Skip to content

Release NotesπŸ”—

Apheris Hub 1.4.0πŸ”—

  • Release date: 2026-05-28

ApherisFold Application: 0.60.0πŸ”—

  • OpenFold3
    • openfold3:0.60.0
  • Boltz-2
    • boltz2:0.60.0
  • Protenix
    • protenix:0.60.0
  • Mock Model
    • mock:0.60.0

HighlightsπŸ”—

The Apheris Hub 1.4.0 release introduces affinity prediction support for OpenFold3 and Boltz-2, bringing quantitative binding affinity estimates alongside structural predictions. This release also delivers the next iteration of fine-tuning with full start, cancel, and continue capabilities, including checkpoint deployment to new model weights and real-time training metrics. Additionally, this version includes numerous stability and usability improvements across the prediction, benchmark, and fine-tuning workflows, along with important security fixes.

New FeaturesπŸ”—

  • Affinity Prediction for ApherisFold (OpenFold3 & Boltz-2): Added support for predicting protein-ligand binding affinity alongside structural predictions. When an affinity-capable model and weights are selected, affinity is automatically predicted for the ligand in single-ligand queries; when multiple ligands are present, users designate for which ligand chain to predict affinity. Results display model-specific affinity metrics ("Predicted affinity") in the results stats table and are included in downloaded metrics files. For Boltz-2, binding probability and delta values are also shown. Affinity capability is contract-driven, ensuring future affinity models appear automatically without UI changes. See more on the documentation.
  • Fine-Tuning (Start, Cancel, Continue): Completed the fine-tuning feature set with full lifecycle support: users can now start fine-tuning jobs, cancel running jobs, and continue completed jobs with additional training time. Checkpoints can now be deployed as model weights and used for prediction, benchmarking or fine-tuning. Real-time training metrics are streamed via WebSocket as training progresses, and the fine-tuning chart now includes the baseline model performance at gradient step 0 for direct comparison. See more on the documentation.
  • Updated SMILES/CCD Toggle: Moved the SMILES/CCD input toggle into the ligand data-entry field for a cleaner query builder layout, particularly with the addition of affinity prediction controls. The redesigned pill toggle matches the surrounding input field style for improved visual consistency. See more on the documentation.
  • Automatic Line Break Removal from Sequence Input: Sequences pasted or written into the query builder are now automatically stripped of line breaks, preventing submission failures caused by accidental multi-line paste from external tools.

Breaking ChangesπŸ”—

  • Public MSA Mode Removed from Model Wrappers: The "public" MSA mode in OpenFold3 and Boltz-2 wrappers has been removed. All MSA generation is now exclusively handled by-file or via configured MSA servers in the Hub UI. Deployments or scripts that relied on the wrapper-internal public MSA mode must migrate to using Hub's MSA server configuration.

EnhancementsπŸ”—

  • Prediction and Query Building:

    • Replaced legacy synchronous prediction flow with asynchronous wrapper execution (predict-async endpoints), improving reliability and enabling better status tracking for running jobs.
    • Fixed model weight selection being changeable after a fine-tune job creation form is in progress; the model dropdown is now disabled once a job is created.
    • Fixed MSA server configuration not being loaded on the fine-tuning page when opening in a fresh browser session.
    • Improved session handling to attempt silent re-authentication on token expiration, reducing unexpected logouts during active use.
    • Fixed results state not transitioning to "done" when navigating to job details while a WebSocket update is pending.
    • Fixed model and weight fuzzy search returning incorrect results or ghost group headers.

  • Results and Visualization:

    • Fixed zero metric values being displayed as a dash (-) on benchmark detail pages; they now render as "0.00" consistently.
    • Fixed fine-tuning job duration not being displayed on the job details page after completion.
    • Fixed fine-tuning list showing status text for completed jobs; duration and a progress bar are now shown instead.
    • Fixed 3D structure viewer not rendering when pLDDT confidence data is unavailable; the viewer now shows with an informational notice instead of an error.
    • Added comparison of checkpoint performance against the base model (gradient step 0) directly in the fine-tuning metrics chart.
    • Real-time fine-tuning metrics updates via WebSocket, so charts update as training progresses without a page refresh.

  • Benchmark Workflow:

    • Fixed benchmark page going blank after cancelling a running benchmark job.
    • Fixed benchmark state not updating on-screen after cancellation without a manual reload.
    • Fixed cancel button not appearing during the "preparing" state of a benchmark.
    • Fixed MSA error snackbar being shown after user-initiated benchmark cancellation.
    • Fixed query status text being truncated on the benchmark results detail page.
    • Fixed MSA matching for CIF files that share chain sequences across structures.
    • Fixed duplicate MSA files being rejected when the same sequence appears across benchmark structures; redundant files are now silently skipped and reused.
    • Fixed stale benchmark draft state persisting when navigating back to the benchmark page.
    • Fixed model parameters not being updated when changing models after adding a benchmark structure.

  • Fine-Tuning Workflow:

    • Fixed form layout issue causing upload and remove buttons to be hidden when using validation files with SDF and MSA uploads at narrow breakpoints.
    • Improved scheduler to correctly route fine-tuning jobs based on available backends with fine-tuning capability.
    • Fixed reset of fine-tune form when the selected model becomes unavailable mid-session.
    • Fixed consistent status colors between prediction/benchmarking and fine-tuning job states.

  • Infrastructure and Security:

    • Added writable weights volume support so the Hub coordinator can deploy fine-tuned weights directly to model instances without requiring a restart.
    • Refactored model discovery and routing to support adapter-backed inference targets, enabling future integrations with external inference providers.
    • Added NetworkPolicy to restrict model pods to communicate only with the Hub, preventing unintended external network access.

  • Model Wrapper (ApherisFold Application):

    • Added affinity prediction support for OpenFold3 (via SandboxAQ integration) and Boltz-2.
    • Added asynchronous predict endpoints (/api/v1/predict-async) to all model wrappers.
    • Added dynamic weight loading: newly deployed fine-tuned weights are discoverable by running model instances without requiring a restart.
    • OpenFold3 model weights are now persisted in GPU memory between predictions, significantly reducing latency for sequential inference requests.
    • Added top-K checkpoint retention for fine-tuning runs to prevent uncontrolled disk usage from checkpoint accumulation.
    • Added fine-tuning cancel/continue support to the Mock model, enabling full lifecycle testing without GPU resources.
    • Fixed extraction of custom ligands directly from CIF files in certain conditions, removing the requirement to supply separate SDF files for already-embedded custom ligands.
    • Improved error messages for invalid CIF files uploaded via the "Start from CIF" workflow.
    • Fixed error detail exposure for failed async prediction jobs; the actual error is now returned in the API response instead of a generic message.

  • Documentation:

    • Added notes on chown correction for OpenFold3 fine-tuning checkpoint deployment permission issues.
    • Added documentation for cuequivariance kernel compilation delay on first OpenFold3 inference or fine-tuning forward pass after container start.

Apheris Hub 1.3.2πŸ”—

  • Release date: 2026-04-30

ApherisFold Application: 0.49.0πŸ”—

  • OpenFold3
    • openfold3:0.49.0
  • Boltz-2
    • boltz2:0.49.0
  • Protenix
    • protenix:0.49.0
  • Mock Model
    • mock:0.49.0

EnhancementsπŸ”—

  • Benchmark Workflow:
    • Added support for multiple MSA files per structure
  • Fine-Tuning Workflow:
    • Added support for multiple MSA files per structure

Apheris Hub 1.3.1πŸ”—

  • Release date: 2026-04-02

ApherisFold Application: 0.49.0πŸ”—

  • OpenFold3
    • openfold3:0.49.0
  • Boltz-2
    • boltz2:0.49.0
  • Protenix
    • protenix:0.49.0
  • Mock Model
    • mock:0.49.0

EnhancementsπŸ”—

  • Documentation:
    • Updated the required CUDA version to avoid potential issues when using Boltz-2.
    • Added a known issue on Dataset Generation for SDF files with explicit hydrogens.

Apheris Hub 1.3.0πŸ”—

  • Release date: 2026-03-31

ApherisFold Application: 0.49.0πŸ”—

  • OpenFold3
    • openfold3:0.49.0
  • Boltz-2
    • boltz2:0.49.0
  • Protenix
    • protenix:0.49.0
  • Mock Model
    • mock:0.49.0

HighlightsπŸ”—

The Apheris Hub 1.3.0 release introduces fine-tuning capabilities and expands model support with Protenix integration and new OpenFold3 weights aligned with the March 2026 release. This release also includes multiple improvements to MSA server management and support, including NVIDIA NIM integration, global pre-configured MSA servers and enhanced MSA service architecture.

New FeaturesπŸ”—

  • Fine-Tuning V1 (Beta): Introduced fine-tuning capabilities to enable model customization with user-specific datasets for advanced workflows and large-scale personalization. Users can submit fine-tuning jobs by specifying model endpoints, passing data in standardized formats, and configuring training parameters. Fine-tuning is currently supported for OpenFold3 and Mock models. The fine-tuning workflow automatically handles MSA preparation for protein chains when needed, integrating with the existing MSA service infrastructure to fetch alignments. This feature is currently in beta and under active development. See Fine-tuning for more information.
  • Protenix Model Support: Added Protenix-v1 to ApherisFold, providing users access to the latest open-source co-folding methods. The model includes two sets of weights (1.0.0 and 1.0.1) that users can choose from on the model page, with availability for inference. This addition expands modeling options and gives users another high-performing alternative for structure prediction. See Models for more information.
  • New OpenFold3 Consortium Release Weights: Added new OpenFold3 model weights from the latest public release, providing users with improved prediction capabilities. Two new weight sets are now available: version 3.0.0 (OpenFold3-preview model trained with a 2021-09-30 PDB data cutoff) and version 4.0.0 (OpenFold3 March 2026 release model with the same PDB data cutoff but incorporating architectural and training improvements). These weights represent the main model architecture moving forward for OpenFold3 and offer enhanced prediction accuracy and performance.
  • NVIDIA NIM MSA Server Integration: Extended MSA server support to include NVIDIA NIM (v2) for ColabFold MSA generation. This integration enables seamless connection to enterprise MSA infrastructure for organizations using NVIDIA BioNeMo platforms. See the MSA Service documentation for more information.
  • Global Pre-Configured MSA Servers: Added deployment-time configuration for organization-wide MSA servers. Administrators can now configure multiple read-only MSA servers that are available to all platform users, specify a default active server, and ensure at least one MSA server is configured when the MSA component is enabled. See the Kubernetes deployment or Docker deployment documentation for configuration details.
  • Enhanced Query Building from CIF Files: Improved the workflow for creating queries from structure files by adding drag-and-drop support directly into the query builder and renaming the "PREVIOUS" button to "START FROM" with expanded functionality. Users can now drop CIF files into the query builder to automatically generate queries, with multiple CIFs creating multiple queries up to the maximum allowed. The START FROM modal includes a "Build query from CIF" option at the top for selecting files from disk, while maintaining access to previous runs at the bottom. See the Query Builder documentation for more information.

Breaking ChangesπŸ”—

  • MSA Server Configuration Changes: MSA server configuration has been moved from individual user settings to deployment-time configuration for organization-wide servers, with user settings limited to enabling/disabling and selecting from pre-configured options. Existing MSA server configurations may need to be migrated to the new deployment-time configuration. See the Kubernetes deployment or Docker deployment documentation for migration guidance.

EnhancementsπŸ”—

  • Prediction and Query Building:

    • Added support for pre-filled examples via URL, enabling users to share links that open the prediction form with example queries pre-populated.
    • Prevented cursor jumping in JSON editor when typing invalid JSON, allowing users to add missing properties without interference.
    • Fixed molecule type validation to update before async validation, preventing incorrect "Please add MSA asset" messages on ligand chains.
    • Added background to MSA asset error messages to prevent text overlap with other UI elements.
    • Enforced minimum of one chain ID in model schema to prevent submission failures.

  • Results and Visualization:

    • Improved the PAE/PDE matrix view to create three synchronized selections when users select a region: identity sequence (if identity was selected), aligned sequence, and scored sequence. All three selections appear as colored highlights in both the sequence viewer and molecular viewer, matching the behavior of the AlphaFold Server and providing more intuitive structure exploration.
    • Standardized the display of pLDDT metrics across the platform to use the lowercase "p" format (pLDDT) consistently, matching common scientific convention and the format shown in the 3D viewer information panel.
    • Fixed exact ID search to bypass filter mode in results and fine-tune lists, ensuring searches work regardless of active filters.
    • Added conditional rendering for optional metrics (PAE/PDE/top-level stats) with clear fallback UI when data is missing.
    • Fixed results details to automatically load when job status transitions to done via websocket.

  • Benchmark Workflow:

    • Disabled submit button during benchmark start request to prevent duplicate submissions.
    • Added per-chain MSA failure reason display on benchmark structure rows.

  • User Interface Improvements:

    • Reduced model weights first column width and softened navigation divider.
    • Cleaned up model selection dropdown with improved layout and removed "show all" option.
    • Added model ID to dropdown selection for clearer identification.
    • Updated predict "Previous runs" to "Build query from previous run" with clearer context.
    • Filtered predict inputs for "start from" to display predictions only, not benchmarks.
  • Infrastructure and Configuration:
    • Added provisions for paid licensing to the Hub Terms & Conditions, preparing for future commercial usage models. The updated terms indicate that Apheris reserves the right to limit usage based on conditions and consider trial or evaluation variants of the Hub.

  • Model Wrapper Enhancements:

    • Activated confidence matrix in model responses.
    • Fixed RNA sequence parsing in mmCIF loading.
    • Extended /query_from_mmcif response to return missing ligands requiring custom SDF files.

DocumentationπŸ”—

  • Custom Model Weights: Expanded documentation for configuring custom weight versions and managing multiple weight sets. See Customizing Model Weights for more information.

Apheris Hub 1.2.6πŸ”—

  • Release date: 2026-02-19

HighlightsπŸ”—

The Apheris Hub 1.2.6 release adds per-model nodeSelector and affinity support to the Helm chart, enabling explicit scheduling and node placement for model deployments such as OpenFold3.

New FeaturesπŸ”—

  • Per-Model nodeSelector and affinity Support in Helm Chart: Added support for configuring nodeSelector and affinity at the model instance deployment level (models.instances.<model>.deploy.nodeSelector and models.instances.<model>.deploy.affinity) and wired both through chart templates and schema validation. This allows targeting specific node pools and expressing scheduling rules for individual model deployments.

Apheris Hub 1.2.5πŸ”—

  • Release date: 2026-02-16

HighlightsπŸ”—

The Apheris Hub 1.2.5 release adds support for custom HTTP headers on MSA server configurations.

New FeaturesπŸ”—

  • Custom HTTP Headers for MSA Servers: Added support for configuring custom HTTP headers on MSA server configurations. Users can now add key-value header pairs that will be included in all requests to the MSA server, useful for custom authentication or proxy requirements.

Apheris Hub 1.2.4πŸ”—

  • Release date: 2026-02-12

ApherisFold Application: 0.37.1πŸ”—

  • OpenFold3
    • openfold3:0.37.1
  • Boltz-2
    • boltz2:0.37.1
  • Mock Model
    • mock:0.37.1

HighlightsπŸ”—

The Apheris Hub 1.2.4 release introduces ForgeRock identity provider support, NVIDIA NIM ColabFold MSA server integration, and updates default model wrapper versions from 0.37.0 to 0.37.1 across Kubernetes and standalone deployment defaults.

New FeaturesπŸ”—

  • ForgeRock Identity Provider Support: Added first-class ForgeRock OIDC integration for Hub authentication, including provider-specific configuration via hub.auth.providerType=forgerock and support for ForgeRock login/token exchange flows through Hub endpoints to improve compatibility with enterprise ForgeRock deployments.
  • NVIDIA NIM ColabFold MSA Server Integration: Added support for NVIDIA NIM ColabFold as an MSA server type in Hub and UI workflows, including configuration support and related API updates for server type handling.

EnhancementsπŸ”—

  • Model Version Update: Updated default ApherisFold model image tags to 0.37.1 for OpenFold3, Boltz-2, and Mock model deployments in Helm chart defaults and standalone deployment configuration. This update fixes an issue where wrong CUDA libraries were being picked up in some circumstances.

Apheris Hub 1.2.3πŸ”—

  • Release date: 2026-02-02

ApherisFold Application: 0.37.0πŸ”—

  • OpenFold3
    • openfold3:0.37.0
  • Boltz-2
    • boltz2:0.37.0
  • Mock Model
    • mock:0.37.0

HighlightsπŸ”—

The Apheris Hub 1.2.3 release adds Kubernetes support for customizable model weights, enabling operators to mount and configure multiple weight sets without rebuilding model images.

New FeaturesπŸ”—

  • Customizable Model Weights for Kubernetes: Added Helm values and templates to mount custom weight files and configure weight selection for OpenFold3 and Boltz-2 deployments. Documentation includes configuration examples and operational guidance.

DocumentationπŸ”—

  • Custom Model Weights: Expanded guidance in the ApherisFold Application documentation with configuration methods, Kubernetes examples, and verification steps.

Apheris Hub 1.2.2πŸ”—

  • Release date: 2026-01-30

ApherisFold Application: 0.37.0πŸ”—

  • OpenFold3
    • openfold3:0.37.0
  • Boltz-2
    • boltz2:0.37.0
  • Mock Model
    • mock:0.37.0

HighlightsπŸ”—

The Apheris Hub 1.2.2 release introduces custom Certificate Authority (CA) certificate support for enterprise environments with internal PKI infrastructure, enabling secure integration with identity providers and internal services using enterprise TLS certificates across both Kubernetes and Docker deployments.

New FeaturesπŸ”—

  • Custom CA Certificate Support: Added support for mounting custom CA certificates in both Kubernetes (via Helm values) and Docker deployments. Certificates are automatically trusted alongside system CAs, enabling secure connections to identity providers and internal services using enterprise TLS. Documentation includes verification steps for both deployment modes.

DocumentationπŸ”—

  • Dex Integration Guide: Restructured with two clear deployment scenarios (same-URL and internal/external URL patterns), prerequisite checks, practical verification commands, and complete configuration examples.
  • Custom CA Certificates: New documentation for Kubernetes (ConfigMap creation, Helm configuration, verification via debug containers) and Docker (configuration and verification).

Apheris Hub 1.2.1πŸ”—

  • Release date: 2026-01-28

ApherisFold Application: 0.37.0πŸ”—

  • OpenFold3
    • openfold3:0.37.0
  • Boltz-2
    • boltz2:0.37.0
  • Mock Model
    • mock:0.37.0

HighlightsπŸ”—

The Apheris Hub 1.2.1 release focuses on stability improvements and bug fixes that enhance the user experience when working with benchmarking workflows and prediction forms. In addition it introduces configurable MSA Server timeouts and enhanced logging capabilities to improve the debugging process of integrating an MSA Server with the Apheris Hub.

New FeaturesπŸ”—

  • Configurable MSA Server Timeouts & Logging: Extended MSA server configuration to allow operators to configure request timeouts via Helm values. Added comprehensive logging that shows request submission, status, duration, and errors for both Foldify and ColabFold server types. This enhancement enables users to adjust timeouts based on their specific server performance and network conditions, while providing better visibility into MSA generation processes for debugging slow or failed requests.

EnhancementsπŸ”—

  • Prediction and Query Building:

    • Fixed issue where clicking on benchmark examples multiple times would add files multiple times to the form.
    • Fixed model schema validation that previously prevented form submission after switching the active MSA server on the benchmark form.
    • Improved asset recovery when creating a new prediction from a previous job result, ensuring input assets are properly restored.
    • Added warning indicator when model settings are updated, helping users identify when model parameters may need adjustment after changing models.
    • Fixed SMILES to CCD code toggle validation to trigger immediately upon switching and allow proper toggling back without requiring input clearing.
    • Improved ColabFold response parsing to handle string responses correctly.
  • Model Wrapper Enhancements:
    • Fixed bugs in Foldify MSA truncation that occasionally caused benchmark queries to fail with array shape broadcasting errors.

Apheris Hub 1.2.0πŸ”—

  • Release date: 2026-01-15

ApherisFold Application: 0.36.0πŸ”—

  • OpenFold3
    • openfold3:0.36.0
  • Boltz-2
    • boltz2:0.36.0
  • Mock Model
    • mock:0.36.0

HighlightsπŸ”—

The Apheris Hub 1.2.0 release introduces comprehensive benchmarking capabilities. This release focuses on enabling users to evaluate model performance against ground-truth structures through an intuitive benchmarking workflow.

New FeaturesπŸ”—

  • Benchmarking: Introduced a complete benchmarking workflow that allows users to upload multiple ground-truth structures as CIF files, automatically validate and process them, and compare predictions against experimental structures. The benchmark form supports uploading structure files with optional MSA files or using MSA servers (Foldify/ColabFold) for automatic MSA generation. Results are displayed with both per-structure metrics and aggregated statistics across all benchmark submissions, including metrics such as Ligand RMSD, Protein CΞ± LDDT, and Ligand-Protein Interaction LDDT. The benchmarking feature helps users assess whether models are applicable to their specific protein targets and fall within the model's domain of applicability. For full details about this new feature check our new Benchmarking documentation section.
  • Production Authentication Support: Extended authentication capabilities to support multiple identity providers: Auth0, DEX and Microsoft Entra ID. The authentication system now provides flexible integration with external frameworks using OAuth protocol allowing enterprise-ready deployments. When authentication is enabled, complete user data segregation ensures that users can only access their own prediction queries, results, and benchmark data. See the new Authentication Setup documentation section for more information.
  • Configurable MSA Browser Timeout: Added user-configurable timeout setting in the Settings page for MSA server requests made directly from the browser. This setting controls how long the browser will wait for a response from ColabFold MSA servers before timing out, allowing users to adjust timeouts based on their specific server performance and network conditions.
  • Reuse Previous Job Configurations: Added capability to create new prediction jobs from previous results through a Set up new prediction button on the results page. All job details including sequences, model parameters, and tags are pre-populated and fully editable, streamlining iterative refinement of prediction parameters.
  • Dynamically Changeable Model Weights: Added support for modifying available model weights without requiring model code changes. Users can now mount custom weight files into the model container and provide a configuration file that defines the available weight sets. This enables seamless switching between different weight configurations such as public weights, fine-tuned weights, or federated weights, making it easier to evaluate and deploy custom-trained models alongside pre-packaged weights. See our documentation for more information on how to use Custom Model Weights.

Breaking ChangesπŸ”—

  • Renamed Navigation: Updated navigation terminology to be more familiar to chemistry users, replacing "Inference" with "Predict" throughout the interface. The main navigation now uses clearer terms: MODELS, PREDICT, BENCHMARK and RESULTS. Renamed /api/v1/applications endpoints to /api/v1/models for clearer terminology.

EnhancementsπŸ”—

  • Prediction and Query Building:

    • Added validation to allow 'X' as a valid amino acid sequence character.
    • Enhanced validation to trigger across all sub-components of forms preventing partial validation states.
    • Fixed SMILES renderer to prevent element ID overlaps when multiple structures are displayed.
    • Increased maximum amino acid sequence length limit.

  • Results Visualization:

    • Fixed superposition structure coloring to properly gray out ground truth when using chain selection buttons.
    • Improved sparkline visualization to properly resize when window dimensions change.

  • Infrastructure and Configuration:

    • Enhanced authentication configuration validation to occur at application startup rather than at runtime.
    • Improved Helm chart to support global label additions across all resources.
    • Added readiness probe to deployment for better health monitoring.
    • Enhanced support ZIP to include registry.yaml and correct nvidia-smi availability detection.
    • Improved API key handling in Helm chart for consistency with Docker standalone deployment.
    • Fixed model image pull secrets to properly inherit from global configuration.

  • Model Wrapper Enhancements:

    • Improved error handling to store full stack traces in error files for better debugging.
    • Fixed CIF to query conversion to properly return responses even when only errors occur.
    • Made query name hashing case-insensitive to prevent duplicate processing issues.
    • Enhanced bond inference for unknown ligands using 3 standard deviations for improved structure quality.
    • Fixed Boltz-2 error handling to prevent crashes with invalid argument counts.
    • Corrected Boltz-2 sample indexing to start at 1 instead of 0.

Apheris Hub 1.1.2πŸ”—

  • Release date: 2025-12-15

ApherisFold Application: 0.28.0πŸ”—

  • OpenFold3
    • openfold3:0.28.0
  • Boltz-2
    • boltz2:0.28.0
  • Mock Model
    • mock:0.28.0

EnhancementsπŸ”—

  • Documentation:
    • Documentation has been reviewed and updated to reflect the changes around the Hub Coordinator mode.

Apheris Hub 1.1.1πŸ”—

  • Release date: 2025-12-09

ApherisFold Application: 0.28.0πŸ”—

  • OpenFold3
    • openfold3:0.28.0
  • Boltz-2
    • boltz2:0.28.0
  • Mock Model
    • mock:0.28.0

HighlightsπŸ”—

No relevant changes were done in this version.

Apheris Hub 1.1.0πŸ”—

  • Release date: 2025-12-05

ApherisFold Application: 0.28.0πŸ”—

  • OpenFold3
    • openfold3:0.28.0
  • Boltz-2
    • boltz2:0.28.0
  • Mock Model
    • mock:0.28.0

HighlightsπŸ”—

The Apheris Hub 1.1.0 release introduces major architectural improvements with Kubernetes-native deployment and coordinator mode, enabling more flexible and secure enterprise deployments. This release marks a significant shift in deployment architecture, moving away from Docker API dependencies to support Kubernetes orchestration.

New FeaturesπŸ”—

  • Kubernetes-Native Deployment with Helm Charts: The Hub now offers first-class support for Kubernetes deployments via Helm charts, designed for production environments in pharmaceutical and enterprise settings. This deployment method allows seamless integration with existing Kubernetes infrastructure, provides better scalability, and eliminates dependencies on the Docker API. Cloud-based (AWS EKS) Kubernetes deployments are fully supported with comprehensive documentation while we are still maintaining support for simplified Docker-based deployments through auxiliary deployment scripts.
  • Hub Coordinator Mode: Introduced a new architectural pattern where the Hub operates in coordinator mode, decoupling model lifecycle management from the Hub itself. In this mode, models and their wrappers can be deployed and managed independently, with the Hub discovering and coordinating access to already-running models. This removes the requirement for docker.sock access, addressing critical security concerns in enterprise environments, and enables better alignment with orchestration technologies like Kubernetes and job schedulers like Slurm. For more on this, see the documentation.
  • Model Weight Management: Added comprehensive support for selecting between multiple available model weight sets within a single model deployment. Users can now select from different weight configurations (that in the future can be public weights, fine-tuned weights or federated weights) through both the UI and API. A new /weights endpoint allows querying available weight sets, with each set including metadata like version and description for easy identification.

Breaking ChangesπŸ”—

  • Deployment Mode Changes: CloudFormation and single Docker container deployment methods are no longer maintained. Supported deployment flows are now limited to Kubernetes with Helm charts (for production and development) and auxiliary Docker scripts (for local evaluation). Users relying on deprecated deployment methods must migrate to one of the supported options.
  • Docker API Dependency Removed: The Hub no longer requires or uses Docker API integration for managing model lifecycle. Model installation, starting, and stopping through the Docker API has been removed.
  • PostgreSQL Database Required: SQLite is no longer supported as the backend database. The Hub now requires PostgreSQL for improved performance and better concurrency and state handling.
  • Weight Version Parameter Required: The /predict endpoint now requires a weightVersion parameter to specify which model weights to use. Requests without this parameter will fail. This change enables multi-weight support but requires updates to existing prediction scripts.
  • Chain IDs Type Enforcement: The chain_ids field in prediction requests now strictly requires an array of strings (string[]). Previously accepted weak typing like single strings is no longer supported to prevent ambiguous interpretations.

EnhancementsπŸ”—

  • Prediction and Query Building:

    • Added validation to prevent duplicate chain IDs within a query.
    • Implemented prevention of duplicate file attachments in query builder.
    • Added validation requiring at least one query per prediction request.
    • Improved JSON request parameter validations including maximum query limits.
    • Fixed enter key behavior in query builder to prevent unintended actions.
    • Trimmed whitespace from sequences and SMILES inputs automatically.

  • Results and Visualization:

    • Fixed SMILES to CCD code toggle functionality in ligand chain configuration.
    • Added ligand atom pLDDT coloring in molecular viewer.
    • Improved PAE/PDE plot legend to reflect actual displayed data.

  • Generic User Interface Improvements:

    • Added clear button to search input fields.
    • Fixed issues with CodeMirror panels overlapping modals.
    • Added visual indication of the number of running and pending jobs per model version.

  • Model Wrapper Enhancements:

    • Improved typing of API responses removing null types and making fields consistently optional.
    • Enhanced CIF validation response typing for better frontend integration.
    • Improved error messages when invalid sequences are submitted.
    • Enhanced error surfacing when GPUs are configured incorrectly.
  • Documentation:
    • Updated Query Builder documentation for multi-ligand prediction support.

Apheris Hub 1.0.0πŸ”—

  • Release date: 2025-10-28

ApherisFold Application: 0.22.0πŸ”—

  • OpenFold3
    • openfold3:0.22.0
  • Boltz-2
    • boltz2:0.22.0
  • Mock Model
    • mock:0.22.0

HighlightsπŸ”—

The Apheris Hub 1.0.0 release marks a significant milestone with the introduction of a graphical Query Builder and structure prediction evaluation capability. This release focuses on improving user experience for non-technical users while adding powerful features for model performance assessment, including full-screen visualization modes and support for structure comparison workflows. Additionally, infrastructure improvements enable better troubleshooting through support ZIP exports and more robust model management capabilities.

New FeaturesπŸ”—

  • Query Builder for Prediction: Introduced a graphical Query Builder that allows users to construct prediction requests without writing JSON manually. The builder supports adding multiple queries, selecting molecule types from dropdowns (Protein, Ligand, DNA, RNA), defining chain IDs and sequences, and attaching external assets. Users can seamlessly switch between the graphical builder and JSON editor, with automatic validation to prevent errors.
  • Full-Screen 3D Molecular Viewer: Added ability to maximize the molecular viewer to utilize full screen space for detailed structural inspection. The Full-Screen Molecular Viewer includes the 3D viewer, PAE/PDE plot, and sequence bar, with options to minimize individual components as needed.
  • Evaluation with Structure Superposition: Enabled comparison of predicted structures against ground-truth experimental structures. When a ground-truth structure is provided with the prediction request, the system automatically performs structure alignment and calculates key metrics including Average pLDDT, Ligand RMSD, Ligand-Protein Interaction LDDT, Protein CΞ± GDT-HA, Protein CΞ± GDT-TS, Protein CΞ± LDDT, Protein CΞ± RMSD. The aligned structures are displayed as overlays in different colors within the molecular viewer. Check the documentation for Evaluation Examples.
  • Easy export of diagnostic and configuration data to a ZIP file: Added a Support ZIP Archive feature that allows users to download a compressed file containing structured diagnostic and configuration data - from the Apheris Hub Docker container - to streamline external support and debugging.
  • ColabFold MSA Server Configuration: Extended MSA server configuration to support ColabFold servers in addition to the existing Foldify server support. Users can now configure their own private ColabFold server or use the public ColabFold server for MSA generation. When configuring prediction runs, users can select which MSA server to use from their configured options, allowing them to compare performance between different servers or integrate their existing ColabFold infrastructure. Check the documentation sections Configurable MSA Server and MSA Usage to know more about this new feature.

Breaking ChangesπŸ”—

  • Model Version Naming Simplification: Removed the -public-colabfold suffix from model versions. All models now use a single version naming scheme (e.g., boltz2:0.22.0 instead of boltz2:0.22.0-public-colabfold). Users should update any scripts or configurations that reference the old version naming convention. MSA server selection is now managed through the MSA configuration interface rather than through model version selection.
  • Application Renaming: The application has been renamed from "Apheris Co-folding" to "ApherisFold" to better reflect its focus on protein structure prediction.

EnhancementsπŸ”—

  • Prediction Runs / Results:

    • Added real-time updates for humanized time displays (e.g., "2 minutes ago") that refresh every 30 seconds.
    • Enabled persistence of last used model version, so the prediction form remembers previously selected models.
    • Added exact match search capability for job IDs, improving result filtering when searching by job identifier.
    • Fixed pLDDT coloring to work correctly across all model types.

  • Model Management:

    • Added automatic timeout handling for pending prediction requests to prevent queue blockage.
    • Implemented automatic timeout for accepted jobs that remain in accepted state too long.
    • Added automatic cancellation of pending requests when a model is uninstalled.
    • Improved container lifecycle management by removing stale "created" containers before retry attempts.

  • MSA Configuration:

    • Enabled configuration of ColabFold MSA servers in addition to Foldify servers.
    • Added default public ColabFold MSA server configuration for evaluation purposes.
    • Improved MSA configuration validation to prevent editing of preconfigured system servers.
    • Improved wording and user guidance around MSA options in the prediction form.

  • Infrastructure and Logs:

    • Increased database connection pool size and timeout values to handle concurrent write operations better.
    • Enhanced archive endpoint to include input assets alongside job outputs for complete result packages.
    • Improved error message formatting and capitalization for better readability.

  • User Interface - Generic:

    • Redesigned previous inputs modal to handle large numbers of attachments gracefully.
    • Improved tag display in results view with better overflow handling.
    • Prevented premature result loading when jobs are still running to avoid "file not found" errors.
    • Fixed action button behavior to prevent closing settings panels unexpectedly.
    • Updated asset upload interface with clearer wording and improved MSA-specific guidance.

  • Model Wrapper:

    • Enabled by-file MSA support for OpenFold3, allowing users to upload pre-generated MSA files.
    • Added query generation and validation API endpoint for structure file processing.
    • Standardized pLDDT values to 0-100 range across all models for consistency.
  • Documentation:
    • Extended documentation with additional GPU information and requirements.
    • Added section with Apheris Hub configuration parameters.
    • Added Apheris Hub Troubleshooting Guide.

Apheris Hub 0.3.1πŸ”—

  • Release date: 2025-09-12

ApherisFold Application: 0.11.1πŸ”—

  • Boltz-2
    • boltz2:0.11.1
    • boltz2:0.11.1-public-colabfold
  • Mock Model
    • mock:0.11.1
    • mock:0.11.1-public-colabfold
  • OpenFold3
    • openfold3:0.11.1-public-colabfold

HighlightsπŸ”—

The Apheris Hub 0.3.1 release introduces significant enhancements to the Apheris Hub with focus on protein structure analysis and user experience improvements.

New FeaturesπŸ”—

  • MSA Server Configuration: Users can now use different Multiple Sequence Alignment (MSA) servers for structure prediction.

MSA Server Options:

  • Pre-Generated MSA: The base models assume the user will upload pre-generated MSA files; this is the most local option, making no calls outside the Hub to any MSA server.
  • Apheris-hosted Foldify: The base model now supports using a Foldify server, and for a limited time, comes pre-configured with an Apheris-hosted option for free, intended only for evaluation use. This server should be treated as a public MSA server, and users should take the same precautions when using it as they would, for example, the public ColabFold Server.
  • Self-hosted Foldify: You can also configure the Hub to use your own internally hosted Foldify server; adding ColabFold servers is coming soon. Please contact Apheris for support in setting up your own Foldify server via support@apheris.com.
  • Public ColabFold: Until support is available for configuring access to ColabFold servers, there is a public-colab-fold model option that generates MSAs via the public ColabFold server, which is not hosted by Apheris.
  • Real-time Model Installation Feedback: Added progress indicators for model installation, showing percentage completion and status updates in real-time. If installation issues occur, error messages are displayed with troubleshooting information available in model settings.
  • Structure Visualization with Mol: Migrated to Mol for rendering 3D molecules, delivering enhanced visualization and interaction capabilities.

EnhancementsπŸ”—

  • Improved 3D Molecule Visualization:

    • Adopted Mol* for rendering 3D molecules, providing better visualization capabilities and performance.
    • Enhanced molecule coloring and added Chemical Component Dictionary (CCD) code lookups for improved annotation.
    • Removed background from molecule viewer and added hover information for better visualization.
    • Added legend and units for matrix view to improve data interpretation.

  • User Interface Improvements:

    • Added unique naming for downloaded results to prevent filename conflicts and improve organization.
    • Redesigned the job form for a more intuitive user experience.
    • Added job ID display in result list for easier tracking.
    • Enhanced offline connection status indication for better user feedback when connectivity is lost.
    • Duration formatting improved to show minutes and seconds instead of just seconds.
    • Updated model naming conventions to clearly reflect MSA server options. Models now indicate whether they use the public ColabFold server, providing transparency about external dependencies.

  • Prediction Runs / Results Improvements:

    • Enhanced sequence selection highlighting for precise amino acid selection.
    • Added catching of MMSeqs2 MSA server errors from Boltz-2 model.

Bug FixesπŸ”—

  • UI and Configuration

    • Improved error formatting for MSA linting.
    • Updated schema validation to properly handle MSA field requirements.

  • Prediction Runs / Results Improvements:

    • Clarified error messages for mock model prediction results when non-default inputs are used.
    • Fixed OpenFold3 pLDDT vectors for improved accuracy.

Apheris Hub 0.2.1πŸ”—

  • Release date: 2025-08-22

ApherisFold Application: 0.10.1πŸ”—

  • Boltz-2
    • boltz2:0.10.1-private-msa
    • boltz2:0.10.1-public-msa
  • Mock Model
    • mock:0.10.1-private-msa
    • mock:0.10.1-public-msa
  • OpenFold3
    • openfold3:0.10.1-public-msa

HighlightsπŸ”—

  • Enhancements
    • Detection of API connection loss on the UI
  • Bug Fixes
    • Labels are now shown on the PAE/PDE Plot
    • ColabFold server unavailability error messages are clearer when using Boltz-2

Apheris Hub 0.2.0πŸ”—

  • Release date: 2025-08-08

ApherisFold Application: 0.9.0πŸ”—

  • Boltz-2
    • boltz2:0.9.0-private-msa
    • boltz2:0.9.0-public-msa
  • Mock Model
    • mock:0.9.0-private-msa
    • mock:0.9.0-public-msa
  • OpenFold3
    • openfold3:0.9.0-public-msa

HighlightsπŸ”—

The Apheris Hub provides local access to AI-based drug discovery applications. Applications deployed in the Apheris Hub are designed to run entirely locally or in your private cloud or on-prem environments - such that no data ever leaves your environment. They integrate into enterprise R&D stacks and extend rather than replace existing scientific workflows.

Applications in the Apheris Hub focus on a set of drug discovery tasks and are based on one or more foundation models.

Our first application is focused on Co-Folding, starting with prediction - including input validation, and batch execution either programmatically via an API or manually with a scientist-friendly GUI.

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.

Upcoming ImprovementsπŸ”—

Our first release is designed to support early evaluation and adoption of OpenFold3, and we are already making many improvements for follow-up releases. We are investing in more powerful visualizations, more models, job management, and supporting fine-tuning locally and via secure federation.

We plan to release additional support for data preparation, benchmarking, and fine-tuning in our early pipeline. We can already support integrating these models into your existing benchmarking pipelines, but we'd like to add more to ensure the biggest impact on your benchmarking efforts. We'd love to hear any feedback you have around this as well. Please feel free to reach out with feature requests, comments, or suggestions at support@apheris.com.

In future versions, we will allow pointing to an existing private MSA server for OpenFold3. However, today Apheris can also help you set up a private MSA server if you wish. In that case, please contact us at support@apheris.com.

For the first release, submitting additional jobs will put those jobs in a queue. The GPU is fully consumed per job; parallelism and multi-GPU support will come with follow-up releases. The current architecture was built for simplicity in deployment and configuration with more flexibility and scalability coming very soon.