RBPs through $\mathbf{r}$, the softmax above acts as a crude partition function — the physicist's bookkeeping device for competing states — giving a first-pass model of multi-RBP competition for shared 3' UTRs, which is precisely what RBM47-versus-oncogenic-pathway competition in New Insights into the Interplay between Non-Coding RNAs and RNA-Binding Protein HnRNPK in Regulating Cellular Functions and circRNA sequestration in Transcriptome-wide profiles of circular RNA and RNA-binding protein interactions reveal effects on circular RNA biogenesis and cancer pathway expression demand but no current predictor provides.
One contradiction in the source material must be built in, not ignored: RBM47 is a metastasis suppressor in breast cancer Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer yet associates with worse prognosis and immunosuppressive macrophage infiltration in glioma RBM47 is a novel immunotherapeutic target and prognostic biomarker in gliomas. The model therefore conditions the latent-to-energy map on lineage context features, and an explicit evaluation will test whether the same latent RBM47 coordinate produces opposite transcriptomic effects under breast versus glioma-like context — a falsifiable check that the latent space respects known context dependence rather than collapsing it.
Baselines
The naive reference point is a class-conditional copy rule: predict each metastatic sample as the average profile of its metastatic site class, with no transition model at all. This is what "do nothing" produces, and any generative model must beat it on held-out metastatic profiles simply to justify existing; it also serves as the null for the organotropism discrimination test, where it performs at chance. The strongest existing competitor is DCEN RNA Alternative Splicing Prediction with Discrete Compositional Energy Network, which reports a Spearman correlation of 0.623 and a Pearson correlation of 0.651 on the CAPD percent-spliced-in benchmark, against 0.579 and 0.574 for the best SpliceAI-based baseline. It is the strongest competitor because it is the only published model that jointly predicts quantitative splicing from regulatory context, which is the same task our generator's readout must perform; we will run it on identical data splits, with the same input features and the same tuning budget, so that the only difference is whether the regulatory vector is observed and fixed (DCEN) or latent and inferred (ours). Beating it means improving Spearman correlation on held-out metastatic psi prediction by a margin whose confidence interval excludes zero, and, more importantly, doing so on samples where the regulatory context is not observed — the regime DCEN cannot enter by construction. A second comparator for the binding-feature side is RBPNet Towards in silico CLIP-seq: predicting protein-RNA interaction via sequence-to-signal learning, whose reported per-RBP Pearson correlations (up to 0.872 for U2AF2) set the bar for how well continuous binding signals can be reproduced; we use its profiles as input features, so we do not compete with it directly but must not degrade them. Finally, the key ablation isolates the proposal's central ingredient: the same generator and energy head, but with the permutation-invariant set discriminator replaced by a fixed-order vector discriminator on sorted gene identifiers, and separately with the latent RBP bottleneck removed (a direct transcriptome-to-transcriptome map). If the ablated variants match the full model, the physics-imported machinery is doing no work and the idea fails.
Why not just…?
Why not just run standard differential-expression and differential-splicing tools (DESeq2, DEXSeq, rMATS) on primary-versus-metastasis cohorts? These tools answer a different question: they find what differs between two matched groups, as a static list. They cannot produce a generative transition model, cannot be perturbed in silico, and need paired or matched phenotypic cohorts to be meaningful. The file's evidence that this matters is direct: in the hadronization setting, standard summary observables distinguished physics models at AUC near 0.5 while the set-level discriminator reached 0.77 Fitting a deep generative hadronization model, and in splicing, local classification-style architectures underperformed compositional ones (Spearman 0.579 versus 0.623) RNA Alternative Splicing Prediction with Discrete Compositional Energy Network. We will nonetheless run a differential-expression analysis as a feature baseline; if its ranked gene list predicts organotropism as well as our latent states, the latent inference adds nothing and we will say so.
Why not just fine-tune a pre-trained protein or RNA language model (ESM-2, mRNABERT) on clinical outcomes? This is a real threat and we treat it as a genuine second baseline rather than dismissing it. The file gives two concrete reasons it falls short: ESM-2 embeddings alone dropped to an accuracy of 0.67 on RBPs with less than 30% sequence similarity to training data PaRPI predicts RNA-Protein interactions from cross-protocol and cross-batch RNA-binding protein datasets, and flat-sequence language models are blind to the secondary-structure recognition that metastasis-relevant RBPs like Arid4a actually use Arid4a Suppresses Breast Tumor Metastasis by Enhancing MTSS1 Expression via mRNA Stability. More fundamentally, fine-tuning a sequence model predicts outcomes from sequences; it does not infer the unobserved intermediate regulatory state, which is the entire target here. The decisive experiment is the perturbation test below: a fine-tuned outcome predictor has no latent RBM47 dial to turn.
Why not just scan for known binding motifs with position weight matrices? Motif scanning is cheap and interpretable, but the file documents two failures that matter for metastasis: binary peak or motif classifiers miss continuous crosslink distributions and experimental noise that dedicated architectures handle Towards in silico CLIP-seq: predicting protein-RNA interaction via sequence-to-signal learning, and PWMs are structurally blind to the stem-loop conformations that Arid4a binds Arid4a Suppresses Breast Tumor Metastasis by Enhancing MTSS1 Expression via mRNA Stability. We will include a PWM-scoring feature set as an ablation of the structural inputs; if it matches icSHAPE-informed features on our benchmarks, the structural machinery is unnecessary for this task, which would itself be a useful negative result.
A fourth objection deserves a sentence: fitting simple parametric decay curves (single-exponential transcript half-lives) instead of a generative network. The file's own evidence argues against expressivity — transcript stability is dynamically co-regulated by multi-RBP and non-coding-RNA assemblies New Insights into the Interplay between Non-Coding RNAs and RNA-Binding Protein HnRNPK in Regulating Cellular Functions — but for the specific perturbation validation we propose, mono-exponential fits are a legitimate cheap baseline and will be reported alongside.
Proposed Experiments and Decisive Checks
The program has four stages, ordered so that failure is informative at each step. Stage zero is a method-development sandbox: reimplement the two-level GAN on the Herwig 7.2.1 synthetic electron-positron collision data at 91.2 GeV used by both hadronization papers, reproducing the discriminator AUC of 0.77 and approximately 1% precision on hadron flavor frequencies before any biological data is touched. This is cheap, validates the codebase against published numbers, and establishes that our port of the architecture has not silently broken the permutation-invariant training. Archived ALEPH data at the same energy First measurement of anti-kT jet spectra and jet substructure using the archived ALEPH e+e- data at 91.2 GeV is available as a real-data stress test of the trained critic; no reported result found on Lacuna for ALEPH jet-spectrum quantities, so this stage is qualitative.
Stage one trains on unpaired snapshots: primary breast tumors from the TCGA cohort of 748 samples and the 368-case untreated clinical cohort analyzed in Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer, with ENCODE eCLIP profiles (103 RBPs in HepG2, and the 261-dataset, 172-RBP, six-cell-line corpus of PaRPI predicts RNA-Protein interactions from cross-protocol and cross-batch RNA-binding protein datasets) supplying binding features via RBPNet-style continuous profiles, and CAPD (percent-spliced-in values with 3,971-dimensional regulatory vectors across 16 tissues) supplying the splicing supervision for the energy head. Binding features incorporate RNA secondary structure from icSHAPE and RNAplfold, per the PaRPI finding that structure helps.
Stage two is the decisive evaluation, with three checks. First, held-out quantitative splicing: predict percent-spliced-in values in metastatic samples and compare against DCEN's reported Spearman 0.623 / Pearson 0.651, run by us on identical splits and tuning budget. Second, organotropism: train a simple classifier on the inferred latent RBP states to separate brain-relapse from lung-relapse patients, and compare its AUC against the same classifier run on standard differential-expression features; no reported result found on Lacuna for organotropism classification from latent regulatory states, so we establish this number ourselves, with the HadML analog (0.77 versus 0.5) as the effect size worth shooting for. Third — the single most decisive check — the in-silico perturbation: take primary-like tumors whose inferred latent RBM47 is high, dial it to the metastasis-associated low value, and ask whether the generated metastatic profile recapitulates the documented signature of RBM47 re-introduction, namely the abundance shifts in the 102 upregulated and 92 downregulated mRNAs and the destabilization of DKK1 reported in Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer, measured as rank-overlap significance against the published gene lists. A model that recovers a specific, previously published perturbation signature it was never trained on has earned its latent space; one that does not has merely built a smooth interpolator, and we will report which.
Stage three tests generalization honestly: train on one clinical cohort and test on the other, and evaluate behavior on RBP contexts absent from training, using the best cross-context number in the file — PaRPI's accuracy of 0.67 on unseen proteins below 30% similarity — as a sanity bar for degraded-regime performance rather than a matched task baseline. The final analysis is the context-contradiction test: whether the same latent RBM47 coordinate yields breast-suppressive versus glioma-like effects under different lineage conditioning, confronting the documented duality between Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer and RBM47 is a novel immunotherapeutic target and prognostic biomarker in gliomas. If the model passes the perturbation and organotropism checks but fails this one, the honest conclusion is that the latent space captures transition dynamics but not lineage-dependent sign switching — still useful, but a weaker claim than the one staked here.
References
- Fitting a deep generative hadronization model
- Integrating particle flavor into deep learning models for hadronization
- Loss of the multifunctional RNA-binding protein RBM47 as a source of selectable metastatic traits in breast cancer
- Arid4a Suppresses Breast Tumor Metastasis by Enhancing MTSS1 Expression via <scp>mRNA</scp> Stability
- RNA alternative splicing prediction with discrete compositional energy network
- EuRBPDB: a comprehensive resource for annotation, functional and oncological investigation of eukaryotic RNA binding proteins (RBPs)
- PaRPI predicts RNA-Protein interactions from cross-protocol and cross-batch RNA-binding protein datasets
- Towards in silico CLIP-seq: predicting protein-RNA interaction via sequence-to-signal learning
- New Insights into the Interplay between Non-Coding RNAs and RNA-Binding Protein HnRNPK in Regulating Cellular Functions
- Transcriptome-wide profiles of circular RNA and RNA-binding protein interactions reveal effects on circular RNA biogenesis and cancer pathway expression
- RBM47 is a novel immunotherapeutic target and prognostic biomarker in gliomas
- First measurement of anti-kT jet spectra and jet substructure using the archived ALEPH e+e- data at 91.2 GeV