Data loading

This is a tutorial on loading of protein benchmark datasets.

Loading of protein benchmarks

Load the overview table of protein benchmark datasets using the default settings:

import aaanalysis as aa
df_info = aa.load_dataset()
aa.display_df(df=df_info)

	Level	Dataset	# Sequences	Avg length	# Amino acids	# Positives	# Negatives	Predictor	Description	Reference	Label
1	Amino acid	AA_CASPASE3	233	796.587983	185605	705	184900	PROSPERous	Prediction of c...3 cleavage site	Song et al., 2018	1 (adjacent to ... cleavage site)
2	Amino acid	AA_FURIN	71	831.028169	59003	163	58840	PROSPERous	Prediction of f...n cleavage site	Song et al., 2018	1 (adjacent to ... cleavage site)
3	Amino acid	AA_LDR	342	345.754386	118248	35469	82779	IDP-Seq2Seq	Prediction of l...d regions (LDR)	Tang et al., 2020	1 (disordered), 0 (ordered)
4	Amino acid	AA_MMP2	573	546.205934	312976	2416	310560	PROSPERous	Prediction of M...) cleavage site	Song et al., 2018	1 (adjacent to ... cleavage site)
5	Amino acid	AA_RNABIND	221	248.873303	55001	6492	48509	GMKSVM-RU	Prediction of R...(RBP60 dataset)	Yang et al., 2021	1 (binding), 0 (non-binding)
6	Amino acid	AA_SA	233	796.587983	185605	101082	84523	PROSPERous	Prediction of s...PASE3 data set)	Song et al., 2018	1 (exposed/acce...non-accessible)
7	Sequence	SEQ_AMYLO	1414	6.000000	8484	511	903	ReRF-Pred	Prediction of a...ognenic regions	Teng et al. 2021	1 (amyloidogeni...-amyloidogenic)
8	Sequence	SEQ_CAPSID	7935	424.030246	3364680	3864	4071	VIRALpro	Prediction of capdsid proteins	Galiez et al., 2016	1 (capsid prote...capsid protein)
9	Sequence	SEQ_DISULFIDE	2547	241.252454	614470	897	1650	Dipro	Prediction of d...es in sequences	Cheng et al., 2006	1 (sequence wit...ithout SS bond)
10	Sequence	SEQ_LOCATION	1835	399.126975	732398	1045	790	nan	Prediction of s...lasma membrane)	Shen et al., 2019	1 (protein in c...asma membrane)
11	Sequence	SEQ_SOLUBLE	17408	254.611041	4432269	8704	8704	SOLpro	Prediction of s...oluble proteins	Magnan et al., 2009	1 (soluble), 0 (insoluble)
12	Sequence	SEQ_TAIL	6668	400.673365	2671690	2574	4094	VIRALpro	Prediction of tail proteins	Galiez et al., 2016	1 (tail protein...n-tail protein)
13	Domain	DOM_GSEC	126	737.809524	92964	63	63	nan	Prediction of g...tase substrates	Breimann et al, 2024c	1 (substrate), ...(non-substrate)
14	Domain	DOM_GSEC_PU	694	712.570605	494524	63	0	nan	Prediction of g...es (PU dataset)	Breimann et al, 2024c	1 (substrate), ...bstrate status)

The benchmark datasets are categorized into amino acid (‘AA’), domain (‘DOM’), and sequence (‘SEQ’) level datasets, indicated by their name prefix, as exemplified here.

df_seq1 = aa.load_dataset(name="AA_CASPASE3")
df_seq2 = aa.load_dataset(name="SEQ_CAPSID")
df_seq3 = aa.load_dataset(name="DOM_GSEC")
aa.display_df(df=df_seq3, n_rows=8, char_limit=25)
# Compare columns of three types

	entry	sequence	label	tmd_start	tmd_stop	jmd_n	tmd	jmd_c
1	P05067	MLPGLALLLLAA...NPTYKFFEQMQN	1	701	723	FAEDVGSNKG	AIIGLMVGGVVIATVIVITLVML	KKKQYTSIHH
2	P14925	MAGRARSGLLLL...YSAPLPKPAPSS	1	868	890	KLSTEPGSGV	SVVLITTLLVIPVLVLLAIVMFI	RWKKSRAFGD
3	P70180	MRSLLLFTFSAC...REDSIRSHFSVA	1	477	499	PCKSSGGLEE	SAVTGIVVGALLGAGLLMAFYFF	RKKYRITIER
4	Q03157	MGPTSPAARGQG...ENPTYRFLEERP	1	585	607	APSGTGVSRE	ALSGLLIMGAGGGSLIVLSLLLL	RKKKPYGTIS
5	Q06481	MAATGTAAAAAT...NPTYKYLEQMQI	1	694	716	LREDFSLSSS	ALIGLLVIAVAIATVIVISLVML	RKRQYGTISH
6	P35613	MAAALFVLLGFA...DKGKNVRQRNSS	1	323	345	IITLRVRSHL	AALWPFLGIVAEVLVLVTIIFIY	EKRRKPEDVL
7	P35070	MDRAARCSGASS...PINEDIEETNIA	1	119	141	LFYLRGDRGQ	ILVICLIAVMVVFIILVIGVCTC	CHPLRKRRKR
8	P09803	MGARCRSFSALL...KLADMYGGGEDD	1	711	733	GIVAAGLQVP	AILGILGGILALLILILLLLLFL	RRRTVVKEPL

Each dataset can be utilized for a binary classification, with labels being positive (1) or negative (0). A balanced number of samples can be chosen by the n parameter, defining the sample number per class.

# Returns 200 samples, the first 100 positives and the first 100 negatives
df_seq = aa.load_dataset(name="SEQ_CAPSID", n=100)
aa.display_df(df=df_seq, n_rows=8, char_limit=25)

	entry	sequence
1	CAPSID_1	MVTHNVKINKHV...RIPATKLDEENV
2	CAPSID_2	MKKRQKKMTLSN...EAVINARHFGEE
3	CAPSID_3	MRYGGSVISQEL...GAPDKQEVELVD
4	CAPSID_4	MERGDIPFKYVG...LEELAEMDAGLI
5	CAPSID_5	MKRIYLLFAALI...CVGEPITELTHQ
6	CAPSID_6	MARARESQAEAL...RNTISSRISRTW
7	CAPSID_8	MLFKQLFTIISS...VDIIKALIIIKG
8	CAPSID_9	MKILIVEDEPKT...GMGYVLEAPQPR

Or randomly selected using random=True:

# Returns randoms samples, 100 positives and 100 negatives
df_seq = aa.load_dataset(name="SEQ_CAPSID", n=100, random=True)
aa.display_df(df=df_seq, n_rows=8, char_limit=25)

	entry	sequence
1	CAPSID_2351	MESIRGAIWEDR...LHPIEEGNTIHK
2	CAPSID_149	MEMVSGSGSYTA...LRRRQQQDNGTV
3	CAPSID_1161	MFNIDEVYTISS...ALYSNIEKIEKN
4	CAPSID_3899	MAKSLTQILGQA...DWLNDQPRGKSE
5	CAPSID_1424	MTPQATANSLFG...KPPEEEGNGKLE
6	CAPSID_762	MIAGNCRMCLVE...AAGEMSPQALYG
7	CAPSID_3590	MVFDLLRRKTGR...GLPPRTAEVADD
8	CAPSID_3467	MAEPEYLSTRQA...LLGRIAARQRAS

The protein sequences can have varying length:

# Plot distribution
import matplotlib.pyplot as plt
import seaborn as sns

# Utility AAanalysis function for publication ready plots
aa.plot_settings(font_scale=1.2)
df_seq = aa.load_dataset(name="SEQ_CAPSID", n=100)
list_seq_lens = df_seq["sequence"].apply(len)
sns.histplot(list_seq_lens, binwidth=50)
sns.despine()
plt.xlim(0, 1500)
plt.show()

Which can be easily filtered using min_len and max_len parameters:

df_seq = aa.load_dataset(name="SEQ_CAPSID", n=100, min_len=200, max_len=800)
list_seq_lens = df_seq["sequence"].apply(len)
aa.plot_settings(font_scale=1.2)  # Utility AAanalysis function for publication ready plots
sns.histplot(list_seq_lens, binwidth=50)
sns.despine()
plt.xlim(0, 1500)
plt.show()

Loading of protein benchmarks: Amino acid window size

For amino acid level datasets, labels are provided for each residue position, which can be seen by setting aa_window_size=None:

df_seq = aa.load_dataset(name="AA_CASPASE3", aa_window_size=None)
aa.display_df(df=df_seq.head(10), char_limit=25)

	entry	sequence	label
1	CASPASE3_1	MSLFDLFRGFFG...LDLFLGRWFRSR	0,0,0,0,0,0,...,0,0,0,0,0,0
2	CASPASE3_2	MEVTGDAGVPES...LQNPKRARQDPT	0,0,0,0,0,0,...,0,0,0,0,0,0
3	CASPASE3_3	MRARSGARGALL...EMLVAMTTDGDC	0,0,0,0,0,0,...,0,0,0,0,0,0
4	CASPASE3_4	MDAKARNCLLQH...NLGILYILQTLE	0,0,0,0,0,0,...,0,0,0,0,0,0
5	CASPASE3_5	MTSFSTSAQCST...KEIQLVIKVFIA	0,0,0,0,0,0,...,0,0,0,0,0,0
6	CASPASE3_6	MGLGASSEQPAG...PDPEPGLCEGPW	0,0,0,0,0,0,...,0,0,0,0,0,0
7	CASPASE3_7	MANQVNGNAVQL...EFYQDTYGQQWK	0,0,0,0,0,0,...,0,0,0,0,0,0
8	CASPASE3_8	MAKQPSDVSSEC...LRYIVRLVWRMH	0,0,0,0,0,0,...,0,0,0,0,0,0
9	CASPASE3_9	MCTALSPKVRSG...VSASYKAKKEIK	0,0,0,0,0,0,...,0,0,0,0,0,0
10	CASPASE3_10	MFYAHFVLSKRG...IIATPGPRFHII	0,0,0,0,0,0,...,0,0,0,0,0,0

For convenience, we provide an “amino acid window” of length n. This window represents a specific amino acid, which is flanked by (n-1)/2 residues on both its N-terminal and C-terminal sides. It’s essential for n to be odd, ensuring equal residues on both sides. While the default window size is 9, sizes between 5 and 15 are also popular.

df_seq = aa.load_dataset(name="AA_CASPASE3", n=2)
aa.display_df(df=df_seq, char_limit=25)

	entry	sequence	label
1	CASPASE3_1_pos126	QTLRDSMLK	1
2	CASPASE3_1_pos127	TLRDSMLKY	1
3	CASPASE3_1_pos4	MSLFDLFRG	0
4	CASPASE3_1_pos5	SLFDLFRGF	0

Sequences can be pre-filtered using min_len and max_len and n residues can be randomly selected by random with different aa_window_sizes.

df_seq = aa.load_dataset(name="AA_CASPASE3", min_len=20, n=2, random=True, aa_window_size=21)
aa.display_df(df=df_seq, char_limit=25)

	entry	sequence	label
1	CASPASE3_17_pos29	HWPARVDEVPDGAVKPPTNKL	1
2	CASPASE3_36_pos96	TGKRAAEDDEDDDVDTKKQKT	1
3	CASPASE3_4_pos133	VVFVTRKKLVNAIQQKLSKLK	0
4	CASPASE3_199_pos398	KQNGPKTPVHSSGDMVQPLSP	0

Loading of protein benchmarks: Positive-Unlabeled (PU) datasets

In typical binary classification, data is labeled as positive (1) or negative (0). But with many protein sequence datasets, we face challenges: they might be small, unbalanced, or lack a clear negative class. For datasets with only positive and unlabeled samples (2), we use PU learning. This approach identifies reliable negatives from the unlabeled data to make binary classification possible. We offer benchmark datasets for this scenario, denoted by the _PU suffix. For example, the DOM_GSEC_PU dataset corresponds to the DOM_GSEC set.

df_seq = aa.load_dataset(name="DOM_GSEC", n=2)
aa.display_df(df=df_seq, char_limit=25)

	entry	sequence	label	tmd_start	tmd_stop	jmd_n	tmd	jmd_c
1	Q14802	MQKVTLGLLVFL...TPPLITPGSAQS	0	37	59	NSPFYYDWHS	LQVGGLICAGVLCAMGIIIVMSA	KCKCKFGQKS
2	Q86UE4	MAARSWQDELAQ...QIKKKKKARRET	0	50	72	LGLEPKRYPG	WVILVGTGALGLLLLFLLGYGWA	AACAGARKKR
3	P05067	MLPGLALLLLAA...NPTYKFFEQMQN	1	701	723	FAEDVGSNKG	AIIGLMVGGVVIATVIVITLVML	KKKQYTSIHH
4	P14925	MAGRARSGLLLL...YSAPLPKPAPSS	1	868	890	KLSTEPGSGV	SVVLITTLLVIPVLVLLAIVMFI	RWKKSRAFGD

df_seq_pu = aa.load_dataset(name="DOM_GSEC_PU", n=2)
aa.display_df(df=df_seq_pu, char_limit=25)

	entry	sequence	label	tmd_start	tmd_stop	jmd_n	tmd	jmd_c
1	P05067	MLPGLALLLLAA...NPTYKFFEQMQN	1	701	723	FAEDVGSNKG	AIIGLMVGGVVIATVIVITLVML	KKKQYTSIHH
2	P14925	MAGRARSGLLLL...YSAPLPKPAPSS	1	868	890	KLSTEPGSGV	SVVLITTLLVIPVLVLLAIVMFI	RWKKSRAFGD
3	P12821	MGAASGRRGPGL...PQFGSEVELRHS	2	1257	1276	GLDLDAQQAR	VGQWLLLFLGIALLVATLGL	SQRLFSIRHR
4	P36896	MAESAGASSFFP...LSQLSVQEDVKI	2	127	149	EHPSMWGPVE	LVGIIAGPVFLLFLIIIIVFLVI	NYHQRVYHNR

See out Data Handling API and the Tables section for more details.