Why co-folding models are here to stay
Co‑folding models, like AlphaFold 3, Boltz‑2, and OpenFold3, can predict the structure of protein–ligand and protein–protein complexes. While these models perform strongly on public benchmarks, their accuracy often decreases on novel targets that are underrepresented in the training data. The Apheris Co‑Folding Application is a secure, locally hosted tool for using co‑folding models on sensitive proprietary data. By running Boltz-2 and OpenFold3 on in‑house targets, researchers can assess model accuracy in their specific domain and easily integrate these models into their wider workflows. The product provides visualization tools, streamlined data preparation support, and generates auditable, reproducible records for use in regulated environments.
Why co-folding models are here to stay
Co-folding models predict protein–ligand and protein–protein complex structures in a single forward pass by leveraging statistical patterns learned from large structural datasets. Traditional physics‑based approaches, such as molecular docking or molecular dynamics, instead simulate molecular interactions using equations derived from first principles (i.e., physics) and parameterized force fields fitted to experimental and quantum‑mechanical data. Co-folding offers a statistical, template‑free route to complex structures at orders of magnitude faster than physics‑based sampling and with less required user expertise. They are particularly useful for:
Proposing plausible binding poses for ligands or interacting proteins, especially when experimental structures are unavailable.
Generating structural hypotheses to rationalize structure–activity relationships, helping, for example, guide lead optimization in early stages when crystallographic data is unavailable.
Working across diverse modalities such as antibody–antigen interactions, de novo binder design, and other complex biologics. Co-folding models can suggest realistic conformations and interfaces prior to experimental validation.
The current generation of foundation models follows a rapid trajectory:
| Model | License | Key capability | Reference |
|---|---|---|---|
| AlphaFold3 | Research‑only, no commercial use | First unified diffusion model for proteins, nucleic acids & small molecules | https://shorturl.at/j5Y88 |
| Boltz-2 | Permissive Open-Source Software | Opens diffusion architecture, adds explicit affinity head (Boltz‑2) | https://shorturl.at/H2NEU |
| OpenFold3 (OF3) | Full OSS | First community reproduction that reaches or exceeds AF3 accuracy (pre‑print) | Open community reports |
A critical look at model performance
While co-folding methods are rapidly improving and gaining traction as indispensable tools in pharmaceutical, biological, and chemical research, the current generation still has important limitations. Most of these arise from the training data, as current models are trained and evaluated almost exclusively on public Protein Data Bank (PDB) structures. Peter Škrinjar and colleagues examined the consequences of this limited data diversity in their recent study: Have protein‑ligand co‑folding methods moved beyond memorization? The authors constructed an independent benchmark (“Runs N’ Poses”) of 2,600 structures published after the models’ training cut‑offs. They evaluated model performance using stringent success criteria (LDDT‑PLI > 0.8 and ligand‑pose RMSD < 2 Å). They clustered similar structures and analyzed those with few examples in the training data (< 100 examples). They observed an almost linear drop in success rate as training set coverage declined, reaching as low as ~20% for the sparsest bin (Fig. 1 in the paper). The study shows that current co‑folding models make accurate predictions for cases similar to their training data (interpolation) but struggle with distant chemotypes and novel targets (extrapolation).
It is critical to assess the applicability domain of co-folding models and customize the models to your own research tasks
To fully leverage the strengths of co‑folding models while mitigating their limitations, researchers should:
Benchmark models on in-house data: Different models excel at different tasks. Rapid deployment and side-by-side comparison on proprietary targets help identify the best fit for a given use case, especially since public benchmarks reflect training distributions that often differ significantly from industrial data.
Preserve IP and compliance: Proprietary ligands and targets are often absent from the PDB and cannot be shared with public inference servers due to confidentiality and regulatory constraints. Local model deployment is essential to retain control over sensitive assets.
Fine-tune on proprietary data: Adapting publicly trained models to internal datasets improves accuracy for your specific chemical space and target classes, especially where proprietary ligands or sequences diverge from public training data.
Ensure modular integration and traceability: Tooling should be easy to integrate into existing pipelines, version-controlled, and auditable to support reproducible research and compliance in regulated environments.
Consider joining federated data networks: To increase the diversity and size of the training data and hence increase the model's applicability domain and generalizability - an example of this is the AI Structural Biology network
Industrial-Grade Deployment with the Apheris Co‑Folding Application
The Apheris Co‑Folding Application is a self‑contained, on‑premise toolkit that lets computational chemists and biologists deploy and use the latest open models, such as Boltz-2 and OpenFold3, without giving up control over sensitive structures, sequences, or ligand libraries. Curious to see it in action? Watch our 3-minute demo video to get a quick walkthrough of how the Co-Folding Application works in practice. Let’s now look at the main design choices and implementation considerations:
| Pillar | Implementation | Why it matters |
|---|---|---|
| Local deployment | All components are packaged as Docker containers and can be easily deployed with Docker. Can also be deployed using AWS CloudFormation with a few clicks. | No data leaves the corporate network; complies with internal security policies. |
| Stream‑lined data preparation | Data preparation toolkit to support benchmarking and fine-tuning. Supports activities such as deduplication, chain parsing, and MSA validation. | Ensures stable inference and comparable benchmarks across datasets. |
| Model portfolio | Pre‑installed Boltz‑2 & OF3; users may add custom models; more models will be added in the future | Compare architectures head‑to‑head on the same proprietary dataset. |
| Inference API & GUI | Python SDK plus web interface with 3‑D viewer | Effective for both quick visual checks of individual runs and large-scale automated analysis |
| Comprehensive metrics | Each model returns the 3D protein structures (in PDB format) with associated confidence metrics, including pLDDT scores and Predicted Aligned Error (PAE) or Predicted Distance Error (PDE) metrics. These are stored locally and downloadable from the web interface. | Rapid triage of false positives; metrics exportable to SD files or dashboards. |
| Optional fine-tuning on own or partner network datasets | Fine-tune models on own or partner data via the Apheris Gateway | Enables highest accuracy for your industrial use cases and increased model generalizability for new targets without exposing raw structures. |
*This feature isn’t part of the initial release but already prioritized for the next update.
Internally, each run generates a traceable and immutable record, including model version, configuration parameters, and input hashes, ensuring full reproducibility and enabling auditability for regulatory submissions and scientific validation. Curious how this works in a real research setting? Read our "How to Get Started" user story to follow a step-by-step example of the Co-Folding Application in action.
Summary & Outlook
Over the past two years, advances in model architecture have improved the predictive accuracy of deep‑learning–based structure‑prediction models. For example, the algorithmic innovations introduced in AlphaFold 3 relative to its predecessor AlphaFold 2 led to significantly better performance on the PoseBusters benchmark, which evaluates both structural accuracy and physical plausibility of protein–ligand complexes. Across multiple target classes, success rates increased substantially; for antibody–antigen complexes in particular, the success rate rose from roughly one‑third to about two‑thirds. Yet because PoseBusters and related assessments are built from public PDB entries, the observed gains by AlphaFold 3 may largely reflect improved interpolation within regions of structural space well-represented in the training data. Co‑folding models perform well in‑distribution but often struggle with out‑of‑distribution predictions (i.e. extrapolation), such as novel folds or ligands with distinct scaffolds. Because benchmarks rarely reflect proprietary sequence or chemotype space, organizations should evaluate model performance on their own data. Building an internal capability to deploy, benchmark, and compare models is the essential first step to fully leveraging the rapid advancements in model development. The Apheris Co-Folding Application enables you to:
Benchmark models on your in-house targets to determine their strengths and limitations and assess model applicability for your specific research tasks
Standardize data preparation, which is the prerequisite for reliable benchmarks, fine‑tuning and collaborative model training
Generate traceable and fully local predictions - including records that support compliance and enable reproducible research
Once you know how well a model performs on your in-house data, you can choose to improve it by further customizations such as fine-tuning or post-hoc analyses. To further increase the model's applicability domain and generalizability, you can join a federated data network. The AI Structural Biology (AISB) Network is such an example, addressing the challenge of limited availability of protein–ligand structural data in the public domain. Through the network, participants collaboratively train AI models across their distributed proprietary datasets without ever sharing raw data and while keeping data IP fully protected. These collaborative networks are a powerful next step to improve model performance, but they all start with the same foundation: understanding how models behave on your own data. The Apheris Co‑Folding Application gives you the infrastructure to do exactly that. It helps you identify model blind spots, generate high‑quality structured inputs, and prepare your organization for local fine‑tuning or secure federated training later on.
For installation steps, API references, or configuration tips access the full product documentation
Ready to put OpenFold3 and Boltz-2 to work on your own environment? Check out our Co-Folding Application Site.