acceptodds
Under review as a conference paper at ICLR 2027

RNAClip: Auxiliary Protein–Ligand Data Improves Pocket-Free RNA–Ligand Virtual Screening

Abstract

Small molecules targeting RNA offer therapeutic opportunities beyond conventional protein targets, yet current screening methods are constrained by the difficulty of locating binding pockets and the scarcity of validated RNA–ligand interactions. Here, we introduce , a contrastive retrieval framework that brings abundant protein–ligand data into RNA virtual screening. By aligning receptors and ligands in a shared embedding space, the framework enables sequence-based screening without a predefined RNA pocket or binding pose. Auxiliary protein examples contribute both paired supervision and contrastive negatives, allowing RNA screening to draw on data beyond the RNA domain. We also construct a new evaluation set from recently released RNA–ligand structures in the Protein Data Bank(PDB). On this strictly independent test set, outperforms all other evaluated models, achieving state-of-the-art(SOTA) screening performance. A series of ablation experiments further demonstrates that a small subset of auxiliary protein–ligand data can match the performance achieved with the full corpus, and that using auxiliary examples solely as contrastive negatives can preserve the gains from joint training.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA-ligand interaction prediction, virtual screening, contrastive learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bww9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RNACLIP, a framework for RNA-ligand virtual screening. In order to solve the problem of data scarcity, the authors used protein-ligand data to enrich the datasets. The model utilized shared intellifold embedders as pretrained encoders. The authors used ablation studies to prove that additional data helped performance. However, the baseline methods evaluated in experiments are limited, more confidence scores (Ipae, Ipsae et.al) from Alphafold3 / Boltz2 / Protenix2 should be included.

Critical strengths
  • RNACLIP outperforms other evaluated baseline methods by a large margine
  • This paper proves that using additional protein-ligand could further improve RNA-ligand screening performance
Critical weaknesses
  • Notations and equations are hard to follow, especially the definition of m_i
  • Popular interface screening scores from folding models like IPAE, IPSAE are missing from the comparison.
  • The model is evaluated on self-curated eval sets. It also makes sense to compare with other baseline methods on some common benchmark sets.
Questions for authors
  • It is quite confusing that pair-shuffled perform nearly as good as the reference. How is the model doing a great job without being confused by wrong ligand-protein pairs.
  • How are other methods’ training data different from your model? Any chance they are just undertrained?
  • Why is it still necessary to include protein-ligand datasets as additional training source since the pretrained encoder already sees them?
  • Does training on RNA-ligand also help improve protein-ligand screening?
Reviewer_v9v5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RNAClip formulates RNA–ligand virtual screening as retrieval using independently encoded receptor and ligand representations. Frozen IntelliFold trunks supply features to small trainable projection heads; RNA and protein share the receptor tower. A symmetric contrastive objective combines 2,366 RNA–ligand interfaces with auxiliary protein–ligand examples. Inference ranks cached ligand embeddings against an RNA sequence embedding without a supplied pocket or complex pose. The authors curate seven validation and ten test targets using temporal separation, structural redundancy filtering, and property-matched decoys. On the test panel, joint training improves mean per-target AP from 0.255 to 0.468, compared with 0.374 for DeepMIF. Corpus-size, negatives-only, partner-shuffling, and pretraining ablations investigate the source of this improvement.

Critical strengths

The retrieval setup addresses a practical constraint: RNA pockets are often unknown. Independently encoding RNA sequences and ligands avoids requiring a supplied pocket or a pose for every candidate and permits reuse of ligand embeddings. Only about 396,000 head parameters are trained (Appendix B.1).

The auxiliary-data ablations reveal an interesting result. Joint and negatives-only training achieve PT-AP values of 0.468 and 0.465 (Table 2), and even the smallest auxiliary corpus improves over RNA-only training. Intact auxiliary protein–ligand pairing losses are therefore not needed to retain most of the observed gain on this benchmark.

Critical weaknesses

The experiments do not establish what is specifically gained from protein-derived data or from conditioning on the RNA target. The model uses existing frozen representations, projection heads, and a standard symmetric contrastive loss, so the auxiliary-data result carries much of the contribution. Negatives-only and partner-shuffling controls are useful, but retain the protein-receptor population and auxiliary ligand distribution. A matched generic-ligand negative control is missing, as is a receptor-agnostic screening baseline. The results could reflect useful protein-derived information, generic chemical negatives, receptor-modality separation, or a chemical prior for RNA binders. Table 2 establishes that this auxiliary candidate distribution helps, but leaves unclear why protein–ligand data are needed and how much of the improvement depends on the individual RNA query.

The test panel is too small and restricted to establish robust screening generalization. It contains ten targets and eight native-ligand clusters, with one known ligand per target against 99 decoys for nine targets and 79 for the remaining target (Appendices A.2 and C.4). Temporal filtering, structural filtering, the joint similarity audit, and cluster reweighting are useful safeguards. However, taking the minimum of RNA and ligand similarity in the joint audit does not show that each input is individually novel. The decoys are also not measured nonbinders. No training-seed or auxiliary-subset variability is reported. Compared with DeepMIF, Hits@1 changes from 3/10 to 4/10, while both reach 5/10 at Hits@5 (Tables 1 and 5). The AP improvement is encouraging, but this panel provides limited evidence of a stable advantage across targets or realistic screening libraries.

Questions for authors

With RNA supervision, optimizer updates, and candidate-pool size matched, how does negatives-only training compare with generic chemical negatives? How well do a receptor-agnostic ligand scorer and RNAClip with shuffled RNA queries perform on the same panel?

How many independent training seeds and auxiliary-subset selections were evaluated? Could the authors report per-target ranks and variability for the main comparison and negatives-only ablation under the stated checkpoint-selection protocol?

What are the separate maximum RNA-sequence and native-ligand similarities to training data for each target? Are results available with multiple experimentally established actives and measured nonbinders, or on held-out ligand scaffolds?

Reviewer_AUtEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

RNAClip uses frozen IntelliFold encoders, trainable projection heads, and a symmetric contrastive loss for pocket-free RNA-ligand screening. Paper main contribution is to add protein-ligand examples during training. On 10 recent PDB targets, PT-AP increases from 0,255 for RNA-only training to 0.468, compared with 0.374 for DeepMIF. Auxiliary examples used only as negatives achieve nearly identical performance to joint training.

Critical strengths
  • The negatives-only intervention addresses whether auxiliary binding correspondences are necessary.

  • Evaluation curation includes temporal separation, structural filtering, an input-space overlap audit, as well as sensitivity to repeated ligand identities, which is useful

Critical weaknesses
  • Limited novelty and incomplete attribution. Sequence-only contrastive RNA screening already exists in SMARTBind, whose projectors align representations in a shared latent space, where RNAClip adds established structural encoders and auxiliary examples to the standard retrieval objective. Since negatives-only essentially matches joint training, the distinctive contribution depends on showing why protein-derived negatives are useful or not. There is no comparison with generic compound negatives, no separation of auxiliary receptor and ligand effects, no control for target-independent ligand preferences.

  • The evaluation is too small to support broad performance claims. There are 10 targets, 8 distinct native ligands, and 1 positive per target against generated decoys. Without ties, PT-AP here is simply MRR. RNAClip improves Hits@1 from 3/10 to 4/10 over DeepMIF, while both achieve 5/10 at Hits@5. No repeated training runs or confidence intervals establish the reliability of these differences. Likewise, close point estimates do not establish equivalence between auxiliary configurations.

  • Baseline differences limit interpretation. Training data, budgets, representations, and validation procedures vary across methods. The BioLLMNet reproduction also changes the published predictor substantially. These comparisons provide context but do not isolate the contribution of the auxiliary strategy from pre-training and implementation choices.

Questions for authors
  • Does the gain persist with a size- and property-matched unpaired compound library? What happens if auxiliary ligand negatives and protein receptor negatives are added separately?

  • How much performance remains with a ligand-only predictor or shuffled RNA queries? These controls would test whether ranking depends on the target or not.

  • Please report per-target ranks, repeated training and auxiliary-subset seeds, also paired uncertainty estimates. An independent panel with multiple measured actives and inactives would materially strengthen the conclusions.

  • Were RNA exposure and optimizer updates matched in the RNA-only comparison? Please specify the actual checkpoint-selection procedure and report separate receptor and ligand overlap with downstream and pre-training data, including auxiliary protein-associated ligands.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.