When modeling how drugs interact with targets, or how proteins bind to one another, traditional deep learning models often treat the participants as monolithic entities. A drug might be encoded as a single global graph embedding, while a protein is represented by a single vector summarizing its entire amino acid sequence.
However, this holistic framing introduces a fundamental granularity gap. In physical reality, biochemical interactions do not occur between entire molecules as uniform wholes. Instead, they resemble a precise "handshake" where a specific chemical functional group or molecular fragment fits into a localized pocket or microenvironment of a target. When models rely purely on global representations, these critical local signals are washed out by the background noise of the rest of the molecule.
The Granularity Gap in Molecular Modeling
Early conceptual frameworks recognized this limitation. Michel Dumontier in Situational Modeling: Defining Molecular Roles in Biochemical Pathways and Reactions noted that treating entire molecules as participants obscures the precise atoms or functional groups driving chemical changes.
Similarly, in protein structural biology, Liang Zhao et al. showed in Structural and Functional Analysis of Multi-Interface Domains that analyzing proteins at either the whole-molecule or individual amino acid level misses the intermediate scale of specific binding patches—the true drivers of multi-partner coordination. Without modeling this intermediate scale, systems cannot accurately capture how distinct surface regions on a single domain coordinate complex interactions with multiple partners.
Decomposing Molecules into Functional Vocabularies
To bridge this gap, a major shift has emerged toward decomposing complex biological entities into functional subunits. For small-molecule drugs, this means moving away from whole-molecule graphs toward frequent functional units, chemical motifs, or fragments.
The CASTER framework by Kexin Huang et al. (CASTER: Predicting Drug Interactions with Chemical Substructure Representation) breaks down chemical representations into discrete functional units via sequential pattern mining. The authors demonstrate that this granular input captures the underlying mechanisms of chemical reactivity far better than holistic graphs.
This approach is vital for clinical safety. As Chaoqi Yang et al. demonstrate in SafeDrug: Dual Molecular Graph Encoders for Recommending Effective and Safe Drug Combinations, adverse drug-drug interactions are typically triggered by specific local substructures rather than the entire molecule. By isolating these functional groups, models can explicitly flag or avoid dangerous combinations. Similarly, Tengfei Ma et al. in ImageDDI: Image-enhanced Molecular Motif Sequence Representation for Drug-Drug Interaction Prediction combine local chemical motifs with global visual context to ensure that the specific subunits driving reactions are not overlooked.
On the protein side, a parallel decomposition occurs. Rather than evaluating a target structure at a global level—which Xin Zeng et al. note in Joint fusion of sequences and structures of drugs and targets for identifying targets based on intra and inter cross-attention mechanisms can miss fine-grained details of the interaction site—modern architectures isolate specific binding pockets and surface regions.
In SP-DTI: subpocket-informed transformer for drug-target interaction prediction, Jun Xia et al. decompose protein binding sites into atomic-level "subpockets." For protein-protein interactions, Lirong Wu et al. develop MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding, which encodes localized "microenvironments" into a discrete, learned vocabulary. This allows the model to pinpoint the exact residues forming the binding interface instead of outputting a simple binary classification.
Modeling the Substructural Handshake
Once both interactants are broken down into their respective vocabularies of fragments and subpockets, the interaction itself is modeled as a fine-grained matching problem. This is typically achieved using token-level fusion and cross-attention mechanisms.
Holistic Modeling (Traditional):
[ Drug Graph ] ───> Global Pooling ───┐
├───> [ Interaction Prediction ]
[Target Sequence] ─> Global Pooling ───┘
Substructural Modeling (Modern):
[ Drug Fragments ] ───┐
├─── [ Token-level Cross-Attention ] ───> [ Fine-grained Binding ]
[Target Subpockets] ──┘
For example, Zhaohan Meng et al. introduce FusionDTI: Fine-grained Binding Discovery with Token-level Fusion for Drug-Target Interaction, which uses token-level cross-attention to model the direct "handshake" between individual drug atoms and protein amino acid residues. By avoiding global pooling layers until after the cross-attention phase, the model preserves the exact spatial and chemical compatibility of the local sites.
Similarly, Ali Khodabandeh Yalabadi et al. develop FragXsiteDTI: Revealing Responsible Segments in Drug-Target Interaction with Transformer-Driven Interpretation, which uses a learnable "mediator" array to bridge the distinct structural languages of molecular fragments and protein binding pockets. This simultaneous analysis mimics the physical reality of molecular docking, where local structural compatibility dictates binding affinity.
This fragment-oriented matching strategy is echoed by Qi Zhang et al. in FMCA-DTI: a fragment-oriented method based on a multihead cross attention mechanism to improve drug–target interaction prediction, demonstrating that multihead cross-attention over fragments captures essential substructural information that global representations omit.
Why Substructural Modeling Generalizes to the "Cold Start"
The primary reason this approach works so well—especially in "cold-start" scenarios where a drug or protein is entirely novel—is that while global molecular configurations are vast and rare, the underlying vocabulary of functional substructures is compact and highly recurrent.
When a model learns how a specific chemical fragment interacts with a specific protein subpocket, that knowledge transfers perfectly to a brand-new drug that happens to contain the same fragment. Hui Yu et al. leverage this in STNN-DDI: A Substructure-aware Tensor Neural Network to Predict Drug-Drug Interactions by modeling substructure interactions within a 3D tensor space, enabling highly accurate inductive predictions for novel compounds.
This principle extends to genomic scales. Yuanpeng Zhang et al. show in DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction that mapping interactions between chemical fragments and functional genomic pathways captures robust biological mechanisms that remain consistent across different experimental environments, vastly outperforming global gene expression models.
Enforcing Hierarchical Consistency
While local interactions are the primary drivers, they must ultimately align with global biological realities. Modern architectures therefore fuse these granular insights back into a multi-level hierarchy.
In MolHF: Molecular Heterogeneous Attributes Fusion for Drug-Target Affinity Prediction on Heterogeneity, Runze Wang et al. fuse local atom associations, global topology, and sequential bond arrangements simultaneously. To ensure that these different scales do not contradict one another, Claudio Saccà et al. introduce Semantic Based Regularization in Improved multi-level protein¿protein interaction prediction with semantic-based regularization. Their framework enforces biological consistency across hierarchical levels, ensuring that if two whole proteins are predicted to interact, their constituent domains and residues must also exhibit valid binding patterns.
Finally, Yaqing Wang et al. demonstrate in Attending on Multilevel Structure of Proteins enables Accurate Prediction of Cold-Start Drug-Target Interactions that explicitly mapping specific chemical functional groups to their corresponding protein structural levels—from simple sequences to complex 3D folds—allows the system to accurately interpret how local molecular features scale up to dictate global binding affinity.
Key Foundational Work
Situational Modeling: Defining Molecular Roles in Biochemical Pathways and Reactions (2008) — contributed 1 concept elements to this direction.
Structural and Functional Analysis of Multi-Interface Domains (2012) — contributed 1 concept elements to this direction.