[New Model] PoET-2 for DMS Zero-Shot Benchmarks

[timt51]

This PR adds a new baseline model, PoET-2, for the following benchmarks: DMS substitutions and DMS indels.

1. To download the model weights and MSAs used by the model to make predictions, run the following:

   cd proteingym/baselines/PoET-2
   make download
   

   This will save the model weights in the directory ~/.cache/ProteinGym/baselines/PoET-2 and MSAs in the directory ~/.cache/ProteinGym/baselines/PoET. Note that the MSAs are the same as those used for PoET(-1).

2. To make predictions, run

   • scripts/scoring_DMS_zero_shot/scoring_PoET_2_substitutions.sh for the substitutions benchmark or
   • scripts/scoring_DMS_zero_shot/scoring_PoET_2_indels.sh for the indels benchmark

   Predictions will be saved to the directories ${DMS_output_score_folder_subs}PoET-2 and ${DMS_output_score_folder_indels}PoET-2 for the substitutions and indels benchmarks respectively.

   These scripts will download ~12GB of predicted structures from AlphaFoldDB and save them in the directory ${PROTEINGYM_CACHE}/baselines/PoET-2/DMS_AF2_structures_cache. The download may take awhile. If you would like to perform the download step first without performing the model inference step, you can run the scripts with the environment variable SAMPLE_PROMPTS_ONLY=1. When this environment variable is set, GPUs will not be utilized, but an NVIDIA GPU must still be present on the system in order for the imports to run without failure.

   Lastly, note that the scripts will try to utilize all GPUs that are available to them.

3. Using the provided scripts for computing model performance on the DMS benchmarks, I obtained an average Spearman correlation of 0.500 for the DMS substitutions benchmark and 0.573 for the DMS indels benchmark.

4. Precomputed PoET-2 predictions can be downloaded at the following links:

• DMS Substitutions
• DMS Indels

[timt51]

Closing in favor of #89, which includes code for both DMS and clinical unsupervised benchmarks.