Model for psychrophilic¶
- Description: Binary classification of a bacterial phenotype (psychrophily) from Pfam protein-domain content of the genome, for strains in the BacDive-AI release (13,669 strains x 2,047 Pfam families). Random train/val/test split -- the protocol used by prior work. Strains of one species share ~99% of gene content, so this places near-duplicates on both sides and is optimistic; compare against the bacdive_clade variant. Note that trait prevalence ranges from 2.6% to 55%, so accuracy is a weak summary; macro-F1 is the informative metric. External links: https://github.com/JKoblitz/bacdive-AI, https://doi.org/10.5281/zenodo.13757323
Reference(s): https://doi.org/10.1038/s42003-025-08182-w, https://github.com/JKoblitz/bacdive-AI;https://doi.org/10.5281/zenodo.13757323, https://doi.org/10.5281/zenodo.13757323, https://github.com/JKoblitz/bacdive-AI
Model benchmarks
| Model name | Dataset | ACC | Team name | Dataset size | Date submitted | Notes |
|---|---|---|---|---|---|---|
| g2p_logistic | bacdive | 0.9702 | JHU | 13058 | 09-07-2026 | CSV, JSON, run.sh, Info |
| g2p_rf | bacdive | 0.9705 | JHU | 13058 | 09-07-2026 | CSV, JSON, run.sh, Info |
| g2p_multitask | bacdive | 0.9652 | JHU | 13058 | 09-07-2026 | CSV, JSON, run.sh, Info |
| g2p_gbm | bacdive | 0.9725 | JHU | 13058 | 09-07-2026 | CSV, JSON, run.sh, Info |
| g2p_mlp | bacdive | 0.9721 | JHU | 13058 | 09-07-2026 | CSV, JSON, run.sh, Info |