ApherisFold

ApherisFold is a secure, locally deployed application from Apheris that enables pharmaceutical and computational drug discovery teams to run, benchmark, and fine-tune leading co-folding models — including OpenFold3, Boltz-2, and Protenix-v1 — directly within their own infrastructure. By keeping all data, inference queries, and outputs inside the organisation's environment, ApherisFold allows teams to leverage proprietary structural data without exposing it externally.

The platform addresses a core challenge in drug discovery: co-folding models perform inconsistently when applied to novel targets, underrepresented protein conformations, or complex biomolecular assemblies. ApherisFold gives computational chemists and structural bioinformaticians the tools to understand where predictions are reliable and to systematically improve model performance through controlled fine-tuning on internal datasets.

Key Challenges ApherisFold Addresses

• Novel targets or ligand chemotypes: When a protein–ligand complex differs significantly from training data, models often fail to predict accurate binding poses. ApherisFold enables fine-tuning to correct these gaps.
• Underrepresented protein conformations: If a ligand binds to a conformation rarely present in public datasets, models may predict a more common but incorrect conformation.
• Large or complex biomolecular assemblies: Predictions can become unreliable for multimers or complexes with large sequence sizes.
• Allosteric binding sites: Models frequently favour orthosteric binding modes due to their prevalence in structural databases, leading to incorrect predictions for allosteric ligands.

Supported Prediction Tasks

• Monomer structure prediction: Predict the 3D structure of individual proteins to understand folding and domain architecture, using OpenFold3, Boltz-2, and Protenix-v1.
• Protein–protein interactions: Predict 3D structures of multi-protein complexes, including antibody–antigen complexes.
• Protein–ligand complex structure: Predict binding site geometry and 3D pose of small molecules for structure-based drug design.
• Binding affinity estimation: Boltz-2 provides estimates of binding affinity for protein–ligand interactions in addition to structural predictions.
• DNA/RNA complex prediction: Model structural complexes of nucleic acids (dsDNA, RNA), useful for regulatory protein modelling, CRISPR systems, and base-specificity analysis.
• All prediction tasks support lead optimisation and virtual screening workflows. Additional models will be added in future releases.

Fine-Tuning and Benchmarking Capabilities

• Teams can prepare training datasets from proprietary structural data and run reproducible fine-tuning experiments entirely within their own environment.
• Fine-tuned models can be evaluated against base models or federated model checkpoints using consistent benchmarking setups.
• Even small amounts of proprietary structural data can yield meaningful improvements — a case study fine-tuning OpenFold3 on just 10 PDE10A protein–ligand complexes corrected systematic pose errors and improved interface metrics on 17 held-out structures.
• Fine-tuned checkpoints can be deployed across drug programmes to support structure-informed design decisions.

Workflow Integration and User Interface

• Computational chemists can run inference, benchmarking, and fine-tuning workflows programmatically through APIs, integrating results into screening, design, and prioritisation pipelines.
• Medicinal chemists can inspect predicted binding modes interactively in a 3D interface to develop new design hypotheses — identifying new pockets, hydrogen bond opportunities, or interactions that could improve affinity in the next design cycle.
• Teams can evaluate large numbers of candidate compounds, compare predicted binding modes, and prioritise molecules for synthesis in the next DMTA cycle.

Federated OpenFold3 Initiative

• The AISB Network's Federated OpenFold3 Initiative, powered by Apheris and conducted with the AlQuraishi Lab at Columbia University, fine-tuned OpenFold3 across proprietary pharma structural datasets using federated learning.
• Participating partners included AbbVie, Astex, Bristol Myers Squibb, Johnson & Johnson, and Takeda.
• The federated model outperformed both public and locally fine-tuned baselines on interface metrics and demonstrated stronger generalisability across targets and chemotypes.

ApherisFold Lite — Browser-Based Evaluation

• ApherisFold Lite allows teams to explore the product directly in a browser without any infrastructure setup or compute provisioning.
• Users can run OpenFold3, explore prediction outputs, and review how results are organised within the product interface.
• ApherisFold Lite is free to use and designed for early evaluation prior to full deployment.
• The full ApherisFold deployment adds secure inference on proprietary data, systematic benchmarking, API-based workflow integration, and fine-tuning on internal or federated datasets.

ApherisFold can be deployed on-premise or in private cloud environments, ensuring all data and outputs remain within the organisation's infrastructure. The platform is designed for computational chemists, structural bioinformaticians, and medicinal chemists working within pharmaceutical drug discovery programmes who require both scientific rigour and data security when applying state-of-the-art co-folding models.