Genetic screens hand us validated RNA drug targets at a rapid pace — an antisense knockdown of the long noncoding RNA lnc-POTEM-4:14, for example, dismantles a pro-tumor transcriptional program in liver cancer by removing the scaffold that recruits the transcription factor FOXK1 (The LncRNA lnc-POTEM-4:14 promotes HCC progression by interacting with FOXK1). But turning such a target into a pill requires a small molecule, and small molecules do not read sequence: they grip three-dimensional pockets shaped by how the RNA folds. This is the sequence-structure gap in its most consequential form. Current predictors sit on the wrong side of it in one of two ways. Those that use real 3D structures — Augmented base pairing networks encode RNA-small molecule binding preferences and MVRBind: multi-view learning for RNA-small molecule binding site prediction — need solved crystal structures, which almost no disease-relevant RNA has. Those that work from sequence — DeepRSMA: a cross-fusion-based deep learning method for RNA–small molecule binding affinity prediction — collapse the RNA to a single predicted 2D contact map, dropping exactly the non-canonical base-pair geometry that the ablation studies in Augmented base pairing networks encode RNA-small molecule binding preferences proved is mandatory: models stripped of non-canonical interactions performed no better than random at ligand prediction.
The proposal here is to predict small-molecule binding to unsolved RNA structures by marginalizing over a sampled conformational ensemble rather than committing to one fold. Given only a sequence, we sample a handful of conformations with a generative structure model, annotate each as a graph whose edges carry canonical and non-canonical base-pair geometry, weight the states thermodynamically, and fuse them with a molecular graph to output binding affinity and a ligand fingerprint. The result is a fast screening model that inherits the physical realism of structure-based methods without requiring any experimental structure — which is precisely what the pharmacological translation gap demands (Translating Genetic Target Validation into Small Molecule Inhibitors).
Core Claim
The central claim is testable and narrow: RNA–ligand affinity prediction from sequence alone improves substantially when the RNA representation is (i) an ensemble of sampled conformations, not a single minimum-free-energy fold, and (ii) annotated with non-canonical base-pairing edges, not just canonical pairs or plain contact maps. Both ingredients have independent precedent — ensemble weighting lifted protein–RNA binding prediction in Integrating thermodynamic and sequence contexts improves protein-RNA binding prediction, where moving from a single average structure to five Boltzmann-weighted structural experts raised the Pearson correlation on RNAcompete from 0.6058 (the DLPRB-CNN baseline) to 0.6710, and non-canonical edges were shown indispensable in Augmented base pairing networks encode RNA-small molecule binding preferences — but no existing model combines them, and the two most recent sequence-based systems each violate one of them. DeepRSMA uses one predicted contact map per RNA with no edge geometry; RNAmigos needs the holo crystal structure as input and cannot be applied to an unsolved target at all.
Research Program
Three components get built. First, an ensemble annotation pipeline: for an input RNA sequence, a generative conformational sampler — built on the diffusion approach of Towards Unraveling Biomolecular Conformational Landscapes with a Generative Foundation Model (OpenComplex2), which samples equilibrium ensembles orders of magnitude faster than molecular dynamics and reproduces free-energy profiles of millisecond simulations to a mean absolute error of 0.92 kcal/mol — produces $K$ candidate folds. Each fold is converted into an augmented base-pairing graph in the style of RNAmigos, with nodes for nucleotides and typed edges distinguishing Watson-Crick, wobble, and the non-canonical cis/trans Hoogsteen and sugar-edge interactions that define pocket walls. Second, an ensemble-weighted affinity model. With $s$ the RNA sequence, $G_k(s)$ the graph of the $k$-th sampled conformation, $H(m)$ the molecular graph of compound $m$, and $w_k$ a Boltzmann weight derived from the sampler's state energies, the prediction is
$$\widehat{\mathrm{p}K_d}(s,m) \;=\; \sum_{k=1}^{K} w_k \, f\big(G_k(s),\, H(m)\big),$$
where $f$ is a cross-attention network that interleaves nucleotide tokens with atomic tokens, reusing DeepRSMA's cross-fusion design as the student backbone. We will compare fixed Boltzmann weights against learned state-attention, following the probability-weighted expert design that worked for ThermoNet. Third, distillation for chemotype coverage: because the training data (1,439 measured complexes in R-SIM) is small and skewed toward aminoglycosides, we use the generative model to dock underrepresented drug-like chemotypes across sampled RNA ensembles and use agreement-ranked poses as augmentation, then verify on held-out experimental data. The end product is a screening engine fast enough for library-scale use, where direct diffusion docking is not, and applicable to apo targets with no solved structure, where crystal-structure methods are not.
Baselines
The naive reference point is a sequence-only predictor: DeepRSMA's molecular encoders fed with frozen RNA language-model embeddings and no structural input of any kind. This is the "do nothing structural" floor, and RNAmigos's own ablation — sequence-only models performed no better than random at ligand prediction — predicts it should sit near chance on ligand identity tasks while remaining non-trivial on affinity, where short motifs carry some signal. The strongest existing competitors are DeepRSMA for affinity, which reports a Pearson correlation of 0.784 on R-SIM five-fold cross-validation but drops to 0.582 on blind unseen-RNA splits and 0.490 (RMSE 0.920) on the independent HIV-1 TAR set; RNAmigos for ligand fingerprint prediction, which reports a normalized rank of 0.68 and a fingerprint reconstruction mean-squared error of 0.150 on its 773-site, 270-ligand benchmark, placing the true ligand in the 71st–73rd percentile of decoy libraries; and MVRBind for binding-site localization, which reports an AUC of 0.745 on the holo Test18 set, 0.756 on the apo test, and 0.660 on the conformational test. Each will be retrained or evaluated under identical splits, identical molecule encoders, and an equal hyperparameter search budget, so any gap is attributable to the RNA representation alone. The ablation baselines isolate the two claimed ingredients: our own model restricted to a single minimum-free-energy conformation (removing the ensemble), and our model restricted to canonical edge types (removing the non-canonical geometry). Beating DeepRSMA means exceeding its blind-split correlations by at least 0.05 Pearson with overlapping confidence intervals ruled out by paired bootstrap; beating RNAmigos means improving its rank and error on the chemotype-split subset where it is weakest, not just on the aminoglycoside-dominated whole.
Why not just…?
Why not just scale up a sequence language model? This is the cheapest rival and the most serious one. The file's evidence is unambiguous against it for this task: RNA-FM, a purely sequence-based foundation model, solved only 2–3% of EternaV2 design puzzles where the structure-aligned OmniGenome solved 74% (Bridging Sequence-Structure Alignment in RNA Foundation Models), and RNAmigos found sequence-only ligand prediction indistinguishable from random. But the honest answer is that no one has run a very large frozen language model against DeepRSMA on blind affinity splits, so we run it ourselves as the sequence-only baseline above. If it matches the ensemble model, the proposal's premise fails and we will report that.
Why not just fold with RNAfold and match dot-bracket motifs, as Inforna does? Thermodynamic folding is the standard cheap route, and RNAfold reaches a median F1 of 0.492 on bpRNA-1m versus 0.552 for ERNIE-RNA's zero-shot attention maps (ERNIE-RNA: an RNA language model with structure-enhanced representations) — respectable for secondary structure, but secondary structure is the problem. RNAmigos outperformed Inforna across ligand classes, and its ablation reducing representations to canonical pairs fell to random performance; MVRBind found that removing the tertiary view caused its single largest drop. Our canonical-edges-only ablation is exactly this alternative embedded in our pipeline, so the comparison is direct rather than rhetorical.
Why not just dock every library compound with the diffusion model directly? OpenComplex2 can already predict small-molecule poses, handle multi-site targets, and capture cryptic pockets and induced fit (Towards Unraveling Biomolecular Conformational Landscapes with a Generative Foundation Model), and nothing in the file validates it for RNA-targeted inhibitor screening specifically. Its per-system cost in hours rules out million-compound screens, which is why we use it as an ensemble and pose generator that teaches a fast student, not as the screening engine. The risk is real: if its inference becomes cheap enough, the student is unnecessary. Our decisive check is a recall test — screen the library with the student, then run direct generative docking on the top and bottom deciles, and verify the student's rankings survive; if they do not, the distillation is failing and we will say so.
Why not just use antisense oligonucleotides and skip small molecules entirely? ASOs already knock down lnc-POTEM-4:14 in mice (The LncRNA lnc-POTEM-4:14 promotes HCC progression by interacting with FOXK1), so for targets where oligonucleotide delivery and safety are acceptable, this work is genuinely redundant — that concession should be made plainly. The counterargument is the scaffold problem laid out in Translating Genetic Target Validation into Small Molecule Inhibitors: genetic silencing erases the whole molecule and every complex it hosts, while a small molecule binding one folded pocket can disrupt a single oncogenic interface, such as the FOXK1-recruitment site, without abolishing the RNA's other functions; and small-molecule tissue pharmacology differs from oligonucleotide pharmacology in ways that matter clinically. The original authors themselves flag small-molecule screening for this axis as an unmet, earliest-stage need.
Why not just use the single best-predicted structure? ThermoNet showed single-average-structure models lose to Boltzmann-weighted ensembles (0.6058 to 0.6710 Pearson on RNAcompete; 0.836 to 0.864 AUROC on CLIP-seq versus DeepBind), and MVRBind's own numbers degrade from 0.756 AUC on apo structures to 0.660 across conformational ensembles. Our single-conformation ablation reproduces this design choice inside one architecture, so the answer is an experiment, not an assertion.
Proposed Experiments and Decisive Checks
The experimental program runs on four benchmarks with reported numbers. On R-SIM (1,439 complexes with measured $\mathrm{p}K_d$), we evaluate under DeepRSMA's five-fold cross-validation and, more importantly, its blind unseen-RNA and blind unseen-molecule splits, targeting its reported blind-RNA correlation of 0.582 and HIV-1 TAR correlation of 0.490 as the numbers to exceed. On the RNAmigos benchmark of 773 binding sites and 270 ligands, we compare normalized rank and fingerprint error against its reported 0.68 and 0.150, with a pre-registered chemotype split that holds out non-aminoglycoside ligands, since RNAmigos itself reports degraded fidelity on rare drug-like chemotypes — the chemotype-split result is the single most important number in the proposal, because it measures whether ensemble-plus-non-canonical representations generalize beyond the aminoglycoside bias rather than merely re-fitting it. On the MVRBind test suites, we evaluate site prediction on the apo (AUC 0.756) and conformational (AUC 0.660) sets, with success defined as cutting the holo-to-conformational gap (0.745 to 0.660) by at least half. The ablation matrix — single versus ensemble conformations, canonical versus non-canonical edges, fixed versus learned weights, with versus without distilled poses — then attributes any gains. Finally, a case study applies the trained screener to the FOXK1-interacting region of lnc-POTEM-4:14, whose structure is unsolved and for which no reported result found on Lacuna exists for any inhibitor effort; top-ranked compounds would be rescored with the free-energy-perturbation methodology validated for cryptic pockets in How does a small molecule bind at a cryptic binding site?, and the top hits passed to experimental collaborators. The falsifiable predictions are stated in advance: if the canonical-only ablation matches the full model on the chemotype split, RNAmigos's ablation finding does not transfer to affinity prediction and the non-canonical ingredient is dead weight; if the sequence-only baseline matches the ensemble model on blind splits, the structural claim collapses; and if the student's rankings do not survive direct generative re-docking of sampled extremes, the distillation has failed regardless of benchmark scores.
References
- The LncRNA lnc-POTEM-4:14 promotes HCC progression by interacting with FOXK1
- Augmented base pairing networks encode RNA-small molecule binding preferences
- MVRBind: multi-view learning for RNA-small molecule binding site prediction
- DeepRSMA: a cross-fusion-based deep learning method for RNA–small molecule binding affinity prediction
- Integrating thermodynamic and sequence contexts improves protein-RNA binding prediction
- Towards Unraveling Biomolecular Conformational Landscapes with a Generative Foundation Model
- Bridging Sequence-Structure Alignment in RNA Foundation Models
- ERNIE-RNA: an RNA language model with structure-enhanced representations
- How does a small molecule bind at a cryptic binding site?