TL;DR
This comprehensive review charts the evolution of protein-ligand binding affinity prediction from rigid, physics-based scoring equations to flexible, data-driven deep learning models. It highlights how these computational tools are transitioning from isolated molecular simulations into systems-level frameworks like AI Virtual Cells (AIVCs), paving the way for animal-free, personalized drug discovery.
What Problem Does This Paper Address?
At the heart of drug discovery and structural biology lies a crucial question: how tightly does a small molecule (a ligand) bind to a target protein? This "binding affinity" determines whether a molecule can successfully modulate a protein's function, making it a cornerstone for designing therapeutics, engineering enzymes, and mapping biological pathways.
Historically, determining binding affinity has relied on slow, expensive laboratory experiments. While computational methods offer a scalable alternative, they face massive hurdles. Our understanding of chemical physics is incomplete, proteins are highly flexible rather than rigid shapes, and existing experimental datasets are riddled with hidden biases. This paper provides a sweeping overview of how the field is tackling these challenges by shifting from conventional physics-based equations to modern machine learning (ML) and deep learning (DL) architectures.
Methods
The paper categorizes binding affinity prediction methodologies into two main eras:
- Conventional Methods: These rely on human-engineered physics and chemistry principles. They include physics-based methods (calculating free energy changes via molecular dynamics), empirical methods (using statistical regression to weight chemical terms like hydrophobicity and entropy, as seen in AutoDock Vina), and knowledge-based potentials (deriving statistical preferences from known atom-pair distances in structural databases).
- Machine Learning-Based Methods: The modern paradigm splits into two distinct approaches:
- Interaction-free approaches: These use separate neural networks to embed the protein (often using sequence-based Protein Language Models like ESM) and the ligand (using 1D SMILES strings or 2D graphs) without explicitly modeling their 3D spatial interaction. The latent space implicitly captures the binding.
- Interaction-based approaches: These explicitly model the 3D structure of the complex using 3D voxel grids (processed by 3D CNNs) or rotation-invariant 3D graphs and point clouds (processed by GNNs) focused directly on the binding pocket.
Datasets
The authors review several foundational datasets used to train and benchmark these models:
- PDBbind & CASF: The de facto standard benchmarks containing curated 3D structures and experimental affinities. CASF serves as the core testing set.
- Binding MOAD & BindingDB: Large repositories of high-quality crystal structures and protein-compound pairs.
- DUD-E: The gold standard for virtual screening, featuring target proteins paired with active ligands and physically similar but topologically distinct "decoy" molecules to test model discrimination.
- BioLiP, PDSP Ki, and KIBA: Datasets focusing on biologically relevant ligands, psychoactive drugs, and kinase inhibitors, respectively.
- PLAS-20K: A recently released dataset incorporating molecular dynamics simulations to capture dynamic molecular features rather than just static snapshots.
Key Insights and Ideas
- The "Scoring vs. Docking" Paradox: While modern deep learning models excel at scoring power (predicting the exact binding affinity of a known complex), they often perform poorly at docking power (identifying the correct 3D binding pose out of multiple decoys). Conventional physics-based methods still tend to outperform ML models in finding the right geometric pose.
- Hidden Dataset Biases: Popular benchmarks like DUD-E suffer from data leakage. Models often achieve high accuracy simply by memorizing ligand topologies or protein types rather than learning true biophysical interaction rules, causing performance to drop sharply on novel chemical structures.
- The Shift to Dynamics: Static 3D snapshots ignore protein flexibility and entropic effects. The field is rapidly moving toward dynamic-aware datasets (like PLAS-20K) and generative diffusion models to simulate the "ensemble" nature of binding.
- The Rise of AI Virtual Cells (AIVCs): Spurred by the FDA's phase-out of mandatory animal testing, binding affinity prediction is evolving from isolated molecular calculations into a component of full cellular simulations. AIVCs will allow researchers to evaluate binding specificity, metabolic competition, and off-target risks across an entire simulated proteome.
Key Results
The paper synthesizes performance trends across decades of research:
- Conventional scoring functions show a modest Pearson correlation ($R$) of 0.4 to 0.6 with experimental binding affinities, but they successfully predict the correct binding pose 60% to 80% of the time.
- Machine learning and deep learning models significantly boost scoring power on standard benchmarks, but their generalizability drops when tested on out-of-distribution or proprietary industrial datasets.
- The paper maps out the exact architectures and data representations (sequences, 2D graphs, 3D voxels, point clouds) of over 30 major ML models developed since 2018, providing a definitive taxonomy of the current state-of-the-art.
Limitations and Future Work
The authors identify several critical gaps that define the next frontier of the field:
- Data Scarcity for Non-Binders: Current datasets are heavily biased toward successful, high-affinity complexes. To train truly robust models, the community needs more data on low-affinity or completely failed bindings.
- Rigid Protein Assumptions: Most models still treat proteins as rigid shapes, failing to capture "induced fit" mechanisms where a protein changes shape upon ligand entry.
- Flawed Evaluation Metrics: Standard metrics like RMSD for docking or Pearson correlation for scoring don't fully capture real-world utility in drug discovery. More nuanced, application-specific metrics are required.
- Multimodal and Systems Integration: Future work must seamlessly combine 1D sequences, 3D structures, and cellular multi-omics data to predict how a drug behaves not just in a test tube, but inside a dynamic cellular environment.
Summary updated