0.2.2
Released 2026-11-19.
Compatibility
- RFProteína backends. Select the execution backend with
inference_backend:eager,compile, orbioir.compileremains the default, so existing compiled runs keep their behavior.
Breaking Changes
- RFProteína is a numerical engine. BioIR runs the encoder, the velocity network, the integrator, and the decoder. The host application owns generation, partial diffusion, search, scoring, sequence constraints, motif assignment, and structure output. The same host code runs on every backend.
- Public API.
sample,sample_partial,redesign,inverse_fold, andsearchare removed. Advance a prepared state withpartial_simulation, which can resume from the previous clean prediction. For custom feature preparation, integration, or readouts, useRFProteinaTrunk.forward_features. - Explicit sampler settings.
from_reference_state_dictsrequiressimulation_config. The built-in defaults serve tests and toy examples only. They do not match the released sampler settings, so a rollout that relies on them follows a different trajectory. - Batch lengths. Sampling uses the actual length of each batch. Compiler buckets stay inside the native trunk, so seeds that came from the older bucket-padded path can produce different outputs.
Performance
- RFProteína sampling. On an H200 with 60 steps and 180 trunk passes, BioIR is faster than the compiled backend in every configuration we measured. Single monomers run 2.3× to 4.7× faster, and batches of 8 and 32 run 1.4× to 1.9× faster. Peak memory stays within 1.7% of the previous BioIR path.