PPISeq

Installation

1. Clone ppiseq repo.
2. cd ppiseq
3. python -m pip install -U -e .

Training

export CUDA_VISIBLE_DEVICES=0,1

# For full finetuning with BFloat16:
accelerate launch --mixed-precision bf16 train.py

# For ConvBERT training with frozen embeddings:
accelerate launch train.py

# Important Note:
# Set `ddp_find_unused_parameters` parameter to `true` when training ESM2 or ESM3,
# this parameter is in the `train` key in the training/config/config.yaml

Results

Table 1: Top 10 runs ranked by test set Spearman correlation ($\rho$). Metrics (mean $\pm$ standard deviation, 3 seeds): Spearman $\rho$, Pearson r, and RMSE ($pK_d$). PAD: Pooled attention addition; HP: Hierarchical pooling

PLM	Setup	Spearman	Pearson	RMSE	Spearman	Pearson	RMSE
		(Validation)	(Validation)	(Validation)	(Test)	(Test)	(Test)
------------	-----------------	----------------	----------------	--------------	----------------	----------------	--------------
Prot-T5	ConvBERT-PAD	0.48 $\pm$ 0.02	0.48 $\pm$ 0.01	1.52 $\pm$ 0.09	0.48 $\pm$ 0.03	0.51 $\pm$ 0.02	1.42 $\pm$ 0.10
Ankh2-Ext1	Finetuning-HP	0.48 $\pm$ 0.01	0.49 $\pm$ 0.01	1.50 $\pm$ 0.10	0.47 $\pm$ 0.01	0.48 $\pm$ 0.01	1.45 $\pm$ 0.13
ESM2-650M	ConvBERT-HP	0.44 $\pm$ 0.03	0.43 $\pm$ 0.03	1.86 $\pm$ 0.29	0.47 $\pm$ 0.02	0.48 $\pm$ 0.02	1.74 $\pm$ 0.33
Ankh2-Ext2	Finetuning-HP	0.47 $\pm$ 0.01	0.47 $\pm$ 0.01	1.51 $\pm$ 0.07	0.45 $\pm$ 0.01	0.46 $\pm$ 0.02	1.43 $\pm$ 0.06
ESM2-650M	Finetuning-HP	0.45 $\pm$ 0.02	0.44 $\pm$ 0.02	1.75 $\pm$ 0.29	0.44 $\pm$ 0.02	0.45 $\pm$ 0.01	1.68 $\pm$ 0.29
Ankh-Base	Finetuning-HP	0.47 $\pm$ 0.01	0.47 $\pm$ 0.00	1.47 $\pm$ 0.03	0.44 $\pm$ 0.01	0.45 $\pm$ 0.01	1.41 $\pm$ 0.03
Prot-T5	ConvBERT-HP	0.42 $\pm$ 0.02	0.42 $\pm$ 0.02	1.66 $\pm$ 0.23	0.44 $\pm$ 0.01	0.44 $\pm$ 0.01	1.59 $\pm$ 0.25
Prot-T5	Finetuning-PAD	0.44 $\pm$ 0.00	0.44 $\pm$ 0.00	1.65 $\pm$ 0.13	0.44 $\pm$ 0.01	0.45 $\pm$ 0.01	1.57 $\pm$ 0.14
Prot-T5	Finetuning-HP	0.47 $\pm$ 0.02	0.46 $\pm$ 0.02	1.51 $\pm$ 0.05	0.44 $\pm$ 0.01	0.45 $\pm$ 0.01	1.47 $\pm$ 0.08
ESM2-3B	Finetuning-PAD	0.45 $\pm$ 0.01	0.45 $\pm$ 0.01	1.49 $\pm$ 0.07	0.44 $\pm$ 0.01	0.46 $\pm$ 0.00	1.43 $\pm$ 0.08

Note: PAD: Pooled attention addition; HP: Hierarchical pooling