Release Notes

Release 2.6.0

Summary

This release moves the whole prediction path onto the BioNeMo Inference Runtime. Parsing, MSA and template featurization, and structure writing were previously handled by OpenFold code with the runtime used only as an optimized model. They are now handled by the runtime end to end. Predictions are unchanged within measurement error, and the container is substantially smaller.

Key Features

• Security: Addressed all CVE (Common Vulnerabilities and Exposures) issues.
• Single optimized runtime: The NIM no longer carries a separate data pipeline. Accuracy was checked against experimental structures rather than against the previous pipeline: the top-ranked lDDT-Ca difference is +0.0034 with a 95% confidence interval of [-0.0031, +0.0100], that is, no detectable change. Refer to Performance for per-SKU measurements.
• Smaller container: The image builds on stock PyTorch plus one wheel. HMMER, HHsuite, MMseqs2, OpenMM, and DeepSpeed were removed. They served MSA search and relaxation, which this NIM never performs because alignments arrive in the request.

Notes and Limitations

• aligned_confidence_probs in each structure record is deprecated and is always null. Its [N, N, 64] shape serialized to hundreds of megabytes for long sequences. Use predicted_aligned_error instead; the field remains in the response schema so existing clients do not break on a missing key.
• Structural relaxation remains unsupported. relax_prediction must be omitted or false.
• All other features from release 2.5.0 remain supported.