We have used the hand-curated data set of Holland et al. 6, the PROSITE homeodomain data set 10 and the NCBI database 11.

In order to find amino acid patterns that distinguish genes from the ANTP subclasses, as well as genes from the PRD class, from genes belonging to other classes, a fast word discovery program (12; available at 13) was first used to find statistically interesting words. The minimum word length used was two (-m 2) and the minimum number of sequences where the word should occur (-e) was set to 10% of the number of sequences in the data set being considered. The words found were then filtered to discard those words that occurred more than three times in more than one dataset. The software Bioredx (14; available at 13) was then used to find the largest and more general amino acid patterns that distinguish two sets of sequences. This program accepts as input two sequence files and a word seed. Based on the seed given, tries to find patterns that occur in one file (referred as the positive file) and not in the other (referred as the negative file). This software was used to find patterns that occurred less than four times in the negative file and considered patterns up to 20 amino acid residues larger than the initial word seed.

In the reported amino acid patterns, amino acids listed within brackets are those allowed at a given position. For compactness of representation it is also possible to negate the class. The negation is denoted by "^". In this case, the amino acid residues listed are the ones that are not allowed at that particular position. The approach here used to find amino acid patterns does not use a set of aligned sequences. Therefore, in order to make sure that the derived amino acid patterns are found in the same region of the homeodomain sequence, all sequences belonging to a given class and sub-classes were aligned using the PAM 250 scoring matrix 15.

There are many schemes for comparing and grouping amino acids. Nevertheless, none of them can possibly capture the vast number of contexts in which amino acids are found within proteins. The scheme here used is that of Livingstone and Barton 16, based on the amino acid properties size, polarity, hydrophobicity, charge, aliphaticity and aromaticity. For each position and property we calculated the probability of gaining or losing a constraint using the distribution of the amino acid property's observed at that position in all sequences considered. Only changes in the properties values with a probability of occurrence lower than 5% were considered.

A Neighbour-joining tree, using pair-wise deletion, as implemented in the MEGA software 17 was constructed in order to classify the set of 918 homeodomain sequences from a variety of animal species (available at PROSITE 10), into the 12 classes and two sub-classes scheme proposed by Holland et al. 6.

Many characteristic amino acid patterns were detected using the approach described in Material and Methods. Here, we describe the seven patterns that are most pertinent, taking into account their coverage across a group of genes and absence in the other groups.

The pattern [KT] [IV]WFQNRR [AMV]K [DEHKLMQWY] [KR] [KR] (positions 46–58; Fig. 1), named HOXL 1 pattern, is present in all human HOXL homeodomains, and in none of the human NKL and PRD genes.

The pattern LE [AGKNR]E (positions 16–19; Fig. 1), named HOXL 2 pattern, is present in all human HOXL homeodomains except Mnx1. Only two human PRD genes (Pax4 and Pax6) encode this amino acid pattern. None of the human NKL genes encode this amino acid pattern.

The pattern [AKST] [DENPS] [LAST] [Q] [V] (positions 41–45; Fig. 1), named NKL pattern, is present in 36 out of 39 human NKL sequences, in one human HOXL sequence (Mnx1), and in none of the human PRD sequences. The human NKL sequences that do not present this pattern are: HHEX, NANOG, and VENTX.

The pattern L [EINQRV] [^DGHMPTVWY] [^CDGKMNPQR] [FL] [^CFILPTWY] [AEFHKQRV] [ADEGKNSTW] [CHKMPQR] [FHY]P (positions 16–26; Fig. 1), named PRD pattern, is present in all human PRD homeodomains and never present in human HOXL and NKL sequences.

The patterns [HQ] [IV] [AKLT] (positions 44–46; Fig. 1), named ANTP 1 pattern, and [IKTV] [ITV]W [FY]QN [HQR]R [AMNTVY]K, named ANTP 2 pattern, (positions 46–55; Fig. 1) are present in all human HOXL and NKL genes and never present in human PRD homeodomains.

The pattern W [FY] [KQESR] [NK] [HQRKY] [RW] (positions 48–53; Fig. 1), named ANTP-PRD pattern, is found in all human PRD, NKL and HOXL homeodomains.

In order to test the generality of the amino acid patterns derived in the previous section, we used the 356 homeodomain sequences classified by Holland et al. 6, which include the 140 human sequences used above. These sequences include genes classified as HOXL, NKL and PRD, plus other homeobox gene classes (LIM, POU, HNF, SINE, TALE, CUT, PROS, ZF, CERS). It should be noted, that in these analyses, the difficult to classify genes that were excluded above were now included, and these were classified as tentatively suggested by Holland et al. 6. Using amino acid sequences, a neighbour-joining tree (using pair-wise deletion, as implemented in the MEGA software; 17) was built with the 356 sequences, plus 918 homeodomain sequences from a variety of animal species. The latter 918 sequences were obtained from the file PS50071 available at PROSITE 10 after removing all non-animal sequences from the file. Sequences from clusters that are supported by a bootstrap value of 80% or higher, and that include at least one sequence that has been classified by Holland et al. 6, as belonging to a given class or subclass, were classified as belonging to that class or subclass. Using this phylogenetic argument 202 sequences could be classified as HOXL, 204 as NKL, 200 as PRD, 66 as LIM, 83 as POU, 15 as HNF, 23 as SINE, 66 as TALE, 20 as CUT, 3 as PROS, 94 as ZF, and 10 as CERS. Nevertheless, 288 sequences remained unclassified.

Table 1 shows that HOXL 1 pattern is only found in genes classified as HOXL and thus is highly specific. It is also highly representative, because 96% of the HOXL sequences (194/202) used could be classified as such using this pattern. Therefore, the requirement to use one of the amino acid combinations implied by this pattern is a derived feature that appeared in the HOXL lineage early in animal evolution. The HOXL sequences that do not show the expected pattern are: Drosophila melanogaster AbdB, Drosophila melanogaster btn, Strigamia maritima Hox3b, Gallus gallus HMD2 (PROSITE annotation), Danio rerio HXABA (PROSITE annotation), and Salmo salar HXB2 (PROSITE annotation). These sequences do not form a closely related subgroup of sequences. For instance, the Drosophila melanogaster Abd-B and btn genes belong to two different HOXL gene families. Moreover, other genes belonging to these families do show the HOXL 1 amino acid pattern. The Fugu rubripes HXDBB (PROSITE annotation) and the Gallus gallus HXB8 (PROSITE annotation) sequences are incomplete, and thus it is not possible to determine whether they show this amino acid pattern. In addition to the 194 known HOXL genes showing the HOXL 1 pattern, 178/288 'unclassified' sequences also had this pattern (Table 1). Since the HOXL 1 pattern is very specific to the HOXL subclass, it is very likely that the 178 unclassified sequences that show this pattern are HOXL sequences.

The HOXL 2 pattern is not as specific as the HOXL 1 pattern, since it is observed in about 46% of the genes belonging to the POU class and in a few NKL (0.5%) and PRD (8%) genes. All 16 PRD genes showing the HOXL 2 pattern belong to the Pax4/6 gene family, thus it is likely that this is a case of convergent evolution. The human POU homeodomains showing HOXL 2 pattern are POU1F1, POU3F1, POU3F2, POU3F3, and POU3F4. The latter four genes are closely related, although the POU1F1 gene is distantly related to those genes 6. Nevertheless, this is also likely a case of convergent evolution, since the alternative hypothesis (that the need to use one of the amino acid combinations implied by the HOXL pattern 2 is an ancestral feature) implies many independent losses. Therefore we argue that the necessity of using the amino acid combinations implied by HOXL 2 pattern is a derived feature that appeared early in the HOXL lineage. Only three sequences (1.5%) classified as HOXL using a phylogenetic argument, do not show the HOXL 2 pattern, namely: Drosophila melanogaster exex (the Drosophila Mnx1 orthologue; 18), human MNX1 and mouse Mnx1 (PROSITE annotation). Mnx1 genes are peculiar in other ways. For instance, they show both the NKL and HOXL 1 pattern (see below).

The NKL pattern is highly specific but not widely representative of all NKL genes. The pattern is found in 57% of the sequences classified as belonging to the NKL class (116/204; Table 1), but in only ten other classified genes (three HOXL and seven SINE). In addition, 52 of the 288 phylogenetically 'unclassified' genes show the NKL pattern, and of these 36 have already been classified as belonging to the NKL class by PROSITE (data not shown). The three HOXL genes with the NKL pattern are the three Mnx1 genes (from Drosophila melanogaster, Mus musculus, and Homo sapiens). Mnx1 genes also show the HOXL 1 pattern that is highly specific for HOXL genes. Thus, it is unlikely that Mnx1 genes have been misclassified. Therefore, the presence of the NKL pattern in Mnx1 sequences may be the result of convergent evolution.

The seven SINE sequences where the NKL pattern is found are SIX3 from Oryzias latipes, Gallus gallus, Mus musculus and Homo sapiens and SIX6 from Gallus gallus, Mus musculus and Homo sapiens. All SIX human genes cluster together with a high bootstrap value (97%; 6). Nevertheless, the NKL pattern is only observed in the SIX3 and SIX6 genes, two closely related genes. It is thus, likely a case of convergent evolution. The requirement to use one of the amino acid combinations implied by the NKL pattern is a derived feature that appeared likely early in the NKL lineage.

Most of the classified sequences where the NKL pattern is not found have been annotated by Holland et al. 6, and PROSITE as belonging to the NANOG, NOTO, VENTX, EN, DLX, and BARX, families. The EN, DLX and NOTO genes are difficult to classify. Thus, information on these genes was not used to derive the NKL amino acid pattern.

The PRD pattern is highly specific, being found only in sequences classified as PRD, and is also highly representative being found in 86% of the sequences classified as PRD. The PRD genes that do not show this pattern are D. melanogaster OdsH, otp, PHDP, Ptx1, IP09201, HOPX genes from Danio rerio, Homo sapiens, Bos taurus, Rattus norvegicus, Mus musculus, Sus scrofa, and Gallus gallus (PROSITE annotation), human PAX2, PAX5 and PAX8 (which have partial homeobox sequences), Hydra vulgaris Dmbx, OTP genes from Heliocidaris erythrogramma, Heliocidaris tuberculata, Lytechinus variegates, Paracentrotus lividus, and Saccoglossus kowalevskii (PROSITE annotation), OTX from Strongylocentrotus purpuratus (PROSITE annotation), ALX from Strongylocentrotus purpuratus (PROSITE annotation), ANF genes (PROSITE annotation) from Gallus gallus, and Xenopus laevis (two genes), and the PROP gene from Canis familiaris (PROSITE annotation). The PRD sequences that do not show this pattern are not related in any particular way. Therefore, the absence of this pattern in these sequences is likely the result of several independent losses. Therefore, it seems likely that the requirement to use one of the amino acid combinations implied by the PRD pattern is a derived feature that appeared early in the evolution of PRD genes.

Since the PRD-specific pattern is highly specific, it is likely that the 14 unclassified sequences that show this pattern are also PRD sequences. According to PROSITE classification these genes are Gsc from Danio rerio, Xenopus laevis (two genes), Gallus gallus, Mus musculus, Saguinus labiatus, Gorilla gorilla, Pongo pygmaeus, Pan paniscus, and Pan troglodytes, ALX4 from Mus musculus and Bos taurus, and UNC-4 and ceh-36 genes from Caenorhabditis elegans (PROSITE classification). According to PROSITE, 13 of these genes belong to the PRD class; the status of the ceh-36 gene is unknown.

The ANTP 1 pattern is found in most HOXL (98%) and NKL (98%) gene sequences and is almost absent in PRD (0.5%) gene sequences. Nevertheless, such a pattern is also found in sequences from genes belonging to all other classes except the PROS class (Table 1). It should, however, be noted that the sample size of PROS genes is very small (only three sequences). The ANTP sequences that do not show this pattern are Drosophila melanogaster eve, lbe, and lbl, Homo sapiens NOTO, Heterodontus francisci Evx2 (PROSITE annotation), Gallus gallus Hme1 (PROSITE annotation), Fugu rubripes hxdbb, and Caenorhabditis elegans vab7 and hm31 (PROSITE annotation). The HXB8 gene from Gallus gallus (PROSITE annotation) is incomplete, and thus it is not possible to determine whether it shows this pattern. The HOPX sequence (PROSITE annotation) from Danio rerio is the only PRD sequence that shows ANTP 1 pattern.

Although the ANTP 1 pattern is short (only three amino acid positions long), the broad distribution indicates that, very likely, it is not the result of convergent evolution. It is more likely that the ability to use the amino acid combinations implied by this pattern is an ancestral feature of homeobox containing genes. For some likely functional reason, ANTP genes have retained such a pattern, in contrast with PRD genes (the outgroup to ANTP genes) where only 0.5% of all sequences show this pattern.

The ANTP 2 pattern is found in most HOXL (98%) but in only 68% of the NKL sequences. This pattern is not found in sequences from other classes. It is thus, highly specific, although less representative across the NKL subclass of ANTP class genes. The HOXL genes that do not show this pattern are (PROSITE annotations) Hmd2 from Gallus gallus, hxaba from Danio rerio, hxb2 from Salmo salar and hxbb from Fugu rubripes. These sequences are not related in any particular way. It is not possible to determine whether the Hxb8 gene from Gallus gallus (PROSITE annotation) shows this pattern, since this is a partial sequence. NKL sequences not showing the ANTP 2 pattern belong to the Dlx (distalless), En (engrailed) and Noto gene families that have been previously found to be difficult to classify. Hence, information on these genes was not used to derive the ANTP 2 pattern. Once again, genes from these families stand out as oddities (it should be noted that they also do not show the NKL pattern as well; see above).

The ANTP-PRD pattern identifies most HOXL (99%), NKL (98%) and PRD (95%) sequences used. Nevertheless, it is also present in all SINE sequences and in 79% of the LIM sequences. The 16 ANTP plus PRD sequences that do not show the expected pattern are the HOPX genes from Homo sapiens, Mus musculus, Rattus norvegicus, Sus scrofa, Bos taurus, Gallus gallus, and Danio rerio, the HM31 gene from Caenorhabditis elegans, and the Artemia sanfranciscana HMEN gene. It should be noted that HOPX are difficult to classify, that in the latest revision 6 were classified as PRD genes, and these were not used to derive the amino acid patterns being tested. The PAX2, PAX5 and PAX8 sequences from Homo sapiens, as well as the HXB8 gene from Gallus gallus, the HXDBB gene from Fugu rubripes, and the DLX2 and DLX4 genes from Eleutherodactylus coqui, are partial, thus it is not possible to determine whether they show the ANTP-PRD amino acid pattern.

Given the failure to identify about 43% of the classified NKL sequences using the NKL pattern, we performed additional analyses to see whether the NKL pattern could be refined to accommodate the members of the NANOG, NOTO, VENTX, EN, DLX, and BARX, families. The results are shown in Table 2. When the NKL and refined patterns are considered, 98% of all NKL sequences could be classified as such. It should be noted that these refined NKL patterns are almost completely absent in the HOXL, PRD, LIM, POU, HNF, SINE, TALE, CUT, PROS, ZF and CERS classes. The 11 HOXL sequences that contain the EN pattern but not the NKL pattern all belong to the Hox1 gene family. This happens because the EN pattern allows for an asparagine at position 41. Five out of the six genes that show the NANOG pattern but not the NKL pattern belong to the HOXL Gsx gene family. These sequences are found because the pattern allows for a Lysine at position 43. The 8 HOXL sequences that show a hit when using the VENTX pattern but not the NKL pattern all belong to the Gbx family. This happens because a Valine is now allowed at position 43. The three CUT genes that show the EN and VENTX patterns are the Homo sapiens ONECUT1 gene and the orthologues in Mus musculus and Rattus norvegicus. These sequences show a hit because both patterns allow for an Isoleucine at position 45.

Only a few non-bilaterian homeobox sequences (those listed in 6) are contained in the data sets used. Therefore, we collected from the NCBI database a set of 251 non-redundant non-bilaterian homeobox sequences that encompass the regions where the amino acid patterns here reported are located, and that showed the ANTP-PRD pattern derived above (see Appendix). Most bilaterian ANTP and PRD sequences show this ANTP-PRD pattern (Table 1); furthermore, this pattern is only found in ANTP, PRD, LIM and SINE sequences (Table 1). Therefore, by imposing the presence of the ANTP-PRD pattern we hoped to enrich the data set for non-bilaterian ANTP and PRD sequences. Table 3 summarizes the results. As expected, of the retrieved gene sequences only one seems to belong to gene classes other than ANTP, PRD, LIM and SINE.

Most non-bilaterian genes showing the full NKL or PRD signature have already been classified as such. Furthermore, the ANTP Demox class found in Demospongiae and apparently absent in all other animals 19, is characterized by the presence of the ANTP-PRD, and ANTP 1 and 2 patterns, as expected for ANTP genes that do not belong to the HOXL or NKL lineage (Table 3). Although there are no true Hox genes in sponges 20, 13 sequences show the full HOXL signature. Interestingly, all such sequences are non-annotated, thus they must be hard to classify gene sequences. One Ephydatia fluviatilis gene has been labeled as the Msx (NKL) gene. Our pattern analyses show that the corresponding sequence shows a pattern exclusively found in the Mnx1 (HOXL) gene family, thus, also suggesting that non-bilaterian sequences with full signatures may be wrongly identified. About 53% of all non-bilaterian sequences used did not show an easily recognizable signature. We conclude that the reported amino acid patterns must be cautiously used when applied to sequences from non-bilaterian animals.

Key amino acid changes that occurred during the evolution of homeobox containing genes may be revealed as changes affecting a single amino acid position. Thus, we compared, at each amino acid position, the large HOXL (202 sequences), NKL (204 sequences) and PRD (200 sequences) data sets described above, for the following chemical properties: size, charge, polarity, hydrophobicity, aromaticity, and aliphapaty. For completeness, the 178, 52 and 14 sequences that could not be classified using a phylogenetic argument but that show amino acid patterns typical of HOXL, NKL and PRD genes were also used.

Chemical properties observed in more than 99% of the sequences belonging to one class, and observed in less than three-quarters of sequences belonging to another class are shown in Tables 4, 5, and 6. Rules are described relative to the inferred common ancestor situation of the two categories being compared. For instance, when comparing PRD and ANTP class genes the rule "not negatively charged at position 27" means that it was inferred (by comparison with LIM genes) that the common ancestor to these two gene classes could use a negatively charged amino acid at position 27; in this example, ANTP genes subsequently almost completely lost the ability to use a negatively charged amino acid, but 94% of PRD genes retained the use of such an amino acid at this position (Table 4). An almost complete change in amino acid usage regarding polarity and hydrophobicity is also observed for amino acid position 30. Furthermore, about 30% of the PRD genes show a charged amino acid at position 50. HOXL and NKL genes never use charged amino acids at this position, and this is also the case for LIM genes, here used as an outgroup to ANTP and PRD genes. Therefore, the possibility of using a charged amino acid at position 50 seems to be a derived feature that appeared in the PRD lineage. This amino acid position has been previously used to sub-classify PRD genes into three categories 21. A charged amino acid is found in sequences from Caenorhabditis elegans, Strongylocentrotus purpuratus, and Hydra vulgaris, among others, thus this switch is an event that happened early in the evolution of PRD genes.

Genes belonging to the HOXL lineage show derived constraints at amino acid positions 14, 15, 28, 33, 47 and 54, relative to the inferred ancestral state for ANTP class genes (the PRD data set was used as an outgroup; Table 4). Only three sequences classified as HOXL do not follow the general pattern for these genes, namely, the HMA2 gene (PROSITE annotation) from Helobdella triserialis that shows an aromatic amino acid at position 14, and the HXB8 and PAL1 genes (PROSITE annotation) from Sus scrofa and Caenorhabditis elegans, respectively, that show small amino acids at position 28. No derived single amino acid constraints were found for the NKL lineage.

Nematostella vectensis: 32263856; 32263873; 74039490; 74039492; 82621559; 82621587; 82621611; 82621663; 83356315; 156363224; 156363226; 156387518; 156397205

Ephydatia fluviatilis: 438584

Acropora millepora: 117581722; 117581724; 117581726; 117581728; Anemonia erythraea: 158936936; Ephydatia fluviatilis: 438585; 510584; Halichondria bowerbanki: 33641771; Nematostella vectensis: 32816231; 78190373; 82621509; 82621525; 82621533; 82621555; 82621557; 82621585; 82621591; 82621609; 82621625; 82621657; 82621665; 82621667; 110339021; 110339029; 110339121; 156353243; 156363798; 156367335; 156372762; 156397943; 168693291;Suberites domuncula: 34786940; 49659003; Sycon raphanus: 11066243

Anthopleura japonica: 144369330; 144369363; Hydra vulgaris: 2331219; 7635735; Nematostella vectensis: 82570553; 82621543; 82621677; 110339077; 110339133; 156376845; 156402818; 156406963; 156407174

Nematostella vectensis: 82621563; 82621631; 82621655; 156366927; 156402173

Nematostella vectensis: 82621579; 82621647; 82621651; 110339063; 156388033; 156400908; 156407394

Nematostella vectensis: 110339027

Baikalospongia intermédia: 62238212; Ephydatia fluviatilis: 1438870; Ephydatia muelleri: 3184520; Hydra vulgaris: 7635740; Nematostella vectensis: 82621517; 82621615; 82621619; 82621623; 82621635; 82621653; 156375827; 156375891; 156376334; 156401296; 156402692; Podocoryne carnea: 28188799; Potamolepis sp.: 62238208; Suberites domuncula: 49659001

Anthopleura japónica: 144369334; Cladonema californicum: 9652040; Hydra vulgaris: 3021450; 3021452; 6503072; Nematostella vectensis: 38569883; 78190377; 82395396; 82395402; 82395404; 82395406; 82395408; 82395412; 82395414; 82570541; 82570557; 82570559; 82570563; 110339183; 110339213; 110339215; 110339225; 113120207; 156377162; 156615306;Tripedalia cystophora: 33391193

Acropora formosa: 228960; Acropora millepora: 7335704; 7595811; 7595813; 13506878; 117581730; 117581732;Anthopleura japónica: 144369323; 144369326; 144369366; Aurelia aurita: 50841484; Cassiopea xamachana: 4894653; 4894655; 4894659; Cladonema radiatum: 47155918; 47155920; 47155922; Eleutheria dichotoma: 1147626; 91982983; 91982989; Hydra littoralis: 2102728; Hydra magnipapillata: 630481; 630482; 89242120; 144369369; 144369375; Hydra viridis: 7120; 7124; 7130; 83763566; Hydra vulgaris: 4433647; 4838455; 7635731; 7635733; 7635737; 7635742; 9945022; Hydractinia symbiolongicarpus: 2980868; 83272159; Nematostella vectensis: 5081328; 32816237; 74039494; 78190375; 82395398; 82395410; 82570519; 82570521; 82570527; 82570529; 82570537; 82570539; 82570545; 82570549; 82570555; 82570561; 82621507; 82621511; 82621513; 82621515; 82621519; 82621523; 82621527; 82621531; 82621535; 82621539; 82621541; 82621547; 82621549; 82621551; 82621565; 82621567; 82621571; 82621573; 82621575; 82621589; 82621595; 82621601; 82621603; 82621605; 82621613; 82621621; 82621627; 82621629; 82621633; 82621637; 82621643; 82621645; 82621671; 82621675; 83356317; 99030986; 110338989; 110339023; 110339061; 110339099; 110339101; 110339115; 110339171; 110339191; 110339211; 110339247; 113120203; 156356964; 156358580; 156361301; 156361303; 156364678; 156368221; 156371265; 156371439; 156372678; 156374167; 156377158; 156377160; 156377164; 156389076; 156390886; 156393340; 156395806; 156396978; 156398319; 156407912; Podocoryne carnea: 976094; 6118056; 6465862; 7649930; 9964019; 47155914; 62002543; Sarsia sp.: 9988771; Scolionema suvaense: 158936914; Suberites domuncula: 49659005; Trichoplax adhaerens: 38565482