The problem the researchers set out to address has two intertwined parts. Graph-based affinity models typically encode drugs as molecular graphs, where atoms are nodes and chemical bonds are edges, and proteins as residue graphs built from structural or sequence information. Many of the most accurate methods in this space fall into one of two camps. Adaptive-topology approaches learn the graph structure itself during training, refining which nodes should exchange information, but they usually learn the drug graph and the protein graph separately and only merge the two representations at the final prediction layer. Cross-attention approaches, by contrast, explicitly model the interaction between drug and protein nodes, but they do so on top of graph topologies that are either fixed in advance or learned without any knowledge of the partner molecule. In both cases, the topology of the graph — the very skeleton along which information flows — is never conditioned on the specific binding partner.

CGA-DTA closes this gap with a three-stage architecture. The first component, called Adaptive Multi-head Graph Learning, or AMGL, learns multiple soft adjacency heads for each graph. In practical terms, rather than committing to a single fixed set of edges connecting atoms or amino acid residues, the model maintains several candidate connectivity patterns and learns how to blend them. Each head represents a different plausible way of propagating information across the molecular graph, and the model learns a weighting over these heads during training. This soft mixture allows the network to capture multiple modes of interaction-relevant structure that a single hard-coded adjacency matrix would miss.

The second and arguably most distinctive component is the Interaction-Conditioned Topology gate, or ICT. This module takes a global summary of the binding partner — a compressed vector representation of the drug or protein on the other side of the proposed interaction — and uses it to re-weight the mixture of adjacency heads learned by AMGL. The consequence is subtle but important: the effective message-passing kernel, the row-stochastic matrix that governs how information flows between nodes, changes depending on which partner is being considered. A drug candidate’s internal graph representation is therefore not the same for every protein it is screened against; it is reshaped in light of the specific target. The authors are careful to note what this mechanism does and does not do. ICT does not predict physical contact points between the drug and the protein, nor does it reconstruct the topology at the level of individual nodes or edges. It re-weights learned message-passing kernels, which is a statistical operation on the learned connectivity mixture rather than a claim about atomic geometry.

The third component, the Cross-Graph Attention Module or CGAM, then applies bidirectional node-level attention to these interaction-conditioned representations. Attention mechanisms, borrowed from the transformer architectures that have transformed natural language processing, allow each node in the drug graph to attend to relevant nodes in the protein graph and vice versa. Because this cross-attention operates on representations that have already been conditioned on the partner through ICT, the interaction modeling happens on a foundation that is already partner-aware — a coupling the authors argue is missing from prior pipelines where cross-attention sits atop static or independently learned topologies.

On the protein side, the framework makes a deliberate scalability choice. CGA-DTA uses residue embeddings from ESM-2, a protein language model, as alignment-free node features. Traditional pipelines in this field often rely on features derived from multiple-sequence alignment, or MSA, which requires searching large sequence databases to find evolutionary relatives of a target protein. MSA-based features can be informative, but they are computationally expensive and become a bottleneck when screening large libraries or newly annotated targets for which no rich alignment data exists. By drawing node features from a pretrained protein language model, CGA-DTA sidesteps the alignment requirement for its node representations. The authors acknowledge, however, that one MSA-dependent element remains: the initial protein graph topology is built from contact maps produced by PconsC4, a contact prediction tool, so the framework is not yet fully alignment-free.

The empirical results span four benchmark datasets: Davis, KIBA, Metz, and ToxCast. On Davis, CGA-DTA achieved a mean squared error of 0.168 with a concordance index of 0.915; on KIBA, 0.116 with a concordance index of 0.910; on Metz, 0.231 with a concordance index of 0.854; and on ToxCast, 0.149 with a concordance index of 0.920. The concordance index, a standard ranking metric in affinity prediction, measures how well the model orders pairs by binding strength, with values above 0.9 generally considered strong performance. The mean squared error captures how closely predicted affinity values match measured ones. Together, the numbers position CGA-DTA among the competitive methods in the field, though the more instructive finding comes from the ablation studies.

When the authors systematically removed components of the model, a consistent pattern emerged across all four datasets. The Cross-Graph Attention Module contributed the largest individual gain, confirming that explicit bidirectional interaction modeling between drug and protein nodes remains a cornerstone of accurate affinity prediction. The Interaction-Conditioned Topology gate delivered a smaller but consistent additional benefit, supporting the paper’s central thesis that partner-conditioned re-weighting of learned message-passing kernels is a useful design principle. The ablation evidence, the authors stress, supports the value of coupling adaptive graph learning with cross-graph attention, but it does not establish that the learned kernels correspond to physical protein–ligand contact maps — a distinction that matters for anyone hoping to interpret the learned connectivity as structural biology.

The authors outline two clear next steps for the field. The first is to replace the remaining MSA-based contact-map stage, which would make the entire pipeline alignment-free and further improve scalability for large or newly annotated target sets. The second is to validate the learned kernels against experimentally resolved protein–ligand complexes, testing whether the connectivity patterns the model discovers bear any relationship to actual binding geometry. Until such validation is performed, the learned topologies should be regarded as computational devices that improve prediction rather than as interpretable structural hypotheses. The study was supported by the Zhejiang Province Medical and Health Science and Technology Plan Project, and the corresponding author is Wenwen Yu of the School of Information Science and Engineering at East China University of Science and Technology. As virtual screening continues to expand as a first filter in drug discovery pipelines, approaches that let molecular representations adapt to the question being asked — rather than the other way around — are likely to attract growing attention, and CGA-DTA offers a concrete, tested template for how that adaptation can be built into the architecture itself.

Subject of Research: Interaction-conditioned graph topology and cross-graph attention for drug–target affinity prediction

Article Title: CGA-DTA: interaction-conditioned graph topology and cross-graph attention for drug–target affinity prediction

Article References: CGA-DTA: interaction-conditioned graph topology and cross-graph attention for drug–target affinity prediction. (n.d.). https://doi.org/10.1186/s12859-026-06647-1

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06647-1

Keywords: drug–target affinity, graph neural network, cross-graph attention, adaptive graph learning, interaction-conditioned topology, protein language model, ESM-2, virtual screening, drug discovery, BMC Bioinformatics, message passing, contact map

The study was led by Professor Rachel Kolodny and PhD candidate Guy Yanai of the University of Haifa, together with Professor Nir Ben-Tal and graduate student Gabriel Axel of Tel Aviv University, and Specially Appointed Associate Professor Liam M. Longo of ELSI. Kolodny also spent five months as a visiting researcher at ELSI, developing methods to analyze the new model. Their creation, dubbed CLSS for Contrastive Learning Sequence-Structure, is a protein language model designed to overcome a stubborn problem that has limited previous computational approaches: the awkward relationship between what a protein’s sequence says and what its structure actually does.

Scientists have long organized proteins into hierarchical groups based on relatedness, much like the genus and species categories biologists use to classify organisms. These curated systems, such as the widely used ECOD and CATH databases, distill decades of expert knowledge. But with artificial intelligence now capable of generating ’embeddings’—numerical representations in which proteins with similar properties receive nearby coordinates, like postal codes on a map—researchers can visualize relationships across millions of proteins at once, producing what the team calls a protein world map. The catch is that sequence and structure do not map neatly onto each other. Unrelated sequences can fold into similar shapes, while even identical sequences can sometimes adopt wildly different structures.

Most existing protein language models treat sequence and structure as separate worlds, processing one or the other independently. Even hybrid models that incorporate both kinds of data rarely place the sequence and the structure of the same protein at the same location on a global map, leaving researchers with two conflicting atlases of protein space. CLSS was engineered specifically to resolve this discordance. Using a machine learning strategy known as contrastive learning, the model is trained on pairs of protein sequences and their corresponding structures, learning to pull matching sequence-structure pairs together in the embedding space while pushing unrelated pairs apart.

The result is a single shared map in which a protein occupies essentially the same location whether the model is given its sequence or its structure. When benchmarked against other state-of-the-art protein language models, CLSS succeeded in producing a cohesive unified representation, something its predecessors could not achieve. Remarkably, the model’s maps closely reproduced the relationships recorded in the expert-curated ECOD and CATH classification systems, even though those classifications were never shown to the model during training. In direct classification tests, CLSS also performed strongly, demonstrating that merging sequence and structure information yields genuinely more informative protein representations.

Perhaps the most exciting feature of CLSS is its ability to handle fragments. Most protein language models require a complete sequence or structure to generate a meaningful embedding, but CLSS showed that short sequence fragments can in many cases be positioned meaningfully alongside full-length proteins and structures. This capability matters enormously for evolutionary studies, because small pieces of proteins have been repeatedly reused and rearranged throughout the history of life. Some fragments may even have served as the primordial building blocks from which the earliest protein domains were assembled, meaning that similar fragments appearing in otherwise unrelated proteins can hint at ancient evolutionary connections.

The maps produced by CLSS also revealed sweeping patterns across protein space that were previously difficult to see. When the researchers overlaid biological properties onto the maps, proteins associated with organic cofactors turned out to cluster in particular regions, while metal-binding proteins were scattered more broadly. Such patterns illustrate how global protein maps can serve not only as classification tools but as instruments for exploring the interplay between sequence, structure, function, and deep evolutionary history, potentially exposing large-scale patterns invisible to conventional pairwise comparison methods.

‘This gives us a way to look at the protein universe through sequence and structure at the same time, rather than treating them as separate worlds,’ said Longo. ‘What is particularly exciting for us is the possibility of using these maps to uncover large-scale evolutionary patterns that are difficult to recognise using conventional approaches.’ The team ultimately envisions unified sequence-structure representations opening new frontiers in database searches, protein engineering, and the reconstruction of evolutionary trajectories—offering a fresh window onto how the staggering diversity of proteins found in life today emerged over nearly four billion years of evolution.

Subject of Research: A contrastive-learning protein language model that unifies protein sequence and structure representations to map the protein universe

Article Title: Uniting sequence and structure to map the protein universe

Article References: Uniting sequence and structure to map the protein universe. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: protein language model, CLSS, protein evolution, contrastive learning, protein structure, amino acid sequence, ECOD, CATH, protein embeddings, evolutionary biochemistry, protein classification, Institute of Science Tokyo