# esl_dsqdata: efficient sequence database format The `dsqdata` module implements a binary sequence data format. It accelerates sequence data input in four ways, compared to Easel flatfile parsers in `sqio`: * __Asynchronous input.__ Disk and CPU resources are used concurrently, using POSIX threads. A "loader" thread does essentially nothing but read chunks of data. An "unpacker" thread does CPU work to prepare loaded sequence data chunks for consumption. If it takes time $R$ to read and $P$ to process the data, instead of overall time $R+P$, with asynchronous input we only need time $\mathrm{max}(R,P)$. * __Predigitization.__ Sequence data in the `dsqdata` format are already encoded in Easel digital sequence format. User-oriented error checking is done up front when the `dsqdata` file is created. * __Bit packing.__ Disk read time is typically rate-limiting in HMMER and other Easel-based programs, so minimizing data volume is critical. Sequence data are packed bitwise in 32-bit packets to reduce volume by a factor of 1.5 (protein) to 3.75 (nucleic). A packet contains six 5-bit residues (protein or degenerate nucleic) or fifteen 2-bit residues (canonical nucleic) and two control bits. * __Separate metadata.__ Sequence data and metadata (name, accession, description, taxonomy identifier) are stored separately in `.dsqs` and `.dsqm` files. This streamlines unpacking, because these data are handled differently. It also allows a deferred metadata read: sequences may be identified simply by index number during an initial processing sweep, and metadata can be loaded later by random access for a small number of targets of interest. The following table lists the functions in the `dsqdata` API. | Function | Synopsis | |--------------------------------|--------------------------------------------------------------| | `esl_dsqdata_Open()` | Open a digital sequence database for reading | | `esl_dsqdata_Read()` | Read next chunk of sequence data. | | `esl_dsqdata_Recycle()` | Give a chunk back to the reader. | | `esl_dsqdata_Close()` | Close a dsqdata reader. | | `esl_dsqdata_Write()` | Create a dsqdata database | ## dsqdata format's four files The format of a database `mydb` consists of four files: | File | Contents | Description | |-------------|-----------|--------------------------------------------------------------| | `mydb` | Stub | Human-readable information about the data | | `mydb.dsqi` | Index | Disk offsets for each seq in metadata and sequence files | | `mydb.dsqm` | Metadata | Name, accession, description, and taxonomy ids | | `mydb.dsqs` | Sequence | Sequences (digitized, packed) | The database is specified on command lines by the name of the stub file (`mydb`), without any suffix. For example, % myprogram mydb says to open `mydb`. The `esl_dsqdata_Open()` call then opens all four files. ## definition of dsqdata file formats ### the stub file An example stub file: ``` Easel dsqdata v1 x4019752601 Original file: refprot.fa Original format: FASTA Type: amino Sequences: 11432138 Residues: 4358716588 ``` The first line is the only line in the stub file that's parsed by the reader. Its text format matches `/Easel dsqdata v(\d+) x(\d+)/`. The first field is a version number for the format, $\geq$ 1. It is currently unused, but in the future we might need it to parse different versions of the format, if we need to update it. The second field is a 32-bit unsigned integer tag in the range 0..$2^{32}-1$. Each of the four files carries the same randomly generated tag. The tag is used to ascertain that the four files belong together in the same database, as opposed to one or more of them being inadvertently clobbered somehow by the user. After the first line, the rest of the stub file is ignored by the Easel reader, and can contain anything -- even your own notes, if you want to add any. The text here is the useful information that the Easel writer writes by default. ### the .dsqi index file The purpose of the .dsqi index file is to (quickly) give us data offsets (in bytes) we need to randomly access the metadata and sequence data for sequence number 0..nseq-1 in the .dsqm and .dsqs files, or for any specified range of sequences. The header of the binary index file consists of: | name | type | description | |--------------|------------|----------------------------------------------| | magic | `uint32_t` | magic number (version, byte order) | | uniquetag | `uint32_t` | random integer tag (0..$2^{32}-1$) | | alphatype | `uint32_t` | alphabet type code (1,2,3 = RNA, DNA, amino) | | flags | `uint32_t` | Currently 0. Reserved for future flags | | max_namelen | `uint32_t` | Maximum seq name length in metadata | | max_acclen | `uint32_t` | Maximum accession length in metadata | | max_desclen | `uint32_t` | Maximum description length in metadata | | max_seqlen | `uint64_t` | Maximum sequence length | | nseq | `uint64_t` | Total number of sequences in database | | nres | `uint64_t` | Total number of residues in database | The **magic** is used to check that the file is indeed a dsqdata format file, and to detect byte order swapping. Valid values for the magic version/byteorder number are: | value | derivation | description | |------------|---------------------|--------------------------| | 0xc4d3d1b1 | "dsq1" + 0x80808080 | dsqdata version 1 format | | 0xb1d1d3c4 | above, byteswapped | above, byteswapped | The **uniquetag** matches the tag seen in the other files. The dsqdata packet format is only defined for biological sequence alphabets. Valid integer values for the **alphatype** code come from a subset of the codes used in `esl_alphabet.h`: | value | `esl_alphabet.h` | description | |-------|------------------|-------------| | 1 | `eslRNA` | RNA | | 2 | `eslDNA` | DNA | | 3 | `eslAMINO` | protein | The unused **flags** field gives us some flexibility for future versions of the format. The maximum lengths of the names, accessions, and descriptions in the metadata file might someday be useful (in making allocations, for example) but they are currently unused by the Easel reader. Likewise, the maximum sequence length, total number of sequences, and total number of residues in the database may someday be useful (for making decisions about how to partition a parallel search, for example), but they are currently unused too. After the header, the remainder of the file consists of `nseq` records of type `ESL_DSQDATA_RECORD` (defined in `esl_dsqdata.h`): | element | type | description | |----------------|-----------|-------------------------------------------------------------------------| | `metadata_end` | `int64_t` | Position of terminal `\0` of metadata for seq i in .dsqm file, in bytes | | `psq_end` | `int64_t` | Position of final packet for sequence i in .dsqs file, in packets | Storing _end_ positions instead of _start_ positions allows us to determine lengths, without needing an n+1'th sentinel record, albeit at the cost of special casing what happens for the first sequence i=0. For example: ``` len[i] = (i == 0 ? r[i].end + 1 : r[i].end - r[i-1].end) start[i] = (i == 0 ? 0 : r[i-1].end + 1) ``` This is equivalent to treating `r[-1].end = -1`. Some of the Easel reader's code tracks a `last_end` variable for the end of the previous metadata or packed sequence field i-1, which is initialized to -1. This -1 boundary condition is why we use _signed_ `int64_t` types. Packet sequence endpoints are stored in units of unsigned 32-bit binary _packets_, not in bytes. To convert to a disk offset or a length in bytes you multiply by 4 (`sizeof(uint32_t)`). Keeping the size of the dsqdata files as small as possible is critical because the reading speed is limited by the raw size of the data. Therefore we don't store separate positions for the different metadata fields (name/accession/description/taxonomy id); only one position for all the metadata associated with sequence i. The reader reads all of it in one chunk, and parses it for the stored `\0` sentinels. For the same reason, we don't store any information about _unpacked_ sequence lengths, only the bare minimum of information that the dsqdata loader and unpacker need to locate, load, and unpack the packed data for any given sequence i. The unpacker determines the unpacked sequence length when it unpacks the data. ### the .dsqm metadata file The metadata file starts with two header fields, the same two that the index file starts with: | name | type | description | |-----------|------------|------------------------------------| | magic | `uint32_t` | magic number (version, byte order) | | uniquetag | `uint32_t` | random integer tag (0..$2^32-1$) | After the header, the remainder of the file consists of the following data for each sequence i = 0..nseq-1: | field | type | description | |--------------|-----------------------------|-----------------------------------------------------| | name | `char` array ending in `\0` | sequence name (1 word, no whitespace); mandatory | | accession | `char` array ending in `\0` | sequence accession (1 word, no whitespace); or "\0" | | description | `char` array ending in `\0` | sequence description line; or "\0" | | taxonomy id | `int32_t` | NCBI taxonomy identifier; or -1 | The name, accession, and description are variable length strings. The name and accession are single "words" with no whitespace (`\S+`). The description is one line, may contain spaces, but may not contain any newlines. All sequences must have a name, so `strlen(name) > 0`. The accession and description are optional; if they are not present, these are 0-length strings ("\0"). The taxonomy identifier is an integer in NCBI's taxonomy. Valid taxonomy identifiers are $\geq 1$. This field is optional; use a value of -1 to indicate unset. (I cannot find any documentation at NCBI on the maximum range of the taxid, nor can I find a clear statement of whether 0 is valid or not. 0 is currently unused in the NCBI taxonomy. 1 indicates the top level. That makes it look like it's safe to treat 0 as "unset" but it seems even safer to go with -1 and a signed integer. Unless NCBI ends up having more than two billion species. Currently there are about 1.8 million.) These names, types, and semantics match the corresponding fields in an `ESL_SQ`. ### the .dsqs sequence file The sequence file also starts with the same two header fields that the index and metadata files started with: | name | type | description | |-----------|------------|------------------------------------| | magic | `uint32_t` | magic number (version, byte order) | | uniquetag | `uint32_t` | random integer tag (0..$2^32-1$) | After the header, the remainder of the file consists of the packed sequences, with one packet array for each sequence i = 0..nseq-1. Each packet array ends with a specially marked sentinel packet. The packet format is described next. #### packet format Each packet is an unsigned 32 bit integer. The two leading (most significant) bits are control bits. Bit 31 signals EOD (end of data): the last packet in a packed sequence. Bit 30 signals the packet format: 1 for 5-bit, 0 for 2-bit. The remaining bits are the packed residue codes: ``` [31] [30] [29..25] [24..20] [19..15] [14..10] [ 9..5 ] [ 4..0 ] ^ ^ |------------ 6 5-bit packed residues ------------------| | | [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] | | |----------- or 15 2-bit packed residues ----------------| | | | "packtype" bit 30 = 0 if packet is 2-bit packed; 1 if 5-bit packed "sentinel" bit 31 = 1 if last packet in packed sequence; else 0 (packet & (1 << 31)) tests for end of sequence (packet & (1 << 30)) tests for 5-bit packing vs. 2-bit ((packet >> shift) && 31) decodes 5-bit, for shift=25..0 in steps of 5 ((packet >> shift) && 3) decodes 2-bit, for shift=28..0 in steps of 2 ``` Packets without the sentinel bit set are full. They unpack to 15 or 6 residues. 5-bit EOD packets may be partial: they unpack to 0..6 residues. The remaining residue codes are set to 0x1f (11111), indicating EOD within the packet. The only case in which a partial EOD packet encodes 0 residues is a zero-length sequence: there has to be at least one EOD packet. 2-bit EOD packets must be full, because there is no way to signal EOD locally within a 2-bit packet. It can't use 0x03 (11), because that encodes U/T. Generally, therefore, the last packet(s) of a nucleic acid sequence must be 5-bit encoded, solely to be able to use sentinel residues in a partial packet, unless the end happens to come flush at the end of a 2-bit packet. (If we ever needed to pack an alphabet of 2 or 3 residues, we could use 0x03 as a sentinel. This seems unlikely to ever happen, so I'm simply not going to include any code to read EOD 2-bit partial packets.) A protein sequence of length L packs into exactly P $= \mathrm{max}(1, (L+5)/6)$ 5-bit packets. A DNA sequence packs into P $\leq \mathrm{max}(1, (L+14)/15)$ mixed 2- and 5-bit packets. P $\geq 1$ because even a zero-length sequence ($L=0$) requires an EOD packet. A packed sequence consists of an integer number of packets, P, ending with an EOD packet. A packed amino acid sequence unpacks to $\leq$ 6P residues. All its packets are 5-bit encoded. A packed nucleic acid sequence unpacks to $\leq$ 15P residues. The packets are a mix of 2-bit and 5-bit. Degenerate residues must be 5-bit packed, and the EOD packet usually is too. A 5-bit packet does not have to contain degenerate residues, because it might have been necessary to get "in frame" to pack a downstream degenerate residue. For example, the sequence ACGTACGTNNA... must be packed as [ACGTAC][CGTNNA]... to get the N's packed correctly.