Performance

System information

The CAFA3 experiment was executed on a local workstation with the following specifications:

  • CPU: Intel Core i5-12400F (12 cores, 6 performance cores × 2 threads, up to 5.6 GHz)

  • RAM: 32 GB total

  • GPU: NVIDIA GeForce RTX 3060 (12 GB GDDR6)

  • Operating system: Ubuntu 24.04.2 LTS (Noble Numbat), kernel 6.14.0-29-generic

  • Storage: NVMe SSDs

Workload summary

  • Sequences to annotate: 3,328 (from cafa3_gt.fasta)

  • Models enabled: 5 (ESM, ESM3c, Ankh3-Large, Prot-T5, Prost-T5)

  • Layers used: last layer only (final embedding representation)

Embedding stage

During this stage, embeddings were generated for all five enabled models across the benchmark set of ~3,300 sequences. The process lasted about 40 minutes from start to finish, ensuring that each sequence was represented once per model.

This completed the embedding phase, after which the resulting representations were passed on to the lookup and annotation steps.

Lookup stage

The lookup phase started once embeddings were available and lasted about four minutes. A total of 16,640 similarity queries were enqueued and processed in 35 homogeneous batches (batch size 516, grouped by model and layer). During this stage, the system also produced the consolidated FASTA file (sequences.fasta) containing 3,328 queries and 24,992 reference entries. Lookup completed once all queues were emptied.

Post-processing

Immediately afterwards, a post-processing step was applied to the 3,328 benchmark accessions. This phase lasted around two minutes and included filtering, identity calculations and score weighting. It marked the final step of the annotation workflow prior to evaluation.