Paper ReviewAI & Machine Learning

AlphaFold at Five Years: Boltz-2 and the Push Toward Binding Affinity

Five years after AlphaFold2 solved protein structure prediction, the field's frontier has shifted to biomolecular complexes and binding affinity โ€” where AlphaFold3, Boltz-1 (open-source), and Boltz-2 represent successive steps toward the drug discovery application that structural biology always promised.

By ORAA Research
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

When AlphaFold2 debuted at CASP14 in 2020, it effectively solved the single-chain protein structure prediction problem. Five years on, the research community has absorbed the implications and shifted attention to the next bottleneck: predicting not just what proteins look like, but how they interact with other molecules โ€” and how tightly.

This progression matters because structure prediction, remarkable as it is, was always a waypoint. The application that pharma and biotech actually need is binding affinity prediction: given a drug candidate and a protein target, how strongly do they bind? That question remains only partially answered, but the trajectory from AlphaFold2 through AlphaFold3 to Boltz-1 and Boltz-2 shows a field converging on it.

The Research Landscape

AlphaFold3: From Single Chains to Complexes

Abramson et al. (2024) introduced AlphaFold3 with a substantially updated architecture based on diffusion rather than the structure module approach of AlphaFold2. The key capability extension: AF3 predicts complexes involving proteins, nucleic acids, small molecules, and ions โ€” moving from single-chain folding to all-atom biomolecular structure prediction. On CASP16 benchmarks, AF3 demonstrated notable improvements in protein-ligand and protein-nucleic acid complex prediction compared to its predecessor.

However, Abriata and Dal Peraro (2025), analyzing practical outcomes from CASP16, note important nuances. Protein monomer and domain prediction is "largely solved, with barely any space for further improvements at the backbone level." But complex prediction โ€” especially for protein-ligand interfaces โ€” shows substantial room for improvement. The accuracy for ligand pose prediction, while better than previous methods, does not approach the accuracy achieved for protein backbone prediction.

Boltz-1: Open-Source Democratization

A significant development was the release of Boltz-1 (Wohlwend et al., 2024), an open-source model that achieves AlphaFold3-level accuracy on biomolecular complex structure prediction. Boltz-1 introduced innovations in model architecture and speed optimization while making the technology freely available โ€” addressing a major community concern about AlphaFold3's restricted access.

Boltz-1's importance is as much political as technical. By providing open-source access to comparable capabilities, it enabled the research community to experiment with, extend, and build upon the technology in ways that proprietary access alone could not support.

Boltz-2: Binding Affinity as a First-Class Objective

Passaro, Corso, and Wohlwend (2025) introduce Boltz-2 with an explicit focus on the gap between structure prediction and binding affinity prediction. Their core argument: accurately predicting the 3D structure of a protein-ligand complex is necessary but not sufficient for predicting how tightly the drug binds โ€” the thermodynamic quantity that determines pharmacological activity.

Boltz-2 extends the Boltz-1 architecture with modules designed to predict binding affinity alongside structure. Early results show improved correlation between predicted and experimental binding affinities compared to using structural prediction models alone. The paper has already accumulated substantial citations, reflecting the field's interest in this specific capability.

The Broader Affinity Prediction Landscape

Wang, Wu, and Wang (2024) provide context through a comprehensive review of structure-based deep learning models for protein-ligand binding affinity prediction. Their analysis identifies a critical gap: most existing methods achieve moderate correlation with experimental affinities on standard benchmarks (PDBbind) but struggle with generalization to novel protein families and out-of-distribution targets. The "benchmark success versus real-world failure" problem is a recurring theme.

Critical Analysis

<
ClaimEvidenceVerdict
AlphaFold3 extends accurate prediction to biomolecular complexesCASP16 results show clear improvements over AF2 for complexesโœ… Supported โ€” with the caveat that ligand pose accuracy lags behind backbone accuracy
Boltz-1 matches AlphaFold3 accuracy while being open-sourceBenchmarks on CASP targets and PDB test sets show comparable performanceโœ… Supported โ€” an important democratization milestone
Boltz-2 advances binding affinity predictionEarly results show improved affinity correlation; community adoption is rapidโš ๏ธ Promising โ€” independent benchmarking on diverse targets needed
Structure prediction is now sufficient for drug discoveryPredicted structures often lack the sub-angstrom accuracy needed for affinity estimationโŒ Not yet โ€” structure prediction is necessary but not sufficient
The field is converging on a complete computational drug design pipelineEach generation addresses a specific bottleneck (structure โ†’ complex โ†’ affinity)โš ๏ธ Directionally correct โ€” but significant gaps remain (dynamics, entropy, solvation)

What Structure Prediction Cannot Yet Capture

Even perfect static structure prediction would not solve the binding affinity problem completely. Binding is a thermodynamic process involving:

Conformational dynamics: Proteins are not rigid. Binding often involves conformational changes (induced fit) that alter the binding interface. Current models predict a single (or few) static structures.

Entropic contributions: Binding entropy โ€” the loss of rotational and translational freedom upon complex formation, the reorganization of solvent molecules โ€” contributes substantially to binding free energy and is not captured by static structure prediction.

Water-mediated interactions: Bridging water molecules at the binding interface contribute to binding affinity in ways that are difficult to predict from structure alone.

Protonation states: The protonation state of amino acids at the binding interface affects electrostatic interactions and can change upon binding โ€” a subtlety that structure prediction models do not explicitly model.

The CASP16 Reality Check

Abriata and Dal Peraro (2025) provide a measured assessment from CASP16. The competition confirmed that for protein monomers and domains, prediction is effectively solved. For protein-protein interfaces, accuracy has improved but remains below the threshold needed for confident binding energy estimation. For protein-ligand complexes, accuracy varies substantially by target class, with some predictions achieving excellent poses and others placing the ligand in fundamentally wrong orientations.

Open Questions

  • Dynamics integration: Can molecular dynamics simulations be efficiently coupled with deep learning structure prediction to capture conformational ensembles?
  • Prospective drug discovery validation: Most benchmarks are retrospective. Can Boltz-2 or similar models prospectively identify novel drug candidates that are validated experimentally?
  • Covalent and allosteric binding: Current models focus on orthosteric binding. Covalent inhibitors and allosteric modulators involve distinct mechanisms that may require specialized architectures.
  • Training data limitations: Experimental binding affinity data (PDBbind, BindingDB) is biased toward druggable targets and high-affinity binders. Models trained on these data may underperform on novel target classes.
  • Integration with medicinal chemistry: Structure prediction must interface with the practical constraints of drug design โ€” synthetic accessibility, selectivity, pharmacokinetics โ€” which are not captured by binding affinity alone.
  • Closing

    The five-year trajectory from AlphaFold2 to Boltz-2 traces a clear research agenda: from single-chain structure to biomolecular complexes to binding affinity. Each step addresses a specific limitation of the previous generation, and each step moves closer to the drug discovery application that motivates the field. Boltz-2's explicit focus on binding affinity, combined with Boltz-1's open-source accessibility, represents the current frontier. But the gap between predicting a plausible structure and predicting a reliable binding affinity remains substantial, limited by dynamics, entropy, and solvation effects that static deep learning models do not yet capture.

    References (5)

    Abramson, J., Adler, J., & Dunger, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493โ€“500.
    Wohlwend, J., Corso, G., & Passaro, S. et al. (2024). Boltz-1: Democratizing biomolecular interaction modeling. bioRxiv.
    Passaro, S., Corso, G., & Wohlwend, J. et al. (2025). Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv.
    Abriata, L., & Dal Peraro, M. (2025). Practical outcomes from CASP16 for users in need of biomolecular structure prediction. Proteins: Structure, Function, and Bioinformatics.
    Wang, D. D., Wu, W., & Wang, R. (2024). Structure-based, deep-learning models for protein-ligand binding affinity prediction. Journal of Cheminformatics, 16, Article 2.

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 7 keywords โ†’