Apheris Hub - Known Issues

Below, we list known issues with Apheris Hub. If available, we list a workaround to help troubleshoot the issue.

This list will be constantly evolving with input from our different users and with the evolution of our product in future releases.

Known Issues with Apheris Hub 1.4

OpenFold3

Issue	Description	Additional information
Potential Memory Leak with OpenFold3 Benchmarks	When running benchmarks with OpenFold3, some jobs may fail due to a potential memory leak.	This issue is currently under investigation.
OpenFold3 Does Not Support Certain Atom Names	OpenFold3 (and the AlphaFold3 models it derives from) only supports atom names using characters in the range 32-95, which excludes lower case letters. mmCIF files with technically valid lower case atom names will cause the fine-tuning pipeline to crash.	Capitalize all atom names in your mmCIF files before submitting them.

Protenix

Issue	Description	Additional information
Protenix Does Not Support Blackwell GPUs	Protenix fails with "CUDA error: no kernel image is available for execution on the device" when running on NVIDIA Blackwell architecture GPUs (e.g. B200, B100, AWS g7e instances). This is an upstream limitation with no current workaround.	Use an Ampere or Hopper GPU (e.g. A100, H100, AWS g6 instances) for Protenix workloads until upstream adds Blackwell support.

Prediction and Processing

Issue	Description	Additional information
ColabFold MSA Request Payload Not Included in Errors	When ColabFold MSA generation fails, the error message does not include the original request payload.	Use server-side logs to diagnose ColabFold MSA failures.
ColabFold Server Rate Limiting	The public ColabFold server applies rate limiting, which can significantly slow down processing of even simple queries.	If you encounter slower than anticipated query times it is likely due to using the public ColabFold server. Please wait a few minutes and try again or use MSAs from file if the model supports it.
Delayed Execution on Shared GPU	When we have a single GPU being shared between two running model versions, parallel jobs can be started on each of those models. However, because they share the same GPU, the execution can be blocked and delayed.	Additionally, the UI shows the delayed requests as "Running" instead of "Pending", as they are waiting for the GPU to become available.
No validation that MSA file sequence and sequence provided in queries match	Queries containing unknown residues (X character) at the beginning of sequences, particularly when combined with ligands like HEM, may cause MSA generation errors.	Future releases will improve the validations being performed before a user submitting a prediction run known to fail.
Nonstandard residues between parentheses are not supported	Including nonstandard residues in parentheses causes the MSA to crash. Generally, these should be converted to X before sending to an MSA server.	Future releases will add support for Nonstandard residues.
Discrepancies between bare-metal model runs and ApherisFold runs for the same model weights.	ApherisFold uses several optimizations in its codebase, which differ from the model defaults in some cases (e.g. OpenFold3 Preview 1/2), these cause minor differences in structural output. In addition, ApherisFold uses a harmonized post-processing pipeline to ensure direct comparison between models, which differs from specific models' own pipeline and may introduce small differences in the final structure.

Large Structures and Out-of-Memory Handling

Issue	Description	Additional information
Large structures may fail with a GPU out-of-memory error	A prediction can fail with a GPU out-of-memory (OOM) error when a structure is too large for the available GPU hardware. GPU memory use grows steeply with sequence length, and predictions are not currently checked against available GPU memory before they run, so an oversized structure is only detected once the prediction fails.	Use a GPU with more memory, or reduce the input by cropping around the region of interest. See VRAM Requirements for OpenFold3 and Other Models in the Troubleshooting guide for sizing guidance.
GPU out-of-memory reported as "Unknown error occurred"	When a Boltz-2 prediction exceeds the available GPU memory (for example, a very large single-chain input on a smaller GPU), the model detects the out-of-memory condition, but the Hub reports it to the client as a generic Prediction ... failed: Unknown error occurred. A model out-of-memory therefore cannot be easily distinguished from a service failure.	Use a GPU with more memory or reduce the size of the structure. To confirm an out-of-memory condition, check the model worker logs for a "ran out of memory" warning.
Large OpenFold3 predictions may hang instead of reporting an error	For very large complexes, the OpenFold3 data loader can exhaust the container's shared memory, which prevents the prediction from progressing. Instead of failing with an out-of-memory error, the prediction hangs at the data-loading stage and never completes.	Reduce the size of the structure, or increase the shared memory available to the OpenFold3 container in your deployment (for example, to 32 GiB). See OpenFold3 Predictions Hang on Large Structures in the Troubleshooting guide.
Very large predictions may fail while saving results	A prediction for a very large structure can complete successfully on the GPU but then fail while the Hub saves its results, because the raw prediction output exceeds an internal database size limit. When this happens, the job is marked as failed and no structure is delivered through the Hub or CLI, even though inference succeeded.	Reduce the size of the structure. If a prediction is affected, the generated structure may still be recoverable directly from the model worker's output directory.

Benchmarking

Issue	Description	Additional information
Boltz-2 and Protenix Benchmarks May Omit Ligand Metrics	Benchmark runs using Boltz-2 or Protenix may sometimes omit ligand metrics in the results, particularly for predictions with less accurate atom distances.
Cancel Button Available on Terminally Failed Benchmarks	The Cancel button remains visible and clickable on a benchmark that has already terminally failed with no running jobs. Clicking it shows a cancellation confirmation but has no meaningful effect.

Fine-Tuning

Issue	Description	Additional information
Dataset Generation Fails for SDF Files with Explicit Hydrogens	Fine-tuning dataset generation fails when SDF files contain explicit hydrogen atoms. SDF files with explicit hydrogens are not currently supported and will cause the dataset generation phase to fail.	Remove explicit hydrogens from SDF files before uploading them for fine-tuning. Tools such as RDKit or Open Babel can be used to strip hydrogens from SDF files (e.g., using RDKit's RemoveHs() or Open Babel's -d option).

User Interface

Issue	Description	Additional information
PAE Filter Issues on Window Resize	When resizing the browser window or zooming out while viewing results, the PAE filter selection may appear and disappear unexpectedly. Clicking deselect may not work consistently.	If you experience this issue, try avoiding window resize operations while working with PAE plots.
Cannot Switch Users After Logout	After logging out, logging back in with the same organization automatically signs in as the previous user. It is not possible to select a different user without clearing the browser cache or using a new window.	Clear browser cache or use an incognito/private window to log in as a different user within the same organization.
"Set Up New Prediction" Button May Not Restore Query Correctly	On the results page, clicking "Set up new prediction" to rebuild a query from an existing result may not work correctly when the original model used is no longer available.	If the button does not work as expected, set up the prediction manually using the Predict page.

Form and Input Handling

Issue	Description	Additional information
Invalid CIF Upload Accepted Without Error Status	When uploading an invalid CIF file to a fine-tune dataset, the API returns a success status instead of an error. The frontend works around this by checking the response body, but the API response code is misleading.

Installation and Configuration

Issue	Description	Additional information
Port Allocation	The installation script uses a fixed port (8080) which may already be in use on some systems.	As a workaround, ensure port 8080 is available before installation or modify the port in the installation script.
Single GPU Limitation with Multiple Models	When deploying the Hub using Kubernetes and Helm on a machine with only one GPU, only one model (OpenFold3 or Boltz-2) can be started at a time.	Use a machine with multiple GPUs, enable fractional GPU usage, or run one model at a time. This limitation should be considered when planning your deployment infrastructure.
Bash Version Check Missing in Deploy Script	The deploy_apherisfold script requires GNU Bash version 4.0 or higher but does not check this requirement before execution. Running the script with an older version or a different shell results in cryptic error messages.	Ensure you are using GNU Bash version 4.0 or higher before running the deployment script. Verify your bash version with bash --version.
Structure File Uploads May Fail for Larger Files or Over SSH/SSM Tunnels	Uploading larger structure files (for example, ~8 MB), or several at once, can fail with a network error or an upload that never completes. This happens when the deployment's request-handling timeout is not configured with enough headroom for larger structures, and is more frequent when the Hub is accessed through an SSH or SSM port-forwarding tunnel, whose session can be cut during a long upload.	See Structure Uploads Fail or Time Out in the Troubleshooting guide for the configuration workaround.