Author :Scott Lloyd and Quinn O. Snell
Publisher/Conf : 2008 International Conference on Reconfigurable
Computing and FPGAs
Speaker : De-yu-Chen
Data : 2010.5.5
1
Most efforts to accelerate bio-sequence applications with
hardware have focused on database searches. Given a query
sequence, an entire genetic database is scanned looking for other
sequences that are similar.
Accelerating a database search is a simpler problem than
alignment. Only the score for the comparison is computed by
hardware in the forward scan, whereas alignment requires
traceback in addition to the forward scan.
The sequence comparison problem can be mapped to a linear
systolic array of Processing Elements (PEs) requiring
O(min(m,n)) space.
2
Unlike most acceleration methods that focus on sequence
comparison, this research describes and evaluates a spaceefficient, global sequence alignment and architecture that
includes traceback for implementation on reconfigurable
hardware.
The algorithm is based on dynamic programming (NW
algorithm), but partitions the problem into slices for the
FPGA hardware Since sequence lengths are often longer
than the number of PEs available in a systolic array.
3
Let H denote the DP score matrix, The matrix fill occurs in a
scan from upper left to lower right because of dependencies
from neighboring elements.
4
During the forward scan, a pointer p belong to {DIAG,
ABOVE, LEFT} indicates the current selection of the MAX
function in Equation 1.
The value of p is saved to the traceback matrix T, thus T[i, j] =
p, Following the forward scan, traceback proceeds from T[m,n]
to T[0,0], thereby determining the best alignment. The result is a
list of edit operations e belong to {SUBSTITUTE, INSERT,
DELETE}.
5
The forward scan consists of two fundamental scan procedures
ScanPartial and ScanFull.
1. ScanFull(A, B, x, y, T) :
The ScanFull procedure does not partition the DP matrix and
produces a full matrix T of traceback pointers that refer to
adjacent elements of H.
In the other word, ScanFull is run NW algorithm.
6
2. ScanPartial(A, B, x, y, R) :
ScanPartial will compute H score matrix and R row pointer
matrix. Given that p indicates the heritage of element H[i , j],
the following recurrences for 1≦i≦m and 1≦j≦n determine R.
7
Example: S1=GCCCTAGCG
S2=GCGCAATG
R matrix
8
Only the designated columns of R are actually stored, which
correspond to the right-most columns of a slice.(H matrix not be
stored)
9
3. TraceFull(A, B, x, y, T, E) :
The TraceFull procedure alignment, T is traceback pointer
matrix and E is operations(SUBSTITUTE, INSERT, DELETE).
4. TracePartial(A, B, x, y, R, E)
10
Example: S1=GCCCTAGCG S2=GCGCAATG
Suppose PEs=4.
Step1: Call ScanPartial(S1, S2, 9, 8, R) :
Designated
columns
11
Step2: Call TracePartial(S1, S2, 9, 8, R, E) :
12
Step2: Call TracePartial(S1, S2, 9, 8, R, E) :
13
Step2: Call TracePartial(S1, S2, 8, 7, R, E) :
14
Step2: Call TracePartial(S1, S2, 8, 7, R, E) :
15
Step2: Call TracePartial(S1, S2, 8, 7, R, E) :
16
Step2: Call TracePartial(S1, S2, 8, 7, R, E) :
17
Compare with NW algorithm :
18
The global alignment accelerator is implemented using Qnet [9],
an open-source packet-switched network architecture similar to
DIMEtalk [18].
19
Three global alignment implementations are tested in the
evaluation:
1) A software-only version of the algorithm presented in this
paper.(NW algorithm)
2) A version accelerated by the FPGA
3) The Myers-Miller global alignment algorithm for an
additional point of reference.
Seq-Gen [1] produced varying lengths of test sequences ranging
from 128 to 16383 symbols for the evaluation.
20
21
22