Xueyi Yang, Amirhossein Taghavi, Yoshihiro Akahori, Martina Pedrini, Takahiro Ishii, Matthew D. Disney

ACS Chemical Biology

DOI: 10.1021/acschembio.6c00337

Abstract

Disease-associated RNAs are increasingly recognized as promising therapeutic targets for small-molecule intervention. While DNA-encoded libraries (DELs) have long been established for protein ligand discovery, recent studies have demonstrated their feasibility for identifying RNA-binding small molecules. To further advance RNA-targeted ligand discovery, a diverse, solid-phase DEL enriched in privileged RNA-binding scaffolds was constructed and applied to identify ligands of r(G₄C₂)^exp, a toxic RNA repeat expansion implicated in amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). DEL selection outcomes were analyzed through large-scale molecular docking integrated with physicochemical and structure–activity relationship (SAR) analyses. Correlations were observed between docking predictions and experimental enrichment trends, supporting lead identification. The lead compound was subsequently optimized based on rational design, resulting in analogues with enhanced binding affinity and bioactivity. These findings demonstrate that RNA ligand identification can be effectively achieved by combining DNA-encoded library technology with computational approaches for rational design and analysis and highlight a broadly adaptable platform for RNA-targeted small-molecule discovery.

Research Background

Protein–ligand interaction (PLI) prediction is a central task in computational drug discovery. Existing public affinity datasets such as BindingDB and ChEMBL are highly heterogeneous in origin, having been aggregated from thousands of laboratories using many different experimental protocols. As a result, they suffer from systematic bias and substantial standardization challenges, which in turn limit model generalization. In contrast, DNA-encoded library (DEL) technology enables ultrahigh-throughput screening of billions of compounds under unified experimental protocols, providing a new source of large-scale, high-quality training data for PLI modeling.

Model Architecture

Hermes adopts a lightweight Transformer-based architecture. Its main components are:

• Pretrained sequence encoders: ESM2-150M for protein sequences, with only the final four layers trainable, and ChemBERTa-77M-MTR for ligand SMILES strings, with all parameters trainable.
• Joint cross-attention module: alternating self-attention blocks, which process protein and ligand sequences separately, and cross-attention blocks, which enable information exchange between protein and ligand tokens.
• Attention pooling layers: differentiable pooling functions that learn importance scores and aggregate variable-length sequences into fixed-dimensional vectors.
• Prediction head: a multilayer perceptron (MLP) that outputs a binding probability.

For final inference, the authors used an ensemble of nine checkpoints trained with different hyperparameter settings and training-sampling strategies, and averaged their predictions.

Figure 1. Hermes architecture diagram.

Training Data

Hermes was trained on DEL screening data generated from the Kin0 chemical library, which contains 6.5 million members and was constructed as a three-cycle library using 38 cores connected to 384 and 446 building blocks. The dataset covers 239 unique protein targets, approximately two-thirds of which are kinases.

Labels were generated through a binarized hit-calling procedure based on enrichment relative to control screens, including DEL-only, bead-only/no-target, and proprietary controls. To manage class imbalance, the training procedure capped the number of positive samples per protein target, retained the highest-enrichment hits, and paired each positive example with a fixed number of negatives, drawn from both random negatives and hard negatives.

Table 1. Training and evaluation dataset statistics.

Model Evaluation

Hermes was evaluated on four benchmark datasets designed to test different forms of generalization:

1. DEL Protein Split: 164 protein targets not seen during training, screened against the same Kin0 library. This benchmark tests cold-target generalization.

2. DEL Chemical Library Split (STRELKA): 59 protein targets seen in training, but screened against a different 1-million-member benzimidazole library (AMA020). This benchmark tests cold-ligand generalization.

3. Public Binders/Decoys: 403 protein targets, with positives from Papyrus++ and negatives from GuacaMol property-matched synthetic decoys. This benchmark evaluates generalization to external public binding data.

4. MF-PCBA: 26 protein targets from PubChem BioAssay high-throughput screening data, where confirmed dose-response actives are positives and primary-screen inactives are negatives. This benchmark tests performance on heterogeneous public screening data.

The authors compared Hermes against two baselines:

• Boltz-2, a state-of-the-art structure-based deep learning model built on an AlphaFold3-like architecture.
• XGBoost, using concatenated ESM2-650M CLS embeddings and ECFP4 fingerprints as features.

Table 2. Hermes vs benchmarks per-protein AUROC comparison.

Key Findings

The results show clear variation across benchmarks, but several important patterns emerge:

• On DEL Protein Split, both Hermes and the XGBoost baseline outperform Boltz-2, suggesting that DEL-derived training data contains strong information value when the assay system is consistent.
• On DEL Chemical Library Split, all models perform more weakly, indicating that this benchmark is difficult for cold-ligand generalization.
• On MF-PCBA, Boltz-2 substantially outperforms Hermes, although part of this benchmark may overlap with Boltz-2’s training data.
• On Public Binders/Decoys, Hermes clearly outperforms XGBoost, demonstrating stronger generalization to genuinely novel chemical space.

The paper also reports that Hermes performs better on kinase targets than on non-kinases in most benchmarks, which is consistent with the kinase-enriched composition of the training set. This suggests that targeted DEL data generation can improve generalization within a protein family.

Computational Efficiency

Table 3. Inference speed comparison.

Hermes is designed for efficient inference. After correcting for hardware differences, the authors estimate that Hermes is approximately 500–700 times faster than Boltz-2. Its sequence-only design allows protein embeddings to be cached, making it well suited for billion-scale virtual screening campaigns.

Discussion and Conclusions

This study shows that a model trained exclusively on DEL screening data can learn transferable representations of protein–ligand interactions. Without ever being trained on traditional affinity measurements, Hermes generalizes to unseen protein targets, unseen chemical scaffolds, and external datasets derived from different experimental systems. This provides a strong proof of concept for the use of DEL data in PLI modeling.

At the same time, the study identifies several limitations. First, the kinase-heavy training distribution leads to weaker performance on non-kinase proteins. Second, binary label binarization likely restricts model expressiveness. Third, performance on data similar to the training set still appears to be influenced by memorization effects. Future work may benefit from incorporating structure-prediction outputs, modeling continuous enrichment scores instead of binary labels, and improving sampling strategies to better support generalization to unseen protein families.

Overall, as DEL technology continues to mature and generate data at a faster pace than public affinity databases, DEL-trained models such as Hermes are likely to become an important methodology for the next generation of PLI prediction

Reference

Kleinsasser M, Halverson B J , Kraft E ,et al.Hermes: Large DEL Datasets Train Generalizable Protein-Ligand Binding Prediction Models[J].  2026.

Fiona Shilliday,Miguel Gancedo-Rodrigo,Ginto George,Shintaro Aibara,Santosh Adhikari,Syedah Neha Ashraf,Evelyne J Barrey,Paolo A Centrella,Damian Crowther,Paige Dickson,Diana Gikunju,Marie-Aude Guié,John P Guilinger,Anders Gunnarsson,Heather P Harding,Christopher D Hupp,Rachael Jetson,Anthony D Keefe,JeeSoo Monica Kim,Richard J Lewis,Taiana Maia de Oliveira,Jennifer Le-Marshall,Usha Narayanan,Katherine A Nugai,Dušan Petrović,Emma Rivers,David Ron,Daisy Stringfellow,Karl Syson,Lewis Ward,John T S Yeoman,Yan Yu,Ying Zhang,Alisa Zyryanova,David J Baker,Perla Breccia,John E Linley

Nature Communications

DOI: 10.1038/s41467-026-72688-y

Abstract

Eukaryotic initiation factor 2B (eIF2B), a guanine nucleotide exchange factor (GEF), promotes protein synthesis by charging translation initiation factor 2 (eIF2) with GTP. Stress-induced phosphorylation of eIF2 on its α-subunit [eIF2(αP)] inhibits this reaction triggering a protective Integrated Stress Response (ISR). A DNA-encoded chemical library (DEL) screen for modulators of eIF2B, led to the identification of a chemical series that stabilises the inactive state of eIF2B, stimulating the ISR. Cryo-EM of compound-bound eIF2B reveals a conformational switch to the inactive state engaged by eIF2(αP). In cells, compound activity is sensitive to eIF2's phosphorylation state and to a competing eIF2B ligand (ISRIB) that activates the GEF allosterically. These findings establish the feasibility of targeting eIF2B with a drug-like allosteric inhibitor, that serves as an ISR activator (ISRAC), paving the way to explore the therapeutic potential of eIF2B-directed ISR activation.

Daniel Saliba,Eiman A Osman,Abdelrahman Elmanzalawy,Christopher Saab,Son Bui,Serhii Hirka,Shaun Anderson,Violeta Toader,Michael D Dore,Felix J Rizzuto,Donatien de Rochambeau,Maureen McKeague,Hanadi F Sleiman

Nature Chemistry

DOI: 10.1038/s41557-026-02099-5

Abstract

Aptamers are a versatile alternative to antibodies as they are smaller, easier to synthesize and less immunogenic. However, while antibodies are composed of 20 chemically diverse amino acids and are established therapeutics, aptamers are composed of only 4 similar nucleobases, thereby limiting their therapeutic potential. Aptamer chemical modifications are restricted to maintain compatibility with enzymatic selection. Here we introduce aptamer-like encoded oligomers (alenomers), highly chemically modified aptamers that are read and sequenced using a DNA code branching from and corresponding to the target-binding oligomer. We build ~300,000-member DNA-encoded libraries using an automated DNA synthesizer and split-and-pool methods, and screen them for protein binding via next-generation sequencing. In contrast to aptamers, alenomers are not restricted by the need for conservative enzyme-compatible modifications. They can thus explore an almost limitless chemical space, enabling the discovery of highly stable, high-affinity protein-binding aptamers, while offering structural insights into their interactions with target molecules.

Tonglin Yu , Jian Ma , Xiaodi Su , Jianhong Tang , Yujian He , Xiangyu Chen , Li Wu

Advanced Synthesis & Catalysis

DOI: 10.1002/adsc.70499

Abstract

DNA‐encoded libraries (DELs) are a powerful technology increasingly used in drug discovery for screening lead molecules. The key to success is dependent on the chemical space covered by DELs. Enzymes can catalyze complex chemical reactions under mild conditions, making them highly attractive for constructing structurally diverse DELs. However, traditional enzymatic transformations have been considered unsuitable for DEL construction due to their narrow substrate scope, leading to slow progress in this field. Here, we challenge this conventional perception by introducing the catalytic promiscuity of hydrolases into three‐component reactions on DNA. We successfully performed three enzyme‐promoted coupling reactions directly on DNA: the Aza‐Diels–Alder reaction, the Biginelli reaction, and the Mannich reaction.These reactions bring N‐heterocyclic bridged rings, pyrimidines, and α‐branched amine‐based nitrogen‐containing pharmacophores to DELs. Importantly, the entire coupling process on DNA tags does not require the use of harmful metals or stoichiometric organic catalysts, nor does it involve additional immobilization of the DNA strand. Using green and inexpensive hydrolases, the reactions can proceed directly in aqueous mixed solutions. Due to the mildness of enzyme‐catalyzed reaction conditions, all three reactions are highly compatible with DNA tags.

Miklos Feher , Rebecca J. Swett , Ryan T. Walsh , Erin Davis , Christopher I. Williams

ACS Medicinal Chemistry Letters

DOI: 10.1021/acsmedchemlett.6c00141

Abstract

DNA-encoded library (DEL) screening enables identification of small-molecule binders from libraries containing billions of compounds, yet much of the resulting structure–activity relationship (SAR) information remains underutilized. Here, we describe DEL2PH4, an automated ligand-based workflow that converts DEL screening data into three-dimensional pharmacophore models by integrating statistically enriched compounds with structurally related unenriched analogs, which serve as negative examples during model construction. The resulting pharmacophores capture consensus interaction features across DEL families and enable the extraction of actionable 3D SAR information from primary DEL screening data, independent of resynthesis or activity measurements. Application to a MerTK kinase DEL screen demonstrates strong enrichment of positives over decoy molecules in retrospective benchmarking, recovery of known experimentally validated actives from external data sets, and consistency with experimentally determined X-ray binding modes. DEL2PH4 provides a general strategy for translating DEL screening outputs into interpretable 3D models that support virtual screening, scaffold hopping, and medicinal chemistry optimization.