IC-1611Galactica-6.7b fails on protein secondary structure editing tasks, producing hit ratios below random mutation

Shengchao Liu, Jiongxiao Wang, Yijin Yang, Chengpeng Wang, Ling Liu, Hongyu Guo, Chaowei Xiao

SourceConversational Drug Editing Using Retrieval and Domain Feedback

On the two protein editing tasks (more helix, more strand), Galactica-6.7b as the ChatDrug backbone achieves hit ratios of 11.75 and 5.99 respectively, both below the random mutation-3 baseline of 26.90 and 21.44. The authors note that Galactica 'cannot deal with protein editing tasks' while it performs adequately on small molecule and peptide editing. This suggests a domain-specific limitation in the model's ability to reason about protein sequences and secondary structure.

Evidence
correlational
Key metric
task 501 (more helix): 11.75 hit ratio; task 502 (more strand): 5.99 hit ratio; random mutation-3 baseline: 26.90 and 21.44 respectively
Caveat
Only two protein editing tasks were tested. The model is not instruction-tuned, which the authors note requires a different prompt template. Performance is measured within the ChatDrug framework.
Model
Galactica-6.7B
Concepts
Failure mode
Datasets
TAPe [eval]
Methods
ESMFold [eval], ProteinCLAP-EBM-NCE [eval]
Related work
ESMFold [context]
Related findings
IC-1610, IC-1612
Extraction
automatic-extraction