仮説を立てて考えてみよう
Let's Hypothesize and Reason!
井上 克巳 Katsumi Inoue
アンドレイ・ドンチェスク Andrei Doncescu (LAAS-CNRS)
山本 泰生 Yoshitaka Yamamoto
宋 剛秀 Takehide Soh
- Automated hypothesis-finding through
deductively complete methods.
- Intelligent machines:
Thinking like human being.
being
- Automated discovery of scientific knowledge,
in particular biological knowledge.
-Induction of causal laws in action theories,
and applications to systems biology.
- Web-based ILP system.
Background
How Intelligent Machines Think ?
How Human Beings Think ?
Observation
Induction
Everyday Life
Business
Science
Induction
Prediction
Hypothesizing
and
Reasoning
Abduction
Abduction
Deduction
Hypothesis
Generation
Deduction
The genius people are able to mix these
three fundamental modes of reasoning.
Combination of
induction and
abduction
Diagnosis
Design
Characterization
Discovery
Verification
One of the most powerful theoretical
answers for the next generation of
Intelligent Machine (Inoue 2001,2004)
Logic and Computation
IE for Abduction
• SOLAR (Nabeshima, Iwanuma & Inoue 2003)
B: full
f ll clausal
l
l theory
th
E: conjunction of literals (￢ E is a clause)
H: conjunctions of literals (￢ H is a clause)
Abduction and Induction: Logic
Input:
B : background theory
E : examples ／observations
Output:
H : hypothesis satisfying that
Example: graph completion problem – pathway finding
Find an arc which enables a path from a to d.
1. B ∧ H ٧ E,
Axioms: [￢node(X),￢node(Y), ￢arc(X,Y), path(X,Y)].
[￢node(X), ￢node(Y), ￢node(Z), ￢arc(X,Y), ￢path(Y,Z), path(X,Z)].
2. B ∧ H is consistent.
[node(a)]. [node(b)]. [node(c)]. [node(d)].
[[arc(a,b)].
( , )] [[arc(c,d)].
( , )]
Inverse Entailment (IE)
a
c
Negated
Observation:
[￢path(a,d)].
Computing a hypothesis H can be done deductively by:
Production_field: [￢arc(_,_)].
B ∧￢E ٧￢H
SOLAR outputs four consequences:
[￢arc(a, d)] , [￢arc(a, c)], [￢arc(b, d)], [￢arc(b, c)] b
d
We have good tools for this inverse computation.
B
ILP
machine
E
H
IE for Induction
• CF
CF--induction (Inoue 2004: Yamamoto, Ray & Inoue 2007)
CF-induction is the only existing ILP system that is complete
for full clausal theories.
SAT-based Approach for Pathway Finding
ILP machine
This approach transforms pathways and biological rules into CNF
formulas. We can identify probable paths by finding minimal models.
Correspondence： 井上 克巳 （Katsumi Inoue）／ 国立情報学研究所 情報学ﾌﾟﾘﾝｼﾌﾟﾙ研究系 教授
TEL & FAX: 03-4212-2520
Email : [email protected]
推論による仮説発見とシステム生物学への応用
Inference-based Hypothesis-Finding for System Biology
Andrei Doncescu (LAAS-CNRS)
Taisuke Sato (Tokyo Inst.
Inst Tech)
Yoshitaka Yamamoto
Katsumi Inoue
Takehide Soh
Gabriel Synnaeve (TIMC-IMAG)
- Development of a framework for knowledge
discovery from biological databases using logicbased AI.
- Discover hidden rules in systems biology.
- Explain the relationships between causes
and effects from genotype to phenotype.
- Clarification of the principles of hypothesis
formation and hypothesis evaluation and their
efficient implementation.
- Build generic models in biology,
Saccharomyces Cerevisiae and E. coli.
- Bridge between biologists and computer scientists
Modeling
GLUCOSE External
Phenomenological
g
ªINTRACELLULARE
ª
METABOLIC Analyzer
pools
models
PTS
-g6p
g1p
+fdp
Glucose
ATP
Glucose6-P
glycerol
synthesis
+ ATP
Fructose-P NADH,H
ATP
glycerol
GlycerolP
Sedoheptulose7 P TrioseP
+
NADH,H NADH,H+ H20 + 4H+
ATP
Yield PHB
Erythrose4P
Glycerate3P
NAD
FADH2
PEP
4
MurSynth
+amp
+adp
PFK
G3PDH
dhap
TKb
TIS
TrpSynth
ª Genomic Response
serine
synthesis
NADPH,H+ ATP CO2
CO2
Acetyl CoA
CO2
0
Malate
50
40
30
20
10
0
0.91.0
0.70.8
0.6
0.5
0.4
0.3
0.2
0.1
0.0
ª Simulation
SH-CoA
aKglu
aromatic
amino acid
synthesis
f6p
3pg
pep
PEPCxylase
+fd
+fdp
MetSynth,
TrpSynth
NADH,H+
2
met, trp
synthesis
NADPH,H+
CO2
Structural Models
ª Stoichiometric Modeling
Approaches
Discretization
with
continous
HMMs
Problem Settings
1. Predicting the inhibitory effect of toxins including
hydrazine with qualitative modeling
2 Explaining dynamic behavior of E.
2.
E coli pathways with
kinetic modeling
Pathway Model
(Qualitative / Kinetic)
Pathway Data
(from lab + KEGG)
Observations
BDD--EM
BDD
Katsumi Inoue, Taisuke Sato, Masakazu Ishihata, Yoshitaka Kameya
and Hidetomo Nabeshima, “Evaluating Abductive Hypothesis using
an EM Algorithm on BDDs,” in: Proc. of IJCAI-09, 2009.
-atp
atp
Synth2
ile, ala, kival, dipim
synthesis
PDH
accoa
Flux: the rate at which a material is processed
through a metabolic pathway.
- Explaining and predicting metabolic pathways
Background
Knowledge
PK
Metabolic pathway: sequences of enzyme-catalyzed
reaction steps, converting substrates to a variety of
products to meet the needs of the cell.
Goals
- Hypothesis generation by SOLAR
- Hypothesis evaluation by an EM algorithm on BDDs
- Modeling with discretization and the Michaelis-Menten equation
+fdp
oaa
+amp
pyr
ª Metabolic Constraints
Hypotheses
B∧H٧E
DAHPS
pgp
Synth1
cho, mur
synthesis
Approach
SOLAR
e4p
gap
GAPDH
2pg
Suc-CoA
CO2
GTP
SH-CoA
NADH,H+
ª Optimization
nucleotide
synthesis
ENO
Acétate
IsoCitrate CO
FadH2 Succinate
RPPK
gap
PGluMu
ANABOLISME
NADH,H+
Fumarate
rib5p
sed7p
TA
PGK
SerSynth
1/2 O2
3 H+
HS-CoA
HS
CoA ATP
NADPH,H+ ATP
Citrate
OAA
R5PI
TKa
-pep
fdp
HS-CoAATP
NADH,H+
CO2
1
ribu5p
Ru5P
xyl5p
Pyruvate
2
-nadph
p
PGDH
f6p
tryptophan
synthesis
1/2 O2
H20 + 2H+
FAD
ATP
3
-atp
p
6pg
-6pg
ALDO
Pentose P
5
PGI
mureine
synthesis
-nadph
p
G6PDH
g6p
G1PAT
polysaccharide
synthesis
2NADPH,H+
CO2
PGM
The best
hypothesis
selected by
BDD-EM
Pathway
Analysis
Best
Hypotheses
Correspondence： 井上 克巳 （Katsumi Inoue）／ 国立情報学研究所 情報学ﾌﾟﾘﾝｼﾌﾟﾙ研究系 教授
TEL & FAX: 03-4212-2520
Email : [email protected]