PaperlayerView exact revision on arXiv

Source-linked AI summary

Chemical and geometric representation fidelity improves drug--target affinity prediction

Yixiao Li, Yining Qian, Yefan Chen, Zenghui Chen, Jiayue Sun, Yuhai Zhao, Cheng Tan, An-Yang Lu

arXiv:2609.13230v1q-bio.BMcs.AIcs.LGstat.ML

TL;DR

Drug–target affinity prediction can help prioritize compounds, but it depends on representing binding-relevant molecular information. ReGeoDTA preserves chemical and geometric information before interaction prediction; its benefits appear across benchmarks and diverse architectures, including improved early enrichment under a top-10% screening budget.

  • Problem

    DTA models may compress or obscure affinity-relevant distinctions during drug and protein representation construction, despite the need to capture information relevant to binding.

    Source: §I · p. 1, §III · p. 16

  • Method

    ReGeoDTA preserves heterogeneous atomic properties and continuous relative protein geometry before drug–target interaction modelling.

    Source: §I · p. 2, §III · p. 16

  • Results

    ReGeoDTA’s representation-preserving strategies transferred to alternative DTA backbones, and the gains included stronger global affinity estimation and improved early enrichment under a top-10% screening budget.

    Source: §III · p. 16

  • Takeaways & Limitations

    Preserving chemical and geometric information before interaction prediction is a useful complement to increasingly expressive DTA architectures.

    Source: §III · p. 16

    The EGFR experiment was retrospective and target-held-out, not a prospective screen of a fully external compound library.

    Source: §III · p. 16

Abstract

from arXiv · show

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.

I. INTRODUCTION

DTA models have added richer molecular and protein inputs, but affinity-relevant distinctions may still be compressed during representation construction. ReGeoDTA preserves chemical heterogeneity and protein geometry, and improves results across three benchmarks.

I. INTRODUCTION

  • Experimental affinity measurements are costly and difficult to scale across the chemical and protein spaces relevant to drug discovery.DTA prediction offers a computational way to estimate interaction strength and prioritize compounds before experimental validation.Source: §I · p. 1
  • Drug encoders can compress distinct atomic attributes, while contact-map thresholds can erase continuous protein-distance variation relevant to molecular recognition.For example, a contact-map cutoff can treat 3.0 Å and 7.9 Å residue distances alike, but assign different labels to 7.9 Å and 8.1 Å.Source: §I · pp. 1–2
  • ReGeoDTA preserves context-dependent chemical heterogeneity and continuous protein geometry before integrating representations for affinity prediction.The framework focuses on representation fidelity rather than primarily increasing interaction-predictor complexity.Source: §I · p. 2
  • The study tests representation preservation across benchmarks and cold-start settings, using ablation, degradation, predictor-decoupling, and cross-architecture analyses.It also examines attention and drug embeddings and retrospectively screens EGFR-held-out compounds.Source: §I · pp. 2–3

A. ReGeoDTA improves affinity prediction

  • Across Davis, KIBA, and BindingDB, ReGeoDTA achieved the highest CI and lowest MSE among compared methods, with the best result in 10 of 12 dataset–metric combinations.Its advantage spanned pairwise affinity ranking and numerical prediction accuracy.Source: A. ReGeoDTA improves affinity prediction · p. 3

B. ReGeoDTA generalizes to unseen drugs and targets

Cold-start tests assess performance on unseen drugs, targets, or both, while randomization tests disrupt drug–target correspondence, labels, or protein sequence information. ReGeoDTA performs strongly in the cold-start settings, and its performance degrades under all tested randomizations.

B. ReGeoDTA generalizes to unseen drugs and targets

  • On KIBA, ReGeoDTA achieved the best MSE, CI, and R2 across all three cold-start settings; on Davis, it led all reported metrics in cold-target and all-cold settings.In Davis cold-drug testing, it had the lowest MSE and highest R2, while NHGNN-DTA had a slightly higher CI.Source: B. ReGeoDTA generalizes to unseen… · p. 5

C. Randomization tests confirm dependence on intact drug–target relationships

  • Randomizing affinity labels caused the strongest degradation: across Davis, KIBA, and BindingDB, CI fell to 0.536, 0.504, and 0.504, while R2 fell to 0.004, 0.003, and 0.001.AUPR also fell from 0.769, 0.850, and 0.871 to 0.106, 0.209, and 0.217, respectively.Source: C. Randomization tests confirm… · p. 5, Figure 2 · p. 6
  • Drug, protein, and sequence randomization also impaired performance, with MSE increasing across all three datasets.Together, the corruptions disrupted drug–target pairing, sample–label correspondence, or protein sequence information.Source: C. Randomization tests confirm… · pp. 5–6
  • The degradation under all tested corruptions indicates that benchmark performance depended on intact drug–target relationships rather than random associations or spurious dataset-specific correlations.Source: C. Randomization tests confirm… · p. 6

D. Performance gains arise from preserving representations

The analyses examine whether ReGeoDTA benefits from preserving drug- and protein-side information before interaction prediction, rather than from a single component or downstream prediction head.

D. Performance gains arise from preserving representations

1) Component ablation reveals contributions of drug and protein representation modules:

ReGeoDTA’s representation components contribute to affinity prediction, and cumulative representation degradation worsens results across three datasets. Enhanced initial representations also outperform weaker GIN-based representations across matched predictor comparisons.

1) Component ablation reveals contributions of drug and protein representation modules:

  • Atom-symbol simplification raised MSE from 0.186 to 0.212 on Davis and 0.409 to 0.431 on BindingDB; removing PAPE modestly reduced R2 across all three datasets.The ablations also found that replacing GMFF with one-hot encoding or removing adaptive gating impaired performance.Source: 1) Component ablation reveals… · pp. 6–7, Table 3 · p. 7

2) Cumulative representation degradation reduces overall DTA performance:

  • Final degradation raised MSE to 0.229, 0.160, and 0.500 from 0.186, 0.127, and 0.409 on Davis, KIBA, and BindingDB, respectively.The final setting also reduced R2 on all three datasets; intermediate changes were not strictly monotonic for every metric.Source: 2) Cumulative representation… · pp. 7–8, Figure 3 · p. 8

3) More complex predictors do not compensate for weaker initial representations:

  • Enhanced-representation linear readouts achieved lower MSE than the strongest GIN-based predictor on Davis (0.186 vs 0.218), KIBA (0.127 vs 0.152), and BindingDB (0.409 vs 0.486).Once enhanced representations were used, more complex predictors provided limited additional benefit.Source: 3) More complex predictors do not… · pp. 8–9, Table 4 · p. 9

E. Representation-preserving strategy enhances DTA prediction across model architectures

GMFF and PAPE improved prediction across diverse DTA backbones when the representation component was modified while retaining the original backbone and prediction head. On Davis, relative geometric positional encoding outperformed alternative positional strategies and aligned attention more strongly with residue proximity.

E. Representation-preserving strategy enhances DTA prediction across model architectures

  • The transfer tests retained each original backbone and prediction head, changing only the representation component needed to incorporate GMFF or PAPE.Enhanced and original variants used the same data splits and training settings on Davis, KIBA, and BindingDB.Source: E. Representation-preserving strategy… · p. 9

1) GMFF improves affinity prediction across diverse drug-side architectures:

2) PAPE provides consistent gains across DTA backbones:

F. Relative geometric encoding aligns protein attention with residue-level geometry

  • On Davis, PAPE achieved the lowest MSE (0.186) and highest CI (0.906) and AUPR (0.807) among seven positional strategies.The position-free baseline had MSE 0.195, CI 0.900, and AUPR 0.798; other positional strategies did not consistently improve prediction.Source: Table 5 · p. 11, F. Relative geometric encoding aligns… · p. 11
  • PAPE’s attention–distance correlation was -0.3919 ± 0.0072, versus -0.1948 ± 0.0082 for the position-free baseline, indicating stronger alignment with spatial proximity.PPE-based absolute encoding had higher attention sparsity than PAPE but worse MSE and CI, so concentration alone did not explain the predictive advantage.Source: F. Relative geometric encoding aligns… · p. 11

G. PAPE-derived attention aligns with known binding-site organization in a representative protein–ligand complex

In the ABL1–Nilotinib complex, PAPE combines semantic and geometric attention into a more selective pattern around annotated binding-site relationships. The fused map also retains long-range, cross-region associations, while comparisons with other encoders are qualitative rather than rankings of interpretability.

G. PAPE-derived attention aligns with known binding-site organization in a representative protein–ligand complex

  • Fig. 6 compares semantic, distance-aware, orientation-aware, and fused maps, then relates fused attention to 22 annotated residues and component complementarity.Image 5 marks high fused attention at ≥0.35 in blue and lower attention in red; Image 6 distinguishes positional from semantic complementarity.Source: Figure 6 · p. 12, G. PAPE-derived attention aligns with… · p. 12
  • PAPE’s fused attention became more selective around annotated binding-site relationships by combining semantic, distance-aware, and orientation-aware signals.Semantic attention was diffuse, proximity alone did not distinguish binding-related pairs from generic neighborhoods, and orientation-aware attention remained distributed; fusion reduced diffuse background associations.Source: G. PAPE-derived attention aligns with… · pp. 12–13
  • The fused map retained high-attention cross-region associations, including the sequence-distant M290–H361 pair, compatible with three-dimensional binding-region organization.This observation does not establish a direct physical interaction between the residues.Source: G. PAPE-derived attention aligns with… · pp. 13–14
  • Compared on the same ABL1 complex, CNN maps followed sequence-local receptive fields and contact-map GNN propagation followed threshold-defined edges, unlike PAPE’s selective residue-pair organization.The comparison is qualitative because attention, convolutional interaction, and graph-propagation maps represent different internal quantities.Source: G. PAPE-derived attention aligns with… · p. 14

H. GMFF-derived drug embeddings better reflect multi-target affinity profiles

GMFF-derived drug embeddings more closely tracked multi-target affinity-profile similarity than one-hot embeddings. In a retrospective EGFR-held-out screen, ReGeoDTA improved affinity estimation and early enrichment, while its broader results support representation preservation as complementary to diverse predictors.

H. GMFF-derived drug embeddings better reflect multi-target affinity profiles

  • Spearman similarity alignment rose from 0.3801 with one-hot embeddings to 0.4677 with GMFF, and Pearson alignment rose from 0.3636 to 0.5093 on Davis.Bootstrap means showed the same pattern: Spearman increased from 0.3771 ± 0.0221 to 0.4652 ± 0.0198, and Pearson from 0.3613 ± 0.0218 to 0.5071 ± 0.0190.Source: H. GMFF-derived drug embeddings better… · p. 15

I. ReGeoDTA improves real discovery

  • On held-out EGFR, ReGeoDTA reduced MSE from 1.421 to 1.244 and MAE from 0.976 to 0.826 versus DeepDTAGen, while Pearson correlation rose from 0.604 to 0.708.The comparison also reports CI increasing from 0.735 to 0.748 and Spearman correlation from 0.643 to 0.683.Source: §I · p. 15
  • At the fixed pIC50 threshold, EGFR recall rose from 0.452 to 0.677 versus DeepDTAGen, while precision fell from 0.636 to 0.488, a recall–precision trade-off.ReGeoDTA identified 21 of 31 experimentally high-activity compounds, compared with 14 for DeepDTAGen.Source: §I · p. 15
  • Among the top 11 held-out EGFR candidates, ReGeoDTA retrieved nine high-activity compounds versus six for DeepDTAGen, raising the hit rate from 54.5% to 81.8%.The enrichment factor increased from 1.918 to 2.877 under the top-10% screening budget.Source: §I · pp. 15–16

III. DISCUSSION

  • Representation degradation impaired prediction, whereas more expressive interaction modules did not compensate for weaker initial representations; gains from GMFF and PAPE also transferred across encoder families.GMFF produced larger gains by preserving chemically distinct atomic attributes; PAPE produced smaller but more consistent improvements through relative protein geometry.Source: §III · p. 16
  • The EGFR-held-out analysis supports improved prioritization within available labelled data, but does not demonstrate experimental discovery of previously uncharacterized binders.The screen was retrospective rather than a prospective screen of a fully external compound library.Source: §III · p. 16

IV. METHODS

The study evaluates ReGeoDTA on Davis, KIBA, and BindingDB benchmarks, using established split protocols for benchmark comparability and repeated random splits for other training-based analyses.

A. Datasets

  • The three benchmarks are Davis, KIBA, and BindingDB, with prediction labels respectively based on pKd, KIBA scores, and pIC50.Davis measurements were converted from Kd to pKd, and BindingDB measurements from IC50 to pIC50.Source: A. Datasets · p. 16

B. Experimental setup

  • Benchmark comparisons used six folds, holding one fold out for testing and using the other five for cross-validation, following DeepDTA, GraphDTA, and related baselines.Cold-start experiments followed the data partitions and evaluation protocol reported in NHGNN-DTA.Source: B. Experimental setup · p. 16
  • Component ablation, representation degradation, predictor decoupling, positional-encoding comparison, and cross-architecture transfer used repeated random splits to reduce dependence on one partition.Each dataset was split 4:1 and the process repeated five times with different random seeds; average performance was reported.Source: B. Experimental setup · p. 17

C. Overall architecture

ReGeoDTA combines parallel drug and protein encoders to produce an affinity score. Its protein representation uses ProtTrans embeddings and predicted structural coordinates to incorporate continuous relative geometry.

1) Target protein encoding:

  • ProtTrans residue embeddings feed parallel CNN and PAPE branches; PAPE uses full-length features for queries and down-sampled features for keys and values.Gated fusion and max pooling combine the PAPE and CNN outputs into the final position-aware protein representation.Source: C. Overall architecture · p. 17

C. Overall architecture

  • The final drug and protein representations are concatenated and passed through a feedforward module to produce the affinity score, with mean squared error as the prediction loss.The architecture comprises five functional blocks, including GMFF for heterogeneous atomic attributes and PAPE for relative geometric encoding.Source: C. Overall architecture · p. 17

2) Protein 3D structure acquisition:

  • ESMFold-predicted Cα coordinates are down-sampled to one-quarter length and used to construct relative distance and orientation matrices for PAPE.The matrices encode continuous relative geometry; average pooling with kernel size 8 and stride 4 reduces geometric anchors and computational cost.Source: 2) Protein 3D structure acquisition · pp. 17–18

3) Gated Multi-Feature Fusion (GMFF) for drug molecular features:

GMFF integrates heterogeneous atomic attributes, including context-dependent numerical features, while PAPE combines protein sequence features with relative geometric biases. Both modules contribute to the drug and protein representations used by ReGeoDTA.

3) Gated Multi-Feature Fusion (GMFF) for drug molecular features:

  • GMFF distinguishes categorical atom symbol and aromaticity features from numerical degree, hydrogen-count, and implicit-valence features.It adds an aromaticity-conditioned augmentation so numerical attributes can differ between aromatic and non-aromatic atoms, reflecting distinct chemical roles such as pyridinic and pyrrolic nitrogen.Source: 3) Gated Multi-Feature Fusion (GMFF)… · p. 18
  • GMFF projects the augmented numerical inputs and combines them with atom-symbol and aromaticity embeddings, then gates feature components with different scales and embedding forms.The augmentation is concatenated with the original numerical attributes before projection.Source: 3) Gated Multi-Feature Fusion (GMFF)… · p. 18

1) Position-Aware Protein Encoder (PAPE):

  • PAPE adds distance- and orientation-aware geometric biases to content-based residue similarity, then gated fusion combines its output with the CNN branch before max pooling.Keys and values derive from down-sampled features while queries retain full-length resolution; down-sampling aligns the similarity and geometric matrices.Source: 1) Position-Aware Protein Encoder (PAPE) · pp. 18–19

2) Drug graph network:

GMFF-derived drug node features are encoded through parallel GCN and GAT–GCN branches. Pooling the outputs of both branches forms the final drug representation.

2) Drug graph network:

  • GMFF-derived drug node features are processed by parallel GCN and GAT–GCN branches.The GCN branch has three layers, while the GAT–GCN branch has one GAT layer followed by one GCN layer.Source: 2) Drug graph network · p. 19
  • The GAT layer uses neighboring nodes, including the node itself, and a trainable attention parameter vector.The branch then applies a GCN layer to the GAT output.Source: 2) Drug graph network · p. 19
  • Average and max pooling are applied to each branch’s node representations, and the pooled outputs are concatenated into the final drug representation.The resulting representation is denoted zdrug.Source: 2) Drug graph network · p. 19

F. Positional encoding variants

The positional-encoding variants compare PAPE with position-free and alternative absolute or relative encodings. The comparison is designed to test whether PAPE’s benefit comes from positional information generally or its relative geometric encoding.

F. Positional encoding variants

  • The experiment compares PAPE with a position-free baseline and five positional encoding variants, including sinusoidal, coordinate-based, RBF-based, and PPE-based schemes.The variants assess whether PAPE benefits from positional information generally or relative geometric encoding specifically.Source: F. Positional encoding variants · p. 19

1) Sinusoidal absolute position encoding:

2) 3D coordinate-based absolute position encoding:

The section describes ways to inject 3D protein coordinates into residue representations, using absolute encodings and distance-based relative encoding. PPE uses the protein centroid as its coordinate origin.

2) 3D coordinate-based absolute position encoding:

3) 3D coordinate-based feature concatenation:

  • For coordinate-based feature concatenation, Cα coordinates are projected to 32 dimensions and concatenated with residue features before encoding.The coordinate-derived feature dimension is kept smaller than the original residue-feature dimension.Source: 3) 3D coordinate-based feature… · p. 19

4) RBF-based relative position encoding:

  • RBF encoding maps continuous distances to an N-dimensional vector whose responses peak near uniformly sampled distance centers; width w controls response sensitivity.The input distance is between a residue and a down-sampled coordinate anchor.Source: 4) RBF-based relative position encoding · pp. 19–20

5) PPE-based absolute position encoding:

  • PPE assigns each residue a centroid-based positional value and is treated as absolute encoding, not a pairwise relative positional bias.The coordinate origin is the protein centroid; PPE was originally introduced in medical image processing.Source: 5) PPE-based absolute position encoding · p. 20
Loading 2609.13230v1…