. 2024 Aug 28;15:7407. doi: 10.1038/s41467-024-51844-2

Table 2.

Task-specific datasets*

Prediction level	Sequence diversity	Prediction task/data set	Number of sequences			Average length (number of residues/tokens)
Prediction level	Sequence diversity	Prediction task/data set	Train	Validation	Test	Average length (number of residues/tokens)
per-protein	Mutational landscapes	GFP^22,66	21,446	5362	27,217	237.0
		AAV^23,67	28,626	3181	50,776	736.3
		GB1^23,68	2691	299	5743	265.0
	Diverse datasets	Stability^22,69	53,614	2512	12,851	45.0
		Meltome^23,70	22,335	2482	3134	544.5
		SubCellLoc^34,71	9503	1678	490	519.9
per-residue		Disorder^19,56	1056	118	117	118.1
per-residue		SecStr^14,73	9712	1080	364	255.0

^*Prediction level: Per-protein predictions make a single prediction for an entire protein; per-residue predictions provide one number for each residue (position) in a protein. Sequence diversity: distinguishes between tasks with experimental data specific for individual proteins (mutational landscapes) and those for which data mixes different proteins from different organisms. Prediction task: as described in Methods; SubCellLoc: sub-cellular location, SecStr: secondary structure prediction (in 3 states: helix, strand, other). A number of sequences (i.e., proteins, note this is NOT the number of samples, e.g., for secondary structure prediction N proteins – number given in table - correspond to over 200*N residues in the data set): typical cross-validation using Train to optimize fine-tuning, Validation to optimize hyperparameters, and Test only to assess performance (Methods).