Research map / Probabilistic and scientific modeling / Molecules and proteins
M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes M²N, a cross-scale 3D network that models drugs (functional-group graph to atom graph) and proteins (residue graph to atom graph) coarse-to-fine, fused by an adaptive module, for drug-target affinity prediction on DAVIS and KIBA.
Problem
Drug-target affinity (DTA) prediction methods operate at a single scale (residue for proteins, atom for drugs), missing cross-scale information such as atom-level protein detail and functional-group-level drug information.
Approach
Builds progressive macro-to-micro 3D graphs: functional-group graphs for drugs and residue graphs for proteins are refined into atom-level graph transformers using knowledge transfer from coarser scales, with an adaptive fusion module and auxiliary multi-scale MSE losses.
Key results
- On DAVIS under S1, M²N reaches MSE 0.316, PCC 0.639, CI 0.816, outperforming HiSIFDTA (MSE 0.359, PCC 0.608, CI 0.803).
- Under S3 (both novel) on DAVIS, M²N achieves PCC 0.345 vs 0.258 for the best baseline HiSIFDTA.
- Ablation on DAVIS S1 shows removing the adaptive fusion module raises MSE from 0.316 to 0.325 and removing target atom-scale features raises it to 0.354.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add drug FG-scale representations (ablation) | M2N · DAVIS S1 | PCC: 0.615 -> 0.639 | Helped |
| add target residue-scale representations (ablation) | M2N · DAVIS S1 | PCC: 0.618 -> 0.639 | Helped |
| add drug atom-scale representations (ablation) | M2N · DAVIS S1 | PCC: 0.611 -> 0.639 | Helped |
| add target atom-scale representations (ablation) | M2N · DAVIS S1 | PCC: 0.608 -> 0.639 | Helped |
| use adaptive fusion instead of direct concatenation (ablation) | M2N · DAVIS S1 | PCC: 0.621 -> 0.639 | Helped |
| progressively integrate drug FG features into atom representation (ablation) | M2N · DAVIS S1 | PCC: 0.623 -> 0.639 | Helped |
| progressively integrate target residue features into atom representation (ablation) | M2N · DAVIS S1 | PCC: 0.620 -> 0.639 | Helped |
| increase auxiliary-task balance coefficients λ1 and λ2 | M2N · DAVIS S1 | PCC | Hurt |
| vary the number of graph neighbors k1 and k2 | M2N · DAVIS S1 | PCC | Mixed |
Surprising findings
- Sequence-based MolTrans achieves much better KIBA MSE than structure-based GraphDTA, and GraphDTA performs poorly on KIBA PCC (0.269 under S1).
Limitations
- Evaluated only on two DTA benchmarks (DAVIS and KIBA).
- Requires 3D structural data gathered for each target and drug, which the original datasets lack.
- Text reports inconsistent claims (e.g., refers to HeHiDTAG; some hyperparameter ranges contain typos).
Experimental setup
Datasets: DAVIS KIBA
Benchmarks: DAVIS KIBA
Metrics: MSE PCC CI
Baselines: GraphDTA MolTrans DeepGLSTM FusionDTA MFRDTA DGraphDTA MSFDTA HiSIFDTA
Builds on: Graph Transformer ESM GraphDTA
Closest papers by meaning
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · similarity 0.88)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.87)
- 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.87)
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.87)
- Binding-Adaptive Diffusion Models for Structure-Based Drug Design (AAAI 2024 · similarity 0.87)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.