AstraROLE & AstraSUIT: Multi-Task Annotation Models for Functional Profiling of Proteins

AstraROLE and AstraSUIT instantly deliver a rich, ten-factor functional profile for any protein sequence — EC class, GO term, pathway, protein category, cofactor use, domain family, likely host, membrane type, transmembrane helix count, and subcellular location — by uniting transformer embeddings with physicochemical features.

Why AstraROLE & AstraSUIT?

Most pipelines juggle multiple single-task models. AstraROLE and AstraSUIT replace them with one call that delivers ten functional labels in seconds—often at or above specialist accuracy—speeding iteration and cutting false leads.

Limited Scope
Limited Scope
Limited Scope

Integrated Coverage

Together the two models predict EC class, GO terms, pathway memberships, protein category, cofactor type, domain, host association, membrane type, TM-helix count and subcellular localisation—giving researchers a full experimental checklist.

Experimental Efforts
Experimental Efforts
Experimental Efforts

Hypothesis Generation

Because predictions are built on ESM-2 embeddings enriched with physicochemical features, the models generalise well to low-homology, engineered or de-novo proteins.

Missed Insights
Missed Insights
Missed Insights

Real-Time Throughput

Upon handing a sequence to AstraROLE and AstraSUIT, within seconds, they predict key annotations on the given protein — replacing the usual tangle of separate tools and keeping the projects moving.

Benchmark-Topping 0.98 F1

The AstraROLE & AstraSUIT heads reach macro F1 scores of 0.82–0.98. In direct benchmarks, the Astra models equal or exceed specialist performance in five of seven tests—particularly excelling in membrane type, and cofactor predictions.

The Architecture

A 1 351-feature input vector (CLS token from the 650 M-parameter ESM-2 model plus physicochemical enrichments) passes through a shared 512-unit transformer encoder. Task-specific linear heads — four in AstraROLE, six in AstraSUIT — apply sigmoid activations and per-label cut-offs, delivering all ten predictions in one forward pass with millisecond latency on a single GPU.

See AstraROLE and AstraSUIT in action.

Curious to see how AstraROLE and AstraSUIT help on understanding the protein function and location from a simple sequence input? Try the demo, or book a meeting, and let us showcase you how to harness AstraROLE & AstraSUIT.