Genomics algorithms on digital NISQ accelerators - 2019-01-25
1. Genomics Algorithms
on digital NISQ accelerators
25th Jan 2019
Universitat Politècnica de València
Aritra Sarkar
aaw.reet.tro syor.kaar
PhD candidate, QuTech
Delft University of Technology
2. QGS Roadmap
Theoretical QGS
Perfect qubits
Many qubits
Theoretical QGS
Perfect qubits
Less qubits
Integrated QGS
Noisy qubits
Less qubits
Hardware QGS
Noisy qubits
Less qubits
Supremacy QGS
Noisy qubits
Less qubits
Useful QGS
FT qubits
Many qubits
NISQ: Noisy Intermediate-Scale Quantum
Phase II/IIIPhase I
3. Quantum accelerator for genomics
Quantum Complexity Theory
Quantum Algorithms
Computing Applications
Architecture-aware Implementations
2-40 EB/year
Genomical Big Data
QISA0
cqIR
5. What it is (not)?
Quantum
Biology
- “if evolution is smart enough to create a creature who understands QM, it must be using it for itself”
naturally occurring QM phenomena advantages, not necessarily for Computational purpose
e.g. photosynthesis, navigation in birds, neurons firing, … (sense of smell, emotions, past life, etc.…. keeps getting weirder)
Quantum
Genomics
Quantum-mechanical
Sequencing
Quantum-accelerated
Analysis
Sequencing
Gen2
NGS
Gen3
SMS
Gen1
Illumina
Roche 454
~100 bp
parallelism
high yield
Sanger
~1000 bp
Pacific Biosciences
Oxford Nanopore
~10000 bp
Overlap
Layout
Consensus
Pairwise
alignment
de Bruijn
k-mer
Analysis
Sorting
Deduplication
Variant
Calling
Reconstruction
De novoAb initio
(reference-based)
alignment/mapping
(reference-free)
assembly
Exact Heuristic
Approximate
Optimal
6. …
for each short read in sample:
do:
find index in reference genome
assess answer
while (result not satisfactory)
save short read matched index
reconstruct sequenced genome
…
Quantum accelerator
QASM
Simulator
multi-qubit regime target algorithm
current techs. have ~50 physical qubits
current Q Processor designs are not well scalable
exponentially difficult to simulate qubits
large planar topology yet to be implemented
full connectivity to specific topology can be compiled
number of gates related to total decoherence of result
gate fidelity guarantee with QEC codes
universal set to allow full domain exploration
Unlimited
Qubits
Unlimited
Gates
space complexity is a
critical design parameter
~ 50 bound for
feasible QX simulation
full connectivity
(complete graph)
time complexity is a
critical design parameter
Gate Fidelity = 1 (no errors)
available gates
(σX/Y/Z, H, CX, CZ, Rθ, Toffoli)
10. 10
• Map-to-reference alignment vs. De-novo assembly - right candidate for near-term Q acceleration?
– Try both for now. More on that to follow...
• Throughput on CPU vs GPU?
– Compare Q algorithm with single-core single-threaded clock cycles
• Will the developed algorithm solve a real bioinformatics problem faster than classical HPC?
1. Make the Q algorithm work for a small artificial sequence (algorithm functionality check)
2. Make the Q algorithm work for a real GS problem (need not be new or faster than sota algorithms)
3. Mapping and Error Correction
4. Step 1 in physical Q hardware
5. Step 2 in physical Q hardware
6. Make the Q algorithm work for a real GS problem (faster than classical algorithm)
7. Make the Q algorithm work for a new GS problem (inaccessible by classical methods)
• How many qubits are ok in NISQ-era?
– Attack from opposite perspective. What is the minimum number of qubits that are required to solve QGS?
• How to load reference genome for each run more effectively from classical data?
– Open: Pipelining or partial cloning. After algorithm’s functional design.
• How accurate or fast should the algorithm be?
– Note: Simulation time scales exponentially with operations (not true for Q hardware)
FAQs
Q supremacy
PhD scope
Algorithm
11. 11
Genomics optimization
Max-Cut
Hamiltonian Path
Hamiltonian Cycle
Vertex Cover
Shortest Common
Superstring
Clique
Boolean
Satisfiability
Knapsack
Travelling
Salesman Problem
Vehicle Routing
ProblemGraph Colouring
QAOA
VQE
De novo
Sequencing
Smith-Waterman
Algorithm
Independent
Set
…
for each short read from classical memory:
map to Hamiltonian Cycle
map to Vertex Cover
map to Max-Cut
convert to quantum compatible datatype
do:
load data in quantum memory
run QAOA
run VQE
measure quantum state
find solution of VQE
find solution of QAOA
find solution of Max-Cut; Vertex Cover; Hamiltonian Cycle
assess solution of Genomics problem
while (result not satisfactory)
reconstruct sequenced genome
…
13. 13
Near-term algorithms
• Peter Shor’s estimates
– Without QEC, Shor’s algorithm needs ~5k qubits to factor cryptographically significant numbers
– With error correction, ~1 million
– ~100 millions gate operations
• Near-Term Quantum Algorithms
– runs on few qubits (low depth circuits) without extensive QEC (small-codes)
– enough qubits to just store the problem (hard to do better)
– still solve useful problems with local constraints
– Adaptable optimization algorithms (easy to map to problem)
• Genetic Algorithm / Evolutionary Programs
• Simulated Annealing
• Deep Learning
• QAOA: The Quantum Master Algorithm
https://www.bcg.com/en-ca/publications/2018/next-decade-quantum-computing-how-play.aspx
14. 14
Variational Quantum Eigensolver
• Quantum/classical Hybrid algorithm
– Parameterised quantum subroutine is run within a classical optimization loop
– Prepare the quantum state | ൿ𝜓 Ԧ𝜃 , often called the ansatz
– Measure the expectation value ൻ𝜓 Ԧ𝜃 ℋ ൿ𝜓 Ԧ𝜃
• Variational theorem
– Expectation value ℋ ۧ|𝑎𝑛𝑠𝑎𝑡𝑧 ≥ λ1 (smallest eigenvalue; lowest energy; ground-state)
• Find an optimal choice of real-valued parameters Ԧ𝜃 such that the expectation value is minimised
• Heuristic
– No general recipe of ansatz definition that universally works well for all VQE problem
• Ansatz Learning
– Selection of the initial state is arbitrary
• Gained popularity as in some cases it is resistant to a quantum gate noise
– However not to a measurement noise
ℋ ≡ Hamiltonian (not Hadamard, in this context)
15. 15
Quantum Approximate Optimization Algorithm
• Approximate algorithm
– NP-Hard problem
– Polynomial-time solution every instance with guaranteed quality
– QAOA is interesting because of its potential to exhibit quantum supremacy
• Structure
– 2 parameterized Ising-type Hamiltonian
• Cost function (problem soft constraints)
• Driver/Mixing function (solution space hard constraints)
– for QA, Hm is fixed by hardware
– Classical parameter optimizer
• Rigetti Grove’s implementation of the QAOA uses the VQE module backend
– pyQAOA package contains separate modules for each type of problem instance: Max-Cut & graph partitioning
16. 16
NP-hard problems
• Vertex Cover: 𝑆 ⊆ 𝑉 (vertex set) such that each edge of the graph is incident to at least one vertex in 𝑆
• Max-Cut: 𝑆 ⊆ 𝑉 (vertex set) such that the number of edges between 𝑆 and ҧ𝑆 is as large as possible
– Analytically: at least one of the solutions of Max-Cut will be the Minimum Vertex-Cover (Min Vertex-Cover has Max-Cut)
– QAOA identifies the ground state of the Hamiltonian by evolving from a reference state
• The reference state is generated by a Hadamard gate on each qubit from all zero state
0
011110
100001
543210
1
2
34
5
111110
000001
101110
010001
Iteration 1 Iteration 6
17. 17
Variational Quantum Search
• Variationally Learning Grover’s Quantum Search Algorithm
• How this performs for approximate search?
Grover search vs. VQS
1 soln. max prob. w.r.t. qubits
18. 18
Hybrid C/Q programming in OpenQL
• Programming Platform Support
– Classical instructions
– Target classical processor
– Real-time parameter specification
– Unitary decomposition
• Survey of available libraries
– Tutorial for each platforms (QWorld Jupyter notebooks)
• Currently QISKIT (IBM) and Forest (Rigetti) done
– Derive requirement for OpenQL
19. 19
Current To-Do Summary
• Algorithm development (primary focus)
– Design a Variational Approximate Quantum Search Algorithm for DNA sequence alignment
– Map De novo Sequencing problem to QAOA
– Can the ansatz be evolved or learned?
• Platform development
– Implement Rigetti Forest’s maxcut, qaoa and vqe libraries and classical optimisers (like Nelder-Meed) on
OpenQL
• How many primitive gates and qubits does it get decomposed to?
• How do the classical and quantum parts interact?
– How to efficiently implement a QGS algorithm on the micro-architecture?
• e.g. format and procedure to load/store the DNA data
20. Genomics Algorithms
on digital NISQ accelerators
Aritra Sarkar
Quantum Computer Architecture Lab
QuTech and Department of Quantum & Computer Engineering
Delft University of Technology