TriMod-DTI: Triple-Modal Interaction Framework

• TriMod-DTI is a triple-modal contrastive learning framework that integrates sequential, graphical, and geometric features from drugs and proteins.
• It employs modality-specific feature extractors with cross-modal alignment to capture universal and complementary representations for drug-target interaction prediction.
• Empirical evaluations on Human, GPCR, and DrugBank benchmarks demonstrate improved AUC and AUPR compared to state-of-the-art methods.

TriMod-DTI is a triple-modal contrastive learning framework for drug-target interaction prediction that incorporates 1D sequences, 2D graphs, and 3D structures of drugs and proteins in order to obtain universal and complementary feature representations. It was introduced to address a limitation in existing methods that often rely on single-modal representations or combine only two modalities while overlooking 3D structural features. The framework couples modality-specific feature extraction with cross-modal alignment in latent space, and experiments on Human, GPCR, and DrugBank benchmarks report improved performance relative to state-of-the-art methods (Xu et al., 28 May 2026).

1. Concept and representational objective

TriMod-DTI is organized around three parallel Feature Extractor strands for each modality of both drugs and targets: 1D, 2D, and 3D. For drugs, the resulting embeddings are denoted $d_1,d_2,d_3$; for proteins, they are denoted $t_1,t_2,t_3$. These embeddings are aligned via contrastive learning and concatenated for interaction prediction (Xu et al., 28 May 2026).

The representational scope is explicitly asymmetric across biomolecular object types where needed. In the 1D branch, the drug input is a SMILES sequence segmented into $l_d$ substructures, while the protein input is an amino-acid sequence segmented into $l_p$ tokens. In the 2D branch, the drug is represented as a molecular graph, whereas the protein is represented through a protein binding-site graph. In the 3D branch, both drugs and proteins are encoded as geometric graphs derived from atomic coordinates. This suggests that the framework is designed to preserve modality-specific inductive biases rather than enforcing a uniform encoder family across all inputs.

The stated objective is not merely multimodal fusion. The model is constructed to recover both universal and complementary information across modalities. The universal component is pursued through latent-space alignment of different views of the same drug or protein; the complementary component is retained through parallel encoding and subsequent concatenation. In that sense, TriMod-DTI combines representation agreement and representation diversity within a single DTI architecture.

2. Triple-strand feature extraction

The 1D sequence encoder uses an embedding layer with shared dimension $\vartheta$, followed by a Transformer encoder with 6 layers, 4 self-attention heads, and hidden dimension $\vartheta$. Global mean pooling over tokens produces sequence embeddings

$$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$

with $d_1,t_1 \in \mathbb{R}^{D_s}$ (Xu et al., 28 May 2026).

The 2D graph encoder uses distinct graph formulations for drugs and proteins. For drugs, the molecular graph is $G_{d_2}=(V,E)$ with $N_a$ atoms, node features $t_1,t_2,t_3$0, and adjacency matrix $t_1,t_2,t_3$1. Drug graph propagation is implemented with 2 GCN layers: $t_1,t_2,t_3$2 Global add pooling over atoms yields

$t_1,t_2,t_3$3

For proteins, each binding-site graph is encoded by a Topology Adaptive Graph Convolutional Network: $t_1,t_2,t_3$4 followed by attention-based pooling across nodes and a final MLP to obtain $t_1,t_2,t_3$5.

The 3D structural encoder constructs a graph $t_1,t_2,t_3$6 with nodes at atomic coordinates and edges defined by inter-atomic distance $t_1,t_2,t_3$7, with node scalar and vector features. It uses Geometric Vector Perceptron layers to process scalar $t_1,t_2,t_3$8 and vector $t_1,t_2,t_3$9 parts, followed by a 2-layer GVP-GNN message-passing scheme: $l_d$0 Global add pooling yields 3D embeddings $l_d$1 and $l_d$2.

Taken together, the three branches cover token-level sequence context, topological neighborhood structure, and local geometric configuration. A plausible implication is that the architecture is intended to reduce the representational blind spots associated with any single modality.

3. Cross-modal contrastive objective

TriMod-DTI aligns the three representations $l_d$3 of the same drug in a shared latent space, and likewise aligns $l_d$4 for proteins. Positive pairs are any two modalities of the same sample, such as $l_d$5, $l_d$6, and $l_d$7. Negative pairs are cross-sample pairs, such as $l_d$8 and $l_d$9 for $l_p$0 (Xu et al., 28 May 2026).

For a modality pair $l_p$1, with batch size $l_p$2, temperature $l_p$3, and cosine similarity

$l_p$4

the contrastive loss is defined as

$l_p$5

Swapping $l_p$6 gives the reverse-direction loss, and averaging over the three modality pairs gives the drug-level objective

$l_p$7

An identical loss $l_p$8 is used for proteins.

The hyperparameterization reported for this component is specific: cosine similarity is used throughout, the temperature is set to $l_p$9, and all three modality pairs are equally weighted with weights $\vartheta$0. The stated purpose of this design is to enhance discriminative ability by constructing cross-modal positive and negative sample pairs. This suggests that the alignment term is intended not only to regularize the embeddings but also to make inter-modal agreement predictive of interaction-relevant identity.

4. Optimization setup and benchmark protocol

Training uses a composite loss consisting of binary cross-entropy for the predicted interaction $\vartheta$1 and two contrastive terms: $\vartheta$2 The binary cross-entropy term is

$\vartheta$3

Optimization uses Adam with learning rate $\vartheta$4 (best on GPCR), weight decay $\vartheta$5, batch size $\vartheta$6, and dropout $\vartheta$7 (Xu et al., 28 May 2026).

After cleaning to ensure complete 3D information, evaluation is performed on three benchmark datasets.

Dataset	Entities	Interactions
Human	2,175 drugs, 1,540 proteins	4,965 interactions (2,330 positive / 2,635 negative)
GPCR	4,976 drugs, 330 proteins	14,376 interactions (7,479 pos / 6,897 neg)
DrugBank	1,441 drugs, 1,262 proteins	13,121 interactions (5,634 pos / 7,487 neg)

The split protocol differs by dataset. Human and DrugBank use 5-fold cross-validation with train:val:test $\vartheta$8. GPCR uses 80% training data, with 20% of that training partition used for validation, and 20% test data. Each split is averaged over 10 runs.

The reported metrics are AUC, AUPR, and Precision at decision threshold $\vartheta$9. AUC is given as

$\vartheta$0

and AUPR as

$\vartheta$1

This evaluation design places explicit weight on both ranking quality and thresholded decision quality.

5. Empirical performance relative to baselines

Across all three benchmarks, TriMod-DTI is reported to outperform or tie the best current methods. On Human, the best baseline is CSCL-DTI with AUC $\vartheta$2 and AUPR $\vartheta$3, whereas TriMod-DTI reaches AUC $\vartheta$4 and AUPR $\vartheta$5, with Precision $\vartheta$6. On GPCR, the best baseline is MGMA-DTI with AUC $\vartheta$7 and AUPR $\vartheta$8, while TriMod-DTI reaches AUC $\vartheta$9, AUPR $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$0, and Precision $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$1. On DrugBank, the best baseline AUC is $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$2 from CSCL-DTI, while TriMod-DTI reaches AUC $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$3, AUPR $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$4, and Precision $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$5 (Xu et al., 28 May 2026).

Dataset	Best baseline	TriMod-DTI
Human	CSCL-DTI: AUC $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$6, AUPR $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$7	AUC $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$8, AUPR $$d_1 = \mathrm{Pool}(\tilde E^d), \qquad t_1 = \mathrm{Pool}(\tilde E^p),$$9, Precision $d_1,t_1 \in \mathbb{R}^{D_s}$0
GPCR	MGMA-DTI: AUC $d_1,t_1 \in \mathbb{R}^{D_s}$1, AUPR $d_1,t_1 \in \mathbb{R}^{D_s}$2	AUC $d_1,t_1 \in \mathbb{R}^{D_s}$3, AUPR $d_1,t_1 \in \mathbb{R}^{D_s}$4, Precision $d_1,t_1 \in \mathbb{R}^{D_s}$5
DrugBank	CSCL-DTI: AUC $d_1,t_1 \in \mathbb{R}^{D_s}$6	AUC $d_1,t_1 \in \mathbb{R}^{D_s}$7, AUPR $d_1,t_1 \in \mathbb{R}^{D_s}$8, Precision $d_1,t_1 \in \mathbb{R}^{D_s}$9

The relative improvements are stated explicitly for two datasets: on Human, AUC improves by $G_{d_2}=(V,E)$0 and AUPR by $G_{d_2}=(V,E)$1; on GPCR, AUC improves by $G_{d_2}=(V,E)$2, AUPR by $G_{d_2}=(V,E)$3, and Precision by $G_{d_2}=(V,E)$4. Paired t-tests over 10 runs indicate that the AUC gains on Human and GPCR are significant at $G_{d_2}=(V,E)$5. This suggests that the effect is not confined to a single benchmark regime, although the absolute margins differ across datasets.

6. Ablation results and practical use in drug discovery

The ablation studies focus on two questions: the contribution of contrastive learning and the contribution of each modality. On GPCR, removing all $G_{d_2}=(V,E)$6 decreases AUC by $G_{d_2}=(V,E)$7 and AUPR by $G_{d_2}=(V,E)$8. Removing any single cross-modal component, $G_{d_2}=(V,E)$9, $N_a$0, or $N_a$1, yields a $N_a$2–$N_a$3 AUC drop. These results are used to validate the contrastive alignment strategy (Xu et al., 28 May 2026).

The modality study on GPCR reports the following AUC values: sequence only $N_a$4 $N_a$5; graph only $N_a$6 $N_a$7; 3D only $N_a$8 $N_a$9; sequence + graph $t_1,t_2,t_3$00; graph + 3D $t_1,t_2,t_3$01; sequence + 3D $t_1,t_2,t_3$02; and all three modalities $t_1,t_2,t_3$03. Removing any modality from the full model leads to a $t_1,t_2,t_3$04–$t_1,t_2,t_3$05 drop in AUC. This suggests that each modality is complementary rather than redundant, and that pairwise fusion does not saturate the available information.

The case studies are framed as evidence of practical potential for DTI prediction and drug discovery. TriMod-DTI ranked Verapamil (DB00661) among the top 10 candidates for the D$t_1,t_2,t_3$06 dopamine receptor (P14416), and 5/10 predicted targets are supported in PubMed. In a molecular docking validation, the top predicted target Glucose-6-phosphate isomerase 2 (P13376) was docked with Verapamil, and key hydrogen bonds were observed at Thr-274, Gly-151, and Ser-149. The reported practical advantages are that the triple-modal embeddings capture both global sequence/graph context and local 3D geometry, and that contrastive alignment reduces modality gaps, producing robust representations for high-confidence DTI screening. The framework is also released at https://github.com/klez1/TriMod-DTI.

7. Acronym ambiguity and disciplinary scope

A recurring source of ambiguity is the acronym “DTI.” In TriMod-DTI, DTI denotes drug-target interaction prediction, and the model operates in computational drug discovery through joint encoding of drug and protein modalities (Xu et al., 28 May 2026).

By contrast, “DTI” is also widely used to denote diffusion tensor imaging in neuroimaging. An example is the multimodal autism study “Multimodal Functional and Structural Brain Connectivity Analysis in Autism: A Preliminary Integrated Approach with EEG, fMRI and DTI,” where DTI is one of the neuroimaging modalities alongside EEG and fMRI, and the methodological focus is structure–function correlation in the brain rather than drug-target interaction modeling (Cociu et al., 2018). This suggests that the shared acronym should not be interpreted as a methodological connection between the two works.

Within its own scope, TriMod-DTI belongs to the line of multimodal DTI predictors that move beyond sequence-only or graph-only formulations by explicitly incorporating 3D structure. Its distinctive claim is therefore not multimodality in the abstract, but triple-modal contrastive learning over sequence, graph, and 3D representations for both drugs and proteins.