Organizations: College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · School of Computer and Communication Engineering, Northeastern University at Qinhuangdao, Qinhuangdao 066004, China · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · Shanghai Artificial Intelligence Laboratory, Shanghai 200232, China
Predicting drug--target binding affinity (DTA) requires models to distinguish subtle chemical and structural determinants underlying molecular recognition. Although recent approaches increasingly incorporate richer drug and protein information, such information may be compressed, homogenized or discretized during representation construction, causing affinity-relevant distinctions to be lost before interaction modelling. We hypothesized that this representation-stage information loss constitutes an upstream bottleneck that cannot be reliably overcome by increasingly complex interaction predictors. To test this hypothesis, we developed ReGeoDTA, a representation-preserving framework that maintains affinity-relevant chemical heterogeneity in molecular representations and continuous geometric relationships in protein structures. Across three benchmark datasets, ReGeoDTA consistently improved affinity prediction, and the proposed representation-preserving strategies retained their benefits across diverse DTA architectures. Controlled representation degradation progressively reduced predictive performance, whereas increasing downstream predictor complexity failed to recover information lost during representation construction. These findings identify representation fidelity as an upstream design principle for accurate and generalizable drug--target affinity prediction, with potential implications for computational compound prioritization.
Figures & tables
Fig. 1: Overview of ReGeoDTA. A Position-Aware Protein Encoder (PAPE) with relative geometric positional encoding. B Protein positional information processing module for generating processed relative geometric information from 3D protein coordinates. C Drug graph encoder with GMFF-based node feature representation. D Gated Multi-Feature Fusion (GMFF) module for heterogeneous atomic feature integration. E Overall architecture of ReGeoDTA.
Dataset
Scenario
Method
MSE ↓ (Std)
CI ↑ (Std)
R2↑ (Std)
Davis
Cold drug
GraphDTA [ 13 ]
0.920 (0.029)
0.678 (0.036)
0.160 (0.019)
GEFA [ 29 ]
0.847 (0.012)
0.709 (0.028)
0.182 (0.015)
FusionDTA [ 30 ]
0.581 (0.094)
0.737 (0.012)
0.187 (0.034)
MGraphDTA [ 31 ]
0.563 (0.065)
0.729 (0.022)
0.192 (0.021)
NHGNN [ 32 ]
0.554 (0.091)
0.752 (0.017)
0.207 (0.030)
ReGeoDTA
0.546 (0.090)
0.744 (0.028)
0.255 (0.030)
Table 2
Fig. 2: Randomization tests. Performance comparison under four corrupted conditions.
Dataset
Experiments
MSE ↓
CI ↑
MAE ↓
R2↑
AUPR ↑
Davis
Full Model
0.186
0.906
0.211
0.756
0.807
Exp. 1
0.187
0.904
0.211
0.755
0.806
Exp. 2
0.188
0.901
0.212
0.751
0.807
Exp. 3
0.200
0.895
0.222
0.722
0.782
Exp. 4
0.212
0.895
0.232
0.689
0.775
Exp. 5
0.195
0.899
0.218
0.724
0.786
Table 4
Fig. 3: Controlled representation degradation analysis across Davis, KIBA, and BindingDB. The complete ReGeoDTA model is used as the reference setting. Step 1 removes the interaction between aromaticity-related information and numerical atomic features; Step 2 removes numerical atomic descriptors; Step 3 removes aromaticity information and retains only atom-type information; Step 4 replaces the proposed drug encoding with conventional one-hot atom encoding; Step 5 replaces relative position-aware protein attention with contact-map-based attention; Step 6 removes protein positional information; and Step 7 replaces ProtTrans-derived protein representations with standard learnable embeddings. Lower MSE and higher CI, R2 , and AUPR indicate better performance. The overall degradation pattern suggests that preserving chemical heterogeneity, continuous protein geometry, and pretrained protein contextual information is important for affinity prediction.
Dataset
Initial Representation
Predictor
MSE ↓
CI ↑
R2↑
AUPR ↑
Davis
Enhanced initial representation
Linear
0.186
0.906
0.756
0.807
Cross-attention
0.211
0.895
0.685
0.771
LRBF
0.189
0.904
0.749
0.803
PMF
0.184
0.907
0.761
0.809
GIN-based initial representation
Linear
0.225
0.898
0.718
0.770
Cross-attention
0.229
0.896
0.712
0.767
Table 6
Fig. 4: Generalization of our drug encoding strategy across different graph models. In the radar plots, values farther from the centre indicate better performance.
Fig. 5: Generalization of the position-aware encoder across baseline models. In the radar plots, values farther from the centre indicate better performance.
Fig. 6: Decomposition and structural localization of PAPE-derived attention in the ABL1–Nilotinib complex (PDB ID: 3CS9). Images 1–4 show the semantic, distance-aware, orientation-aware, and fused residue-pair attention maps, respectively, illustrating how the attention pattern changes as geometric information is incorporated. Image 5 evaluates the fused map against 22 annotated binding-site residues assigned to four structurally defined regions, with blue indicating high fused attention ( ≥0.35 ) and red indicating lower fused attention ( <0.35 ). Image 6 analyses component complementarity: blue denotes other high-weight fused pairs, orange denotes high-weight pairs identified through positional complementarity, and green denotes high-weight pairs identified through semantic complementarity. The black open circle marks the representative M290–H361 long-range cross-region association.
Fig. 7: Attention, convolutional and graph-propagation maps represent different internal quantities and are presented for qualitative structural comparison (PDB ID: 3CS9). Image 1: Overall structure of ABL1. Image 2: Binding site with spatial functional groups. Image 3: Attention map from our proposed method. Image 4: Attention map from TransVAE. Image 5: Attention map from DeepTGIN. Image 6: Interaction map from CNN with a small convolution kernel. Image 7: Interaction map from CNN with a large convolution kernel. Image 8: Graph neural network propagation map under a low-threshold protein contact map. Image 9: Propagation map under a high-threshold contact map.
Fig. 8: Alignment between drug embedding similarity and multi-target affinity-profile similarity on the Davis dataset. Image 1: Alignment between functional and embedding similarity. Image 2: Bootstrap stability of spearman alignment. Image 3: Bootstrap stability of pearson alignment.
Drug--target affinity prediction is pivotal for accelerating drug discovery, yet existing methods suffer from significant performance degradation in realistic cold-start scenarios (unseen drugs/targets/pairs), primarily driven by overfitting to training instances and information loss from irrelevant target sequences. In this paper, we propose LaPro-DTA, a framework designed to achieve robust and generalizable DTA prediction. To tackle overfitting, we devise a latent dual-view drug representation mechanism. It synergizes an instance-level view to capture fine-grained substructures with stochastic perturbation and a distribution-level view to distill generalized chemical scaffolds via semantic remapping, thereby enforcing the model to learn transferable structural rules rather than memorizing specific samples. To mitigate information loss, we introduce a salient protein feature extraction strategy using pattern-aware top-k pooling, which effectively filters background noise and isolates high-response bioactive regions. Furthermore, a cross-view multi-head attention mechanism fuses these purified features to model comprehensive interactions. Extensive experiments on benchmark datasets demonstrate that LaPro-DTA significantly outperforms state-of-the-art methods, achieving an 8% MSE reduction on the Davis dataset in the challenging unseen-drug setting, while offering interpretable insights into binding mechanisms.
Zihan Dun, Liuyi Xu, An-Yang Lu +2
College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · School of Artificial Intelligence, Beihang University, Beijing 100000, China · School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China
Drug-target relation prediction supports candidate screening, drug repositioning, and mechanism analysis. Existing models often use incomplete drug or protein representations, model cross-modal interactions shallowly, or train affinity regression and interaction classification separately, although these tasks describe closely related views of the same drug-target pair. This paper presents DPTM-DT, a dual-pretrained Transformer framework for multitask drug-target prediction. DPTM-DT combines GROVER molecular graph embeddings, ESM protein language-model embeddings, and CTD physicochemical descriptors, then exchanges drug-target information through bidirectional cross-modal attention. A shared pair representation is used for continuous affinity regression, high-affinity binary classification, and six-level affinity classification. Experiments on Davis and KIBA cover random 80/20 and DeepDTA-style standard splits. On the random 80/20 split, DPTM-DT achieves MSE/CI values of 0.193/0.917 on Davis and 0.120/0.918 on KIBA. It also reports binary AUPR/MCC values of 0.727/0.654 and 0.798/0.689, and six-class Macro-F1/Top-2 values of 0.800/0.932 and 0.815/0.962 on Davis and KIBA, respectively. Across the reported regression, binary classification, and multiclass classification settings, DPTM-DT achieves the best overall performance among the compared methods. Results under the standard split show the same relative trend. Ablations indicate that dual target representation, gated fusion, and cross-modal attention each contribute to the final performance. Code and supplementary materials are available at: anonymous.4open.science/r/DPCM-DT-74E0.
Ge Kong
School of Biological Science and Medical Engineering, Beihang University, Beijing 100191, China
Drug-target affinity (DTA) prediction is widely used to prioritize candidate compounds before costly experimental screening. DTA models are often compared under a single data split, even though deployment may require extrapolation to new chemical series, new protein targets, or both. We ask whether the distribution shift used for evaluation changes which architecture appears best. We curate 718,800 unique drug-protein pairs from the ChEMBL and BindingDB datasets. We compare a Morgan-fingerprint + protein-CNN baseline with 12 controlled architectures that combine four drug representations with three ESM-2 interaction modes. Mean validation RMSE increases from 0.950 and 0.945 under scaffold and fingerprint-cluster OOD to 1.299 and 1.321 under protein-cluster and dual OOD. Model rankings are similar across the two chemical shifts (tau = 0.79), but agreement with scaffold OOD falls under protein OOD (tau = 0.39) and reverses under dual OOD (tau = -0.55). Held-out evaluation, repeated seeds, group-aware bootstrap analysis, and a size-matched control support the same conclusion: architecture selection depends on the form of extrapolation, not only on average error or training-set size. DTA benchmarks should therefore match the chemical and target shifts expected at deployment.
Minjae Chung, Clara Li, Malar Paavai Muthukumaran +8