Advanced Al

Advanced Al for Longevity Genomics Dr. Ben Goertzel OpenCog Foundation Hong Kong Poly U Hanson Robotics Aidyia Limited Stevia First EFTA01125174 OpenCog • Open-source Al project aimed at Artificial General Intelligence • Integrated system aimed at controlling autonomous, generally intelligent agents • Components of OpenCog currently in use for various practical applications... • ... such as analyzing genomics data EFTA01125175 if ATLANTIS PRESS Atlantis Thinking Machines Sees Editor. K.- U. Kiihnberger Ben Goertzel Cassio Pennachin Nil Geisweiller Engineering General Intelligence, Part 1 A Path to Advanced AGI via Embodied Learning and Cognitive Synergy ATLANTIS PRESS [ A ng Machines .- U. Kiihnberger Ben Goertzel Cassio Pennachin Nil Geisweiller Engineering General Intelligence, Part 2 The CogPrime Architecture for Integrative, Embodied AGI EFTA01125176 OpenCog stores and manipulates knowledge in the form of complex graphs (weighted, labeled hypergraphs) Perception Action Feeling Nodes pixel at (100,50)1s MA) at 1.42:01, Sept 15, 2006 Joint_53_actuato is oil at 2:42:01, Sept. 15, 2006 Specific Objects, Composite Actions, Complex Feelings alse_arm_55 Abstract Concepts (some corresponding to named concepts, some not) raise arm EFTA01125177 1 PROCEDURAL KNOWLEDGI MOSES (probabilistic; evolutionary learning), hillclimbing' KNOWLEDGE, "deep learning" hierarchy of memory/ processing units KNOWLEDGE Probabilistic Lo ' Networks, concept blendin language comprehension generation ATTENTIONAL & INTENTIONAL KNOWLEDGE economic attention networks, adaptive goal hierarchy EPISODIC KNOWLEDGE internal world `mutation engine OpenCog features multiple cognitive algorithms, each acting on different sorts of knowledge within the common "Atomspace" dynamic knowledge store. The aim is to achieve high levels of general intelligence via "cognitive synergy" between the different cognitive algorithms, cooperating together to help an agent choose actions based on goals, context and experience. EFTA01125178 PROCEDURAL KNOWLEDGE MOSES (probabilistic evolutionary learning), DECLARATIVE KNOWLEDGE robabilistic Logi' Networks, oncept blending language omprehension EPISODIC EP I L KNOWLEDGE internal world imulation engine ‘ , So far, two of OpenCog's cognitive algorithms (MOSES and Probabilistic Logic Networks (PLN) are being used to help understand genomics data. In time, the full integrated OpenCog architecture will be used to serve the role of an "artificial scientist." EFTA01125179 OpenCog Al for Genomics: Two Examples • MOSES for identifying patterns differentiating supercentenarians from healthy —80 year olds based on SNP combinations • PLN (probabilistic logic networks) for using bio-ontologies to identify genes indirectly connected to the longevity phenotype, via a combination of genomic data and ontological knowledge EFTA01125180 phenotype classification of wholegenome sequenced samples with boolean models derived via MOSES supervised machine learning (Mike Duncan & Ben Goertzel) EFTA01125181 abstract • A boolean classification function was constructed using a novel supervised machine learning algorithm to categorize healthy from chronically ill geriatric subjects. From an evenly divided sample set of 783 subjects, a population of boolean functions consisting on average of 130 variables was evolved, with a mean out-of-sample accuracy of 0.851, compared to an in-sample accuracy of 0.860. • The same analysis pipeline was used to distinguish 17 super centenarians from a subset of the above data set consisting of 230 healthy geriatric females. Five significant functions were evolved, four binary and one with a single variable was evolved, with perfect out-of-sample accuracy. These functions consisted of 5 distinct SNP variants. EFTA01125182 meta optimizing semantic evolutionary search (MOSES) • MOSES is a 2 level genetic programing algorithm to search catagorization function space, allowing detailed exploration of multiple local fitness maxima. • Functions from the meta-level population are selected and "mutated" (their neighborhood in function space is searched). • Variants with improved fitness (better at categorizing) are simplified and returned to the meta-population. • In addition, integrated feature selection and multiple tunable search and fitness functions improve on standard genetic programing algorithms EFTA01125183 epistatic boolean classification models • MOSES evolves programs coded in a simple programing language called combo. • Binary variables are valued "0" if a sample is homozygous for the reference allele and "1" for any alternate alleles for a particular variant. A "true" value indicates "case" status. • an example boolean combo program applied to simulated genomic data: or( $rs1234 and( !Srs5678 or( or( $rs2468 $rs7531 ) and( $rs3142 $rs2001 )))) variable sample 1 sample 2 rs1234 ref (0) ref (0) rs2001 ref (0) alt (1) ref (0) rs3142 ref (0) alt (1) ref (0) It (1) rs2468 ref (0) rs5678 ref (0) rs7531 ref (0) program value control (false) case (true) EFTA01125184 supervised machine learning strategy • A cross validation strategy is used where the data set is randomly partitioned into training and testing sets at a ratio of 4:1. • Accuracy scores on training and testing sets are compared for each combo to assess over- fitting. • Ensembles of combos can be averaged to increase accuracy on out-of-sample data. • Ranked lists of variants can be constructed by counting variable occurrence in combo ensembles. EFTA01125185 whole genome variation data sets wellderly and illderly data set • from Scripps • 783 samples aged 80 and above • 342 males and 441 females • 397 wellderly cases and 386 illderly controls • 230 samples in wellderly female subset super centenarian data set • From Stanford • 17 samples aged 110 and above • 16 females and 1 male • 14 whites, 2 Latinas, and 1 African American EFTA01125186 wellderly vs. illderly MOSES analysis 150 sample variant load histogram 4103 1700 1900 4900 variant count eke 51O3 phenotype El Myth example combo variables are gemini db reference ID numbers • There were 900 combos with accuracies significantly greater than the case prevalence (p > 0.05, McNemar's test) • mean of 130 features per combo • The mean out-of-sample accuracy of combo ensembles was 0.884. • Means for all combos in each cross validation set: accuracy precision recall mean out-of-sample 0.851 0.863 0.843 mean in-sample 0.860 0.871 0.850 example combo 0.880 0.863 0.909 and(or(and(or(and(or(and(or(and(or(and(!$X106020 !$X168745) $X763139) !$X735449 !$X297852) and(or($X53710 $X297852 $X552647) !$X766840 $X808350)) or($X14463 $X766840)) and(or(and(or(!$X735449 $X54045) !$X67669) and(or($X135964 !$X808350) $X558377) $X434945) or(and($X735449 $X522743) and($X497883 $X702846) $X431028 !$X480341) or(and($X735449 $X552647) and(!$X135964 !$X194619))) and(or(and($X217849 $X256079 !$X808350) $X297852) !$X735449 $X695782)) or(and(!$X735449 !$X67669 $X434945) !$X427182)) and(or(and(or($X256079 $X808350) $X135964) !$X217849 $X434945) or($X735449 $X766840) $X67669 $X427182) and(or(and($X735449 !$X165425 $X808350) !$X67669 $X434945) or(and($X16581 !$X695782) $X128740 !$X135964 $X434945) !$X14463 $X217849) $X379084) or(and($X14463 $X217849) !$X67669 $X106020 $X379084 !$X427182 !$X480341) or(!$X379084 !$X379090)) a nd(or(and(or(and($X194619 !$X807580) and($X256079 !$X763139) $X217849 $X427182) or(!$X14463 !$X434945) !$X135964) and($X735449 $X431028) and(!$X106020 $X434945)) or(and(or(!$X165425 I $X217849) $X558377) and(or($X807580 !$X808350) $X695782) and(!$X735449 !$X14463) and($X480341 !$X807580) $X135964) or(and($X427182 $X497883) !$X67669) or(and(!$X558377 !$X807580) !$X14463 $X53710 $X427182) !$X297852 $X379084)) or(and(or(!$X379090 $X427182) !$X128740) and(!$X67669 $X135964) and(! $X480341 !$X577132) !$X808350)) EFTA01125187 wellderly vs. illderly top combo SNPs • The top 25 variants ranked by number of occurrences in the 10000 best combos from 10 cross validation runs. • "Category" indicates if variable is negated, i.e. if variant is negated in combo then category is "control" because combos are "true" for cases. Note variables can have different categories in different combos. • Alternate allele frequencies (AAFs) are shown for the data set and 2 reference genome sets: the Exome Aggregation Consortium (ExAC) and the 1000 Genomes. • Annotations are from geminil v1.7.O (ensembl v75, dbSNP v141, ExAC vO.3) 1. httocileemini readthedocs ore/en/latest/index html rs id combo count category cyto-band gene transcript data AAF adjusted ExAC AAF lk Genomes AAF rs10953303 2719 control chr7q22.1 MN ENST00000546213 0.211 0.236 0.198 rs1050348 2206 control chr6q21 LAMA4 ENST00000389463 0.405 0.665 0.758 rs6942733 2120 case chr7q22.1 ZAN ENST00000538115 0.255 0.235 ii 0.199 a1050348 1308 case chr6q21 LAMA4 ENST00000389463 0.405 0.665 0.758 rs6942733 1004 control chr7q22.1 ZAN ENST00000538115 0.255 0.235 0.199 a2243191 1002 control chr1q32.1 IL19 ENST00000270218 0.203 0.748 0.673 rs1688005 901 case chr19q13.12 FXYDS ENST00000588699 0.261 0.322 0.412 a4842978 900 control chr15q25.2 W0R73 ENST00000561447 0.432 NA 0.726 rs1977420 899 control chrllpl3 APIP ENST00000395787 0.359 0.404 0.457 a7905784 802 case chr 10p13 MCM10 ENST00000378694 0.152 0.118 0.064 rs7905784 801 control chr10p13 MCM10 ENST00000378694 0.152 0.118 0.064 a1381057 801 control chr3q13.33 POLQ ENST00000264233 0.327 0.722 0.745 rs1977420 797 case chrllpl3 APIP ENST00000395787 0.359 0.404 MEI 0.457 rs10953303 720 case chr7q22.1 ZAN 6N5100000546213 0.211 0.236 0.198 rs2228331 703 case chr2q37.3 GPC1 ENST00000264039 0.307 0.664 0.664 a4842978 701 case chr15q25.2 W0R73 ENST00000561447 0.432 NA 0.726 rs2397084 604 case chr6p12.2 IL17F ENST00000336123 0.102 0.069 0.033 a6587467 603 control chr1q44 OR2T6 ENST00000355728 0.309 0.721 0.773 rs912174 601 case chr9p24.3 KANK1 ENST00000382293 0.225 0.219 0.204 a671694 601 control chr7p22.2 SDK1 ENST00000404826 0.268 0.752 0.799 rs11895564 600 case chr2q31.1 ITGA6 ENST00000264106 0.294 0.281 0.252 a7386783 600 case chr8q24.22 OC90 ENST00000254627 0.277 0.729 0.737 rs11250 600 control chr4q13.2 CENPC ENST00000273853 0.392 0.657 0.701 a4802648 599 case chr19q13.33 ZNF473 ENST00000595661 0.221 NA 0.197 rs10277 598 case chr5q35.3 CSorf45 ENST00000376931 0.433 0.626 0.688 EFTA01125188 wellderly vs. illderly combo SNPs effects • Predicted translation effects of variants in top combos • Selected from feature set of 13,242 SNPs classified in gemini dbl as "high" and "medium" impact • Sequence Ontology (SO) impact classification determines gemini impact severity for variant filtering. • The combined annotation scoring tool (CAROL)2 combines SIFT3 and PolyPhen-24 nucleotide scores to predict SNP effect on translated protien. 1 httovileemini readthedors ordrn/latesticontent/databace cchema htmliteletails.of-the.imnart.and-imnart-ceveritv.rolumns 2. Lopes MC, Joyce C, Ritchie GRS, John SL, Cunningham F, Asim it 1, Zeggini E. A combined functional annotation score for non-synonymous variants Human Heredity (in press) 3. Kumar P, Henikoff 5, Ng PC. Predicting the effects of coding non-synonymous variants on protein function using the SIFT algorithm Nature Protocols 4(8):1073-1081(2009) doi: 10.1038/nprot.2009.86 4. Adzhubei IA, Schmidt S, Peshkin L, Ramensky VE, Gerasimova A, Bork P, Kondrashov AS, Sunyaev SR. A method and server for predicting damaging missense mutations Nature Methods 7(4):248-249 (2010) SNPdb ID rs1977420 rs10277 rs11250 rs 1688005 rs2228331 rs2397084 rs2243191 rs11895564 rs912174 rs1050348 rs7905784 rs7386783 rs6587467 rs 1381057 r5671694 rs4842978 rs10953303 rs6942733 rs4802648 gene APIP CSorf45 CENPC FXYDS GPC1 IL17F 1119 ITGA6 KANK1 LAMA4 MCM10 OC90 OR2T6 POLO SDK1 WDR73 ZAN ZAN ZNF473 gene name APAF1 interacting protein chromosome 5 open reading frame 45 centromere protein C FXYD domain containing ion transport regulator 5 glypican 1 interleukin 17F interleukin 19 integrin, alpha 6 KN motif and ankyrin repeat domains 1 laminin, alpha 4 minichromosome maintenance complex 10 otoconin 90 olfactory receptor, family 2T, member 6 polymerase (DNA directed), theta sidekick cell adhesion molecule 1 WD repeat domain 73 zonadhesin (gene/pseudogene) zonadhesin (gene/pseudogene) zinc finger protein 473 category case & control case control case [repeated 3 times] control case case case & control case & control case control [repeated 3 times] case & control [repeated 3 times] case ensembl transcript SO impact CAROL prediction ENST00000395787 ENST00000376931 ENST00000273853 ENST00000588699 ENST00000264039 ENST00000336123 ENST00000270218 ENST00000264106 ENST00000382293 ENST00000389463 ENST00000378694 ENST00000254627 ENST00000355728 ENST00000264233 ENST00000404826 ENST00000561447 ENST00000546213 ENST00000538115 ENST00000595661 missense variant [repeated 15 times] splice region variant missense variant missense variant splice donor variant Neutral (0.876) Neutral (0.773) Neutral (0.000) Neutral (0.705) Neutral (0.000) Deleterious (1.000) Neutral (0.000) Neutral (0.724) Neutral (0.000) Neutral (0.307) Neutral (0.380) Neutral (0.436) Neutral (0.670) Neutral (0.000) Neutral (0.200) nan Deleterious (0.999) Neutral (0.036) nan EFTA01125189 super centenarian vs. wellderly female MOSES input data sample variant load histogram 40 4560 5000 5500 variant count 6000 Plumb/Pe ■wendedy super centenarian • Super centenarian cases were matched to wellderly female controls to attempt to find SNPs associated with extreme longevity. • Feature set constructed from intersection of SNPs in case & control sets. • Cases have almost 30% more variants per sample than controls. EFTA01125190 super centenarian vs. wellderly female MOSES results • There were 5 combos with accuracies significantly greater than the case prevalence (p > 0.05, McNemar's test) • The out-of-sample accuracy for all significant combos was 1.0 1. and(!$rs17521570 $rs5905720) 2. and(!$rs2230681 $rs5905720) 3. and(!$rs557337 $rs5905720) 4. and($rs2230681 $rs5905720) 5. $rs5905720 gene name location transcript data ExAC2 AAF' AAF , 1kG3 AAF rs5905720 MAGIX MAGI family member, X-linked chrXp11.23 ENST00000425661 0.021 0.002 0.0003 rs2230739 ADCY9 adenylate cyclase 9 chr16p13.3 ENST00000294016 0.302 0.357 0.260 rs17521570 RAI14 retinoic acid induced 14 chr5p13.2 ENST00000515799 0.116 0.119 0.099 rs2230681 PSMD9 proteasome 26S subunit, non-ATPase 9 chr12q24.31 ENST00000261817 0.123 0.856 0.834 rs557337 TBC1D4 TBC1 domain family, member 4 chr13q22.2 ENST00000377636 0.065 0.083 0.174 'Alternate Allele Frequency 2 Exome Aggregation Consortium 3 1000 Genomes EFTA01125191 summary MOSES can find a diverse set of accurate boolean categorization functions even in data with very large feature sets and highly imbalanced sample category sizes. EFTA01125192 Probabilistic Logic for Connecting Genomic Data Patterns with Bio-Ontologies: A simple example (Eddie Monroe & Ben Goertzel) EFTA01125193 Logic: very general, flexible framework for carrying out abstract reasoning. Encompasses both mathematical and commonsense reasoning. Probability theory: very general, flexible framework for carrying out reasoning based on uncertainty. Used in a huge variety of areas including data mining, robotics, vision processing, etc. EFTA01125194 "Progic" = probability + logic •Various approaches to synthesizing probability and logic exist •Probabilistic Logic Networks (PLN) is a "progic" framework oriented toward artificial general intelligence. EFTA01125195 Probabilistic Logic Networks • OpenCog represents knowledge in its "Atomspace" in terms of nodes and links of various types • PLN contains a set of probabilistic logic rules, that transform sets of nodes/links into other sets of nodes/links • PLN can do deduction, induction, abduction, analogy and other types of reasoning • PLN can reason on any kind of data, including data-patterns ("combo models") learned in genomic data by MOSES, or data imported into OpenCog from bio-ontologies • Due to its ability to process huge amounts of information in subtle ways, PLN can identify data patterns the human mind will miss • A fundamentally different paradigm than currently popular "machine learning" or "deep learning" architectures, with more capability for abstract symbolic understanding — but can work together with more standard ML algorithms EFTA01125196 Term Logic Predicate Logic A -3. B B -, C I- A— C A A --> B - A 33 A I - 33 -> C 1. Induction 23. C B C I I— A —5. 35 Alzoduc flora EFTA01125197 Multiple PLN Relationship Types PLN involves more than a dozen logical relationship types, each with particular semantics. For instance could be interpreted in many ways including Extensionallnheritance A B Extensionallnheritance B C Extensionallnheritance A C IntensionalInheritance A B IntensionalInheritance B C IntensionalInheritance A C EFTA01125198 "Higher-Order" PLN Following Pei Wang' s usage in NARS, in PLN we refer to logic regarding variables or higher-order functions as "higher-order" ImplicationLink EvaluationLink has($X, mouth) EvaluationLink eats($X, food) EFTA01125199 Quantifying Truth Values Each PLN relationship has a truth value attached to it. PLN supports truth value objects of different types, e.g. • Single probability • SimpleTruthValue: - (s,c) = (probability, confidence level) — (s,n) = (probability, amount of evidence) • Imprecise truth value — (L,U) interval, e.g. (.4,.6) • Indefinite truth value — (L,U,b,k) ... interval plus confidence level b, and "personality parameter" k, e.g. (.4,.6,.9,2) • Distributional truth value - first or second order pdf EFTA01125200 Example PLN rule+formula: deduction B <sB> C <sc> ExtensionalInheritance A B <sAB> ExtensionalInheritance B C <sBc> ExtensionalInheritance A C <SAC> SAC = SAB SBC + ( 1 - SAO ( Sc — sg Sgc ) / (1 — sg ) As given above, this acts on single-probability truth values. It can be extended to other true value forms. EFTA01125201 PLN rules Each rule maps a tuple of relationships into a relationship Example: deduction rule Subset A B Subset B C I - Subset A C PLN formulas Each formula maps a tuple of truth values into a truth value Example: deduction formula SAC = 5AB Sgc + (1 — SAB) ( Sc — ss Sgc ) / (1 — Ss ) EFTA01125202 Inversion (Bayes Rule) A B Subset A B (- Subset B A In PLN, simple first-order induction and abduction are obtained by combining deduction and Bayes rule. More advanced induction and abduction result from using intensional relationships. A 4 B B 4 C I - A4 C 5 4 A 4 I B A 4 B I - A -> C [repeated 3 times] B A A Nt, Abduction EFTA01125203 Glossary of Link Types AttractionLink Indicates the extent to which one concept is a pattern or property helping to characterize another. (AttractionLink A B) indicates the extent to which B is a property that characterizes A. ConceptNode A node representing any concept. ExecutionOutputLink Indicates execution of a function with a list arguments to that function. This allows for atomspace representation of the execution of arbitrary code. GeneNode A node representing a particular gene. EFTA01125204 Glossary of Link Types (cont.) GroundedSchemallode Specifies the name of a predefined procedure that is to be called. ImplicationLink Expresses an if...then... relation, or that the truth of one predicate implies the truth of another. (ImplicationLink A B) denotes that A implies B. IntensionalEquivalenceLink Indicates that the properties associated with one predicate being true are similar to the properties associated with another predicate being true. (IntensionalEquivalancelink A B) denotes that the properties associated with A being true are similar to the properties associated with B being true. IntensionalImplicationLink Expresses an if... then... relation between the properties of 2 predicates. (IntensionalImplicationlink A B) denotes that the properties of A imply the properties of B. EFTA01125205 Glossary of Link Types (cont.) IntensionalSimilarityLink Indicates that two concepts have similar properties. (IntensionalSimilaritylink A B) denotes that the properties of A are similar to the properties of B ListLink Used for grouping Atoms for some purpose, typically to specify a set of arguments to some function or relation. MemberLink Indicates set membership. (MemberLink x S) denotes that element x is a member of set S. The TruthValue associated with a MemberLink is meant to indicate fuzzy set membership. NotLink Corresponds to the negation of a concept or predicate. EFTA01125206 Glossary of Link Types (cont.) PredicateNode Names the predicate of a relation. Predicates are functions that have arguments and produce a truth value as output. SetLink A type of link used to group its arguments into a set (SetLink x y z) simply indicates that there is a set {x,y,z} SubsetLink Denotes extensional inheritance, which is inheritance between sets based on their members. It specifies an "is-an-instance-of" relationship. (SubsetLink A B) specifies that A is an instance of B. EFTA01125207 Example Inference: Goal • Through MOSES analysis, we found overexpression of LY96 appears to distinguish Nonagenarians from controls. • Using PLN, what can we infer about the relationship between LY96 and longevity based on background domain and experimental knowledge? Our target conclusion is: ImplicationLink (ExecutionOutputLink (GroundedSchemallode "scm: make-over-expression-predicate") (GeneNode "LY96")) (PredicateNode "LongLived") Interpretation: "Overexpression of LY96 implies longevity." EFTA01125208 Background Information There is pre-existing evidence that over-expression of gene TBK1 is associated with increased lifespan (Source: Lifespan Observations Database) (IntensionalImplicationLink (sty 0.3 0.7) (ExecutionOutputLink (sty 0.2 0.7) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty .0004 0.9)))) (PredicateNode "LongLived" (sty 0.15 0.8))) Interpretation: "Overexpression of TBK1 implies longevity" Genes are associated with Gene Ontology terms and other categories. ((MemberLink (GeneNode "TBK1" (sty 0004 0.9)) (ConceptNode "GO:0005515" (sty 0.001 0.9))) (MemberLink (GeneNode "TBK1" (sty 0004 0.9)) (ConceptNode "GO:0045087" (sty 0.001 0.9))) Interpretation: "TBK1 is a member of GO category 0005515," "TBK1 is a member of GO category 0045087," ... for each gene category annotation EFTA01125209 Inference Chain Steps (1) Member-to-Subset Rule (Member A B) I- (Subset (Set A) B) Premises: (MemberLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)) (ConceptNode "GO:0051607" (sty 0.001 0.89999998)) ) "TBK1 is a member of GO category 0051607" Conclusions: (SubsetLink (SetLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)) ) (ConceptNode "GO:0051607" (sty 0.001 0.89999998)) ) "The singleton set containing TBK1 is a subset of GO category 0051607" EFTA01125210 Intensional Similarity • We will infer a relationship between the gene LY96 and the predicate LongLived through the similarity of LY96 with gene TBK1, which is already known to be related to longevity. • Intensional similarity is based on common properties of the genes. • Steps 2-5 that follow are needed for creating the IntensionalSimilarity relationship. EFTA01125211 (2) Compare gene properties • We are using GO category annotations for gene properties. • At the start of the inference, we need to get the supersets of {TBK1} and {LY96} and determine the intersection and union of the supersets LY96: member of 25 GO categories TBK1: member of 34 GO categories Common categories (intersection): GO:0005515 protein binding GO:0045087 innate immune response GO:0006954 inflammatory response GO:0010008 endosome membrane GO:0002224 toll-like receptor signaling pathway GO:0002756 MyD88-independent toll-like receptor signaling pathway GO:0007249 1-kappaB kinase/NF-kappaB signaling GO:0034138 toll-like receptor 3 signaling pathway GO:0034142 toll-like receptor 4 signaling pathway GO:0035666 TRIF-dependent toll-like receptor signaling pathway EFTA01125212 (3) Subset NotA B Direct Evaluation (Inheritance A B) I- (Inheritance (Not A) B) For each common category relationship (LinkType A B), create (LinkType (Not A) B) Premises: (SubsetLink (sty 1 0.99999982) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998)) ) "{LY96} is a subset of GO:0045087" Conclusions: (SubsetLink (sty 0.028667862 0.99999982) (NotLink (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) ) (ConceptNode "GO:0045087" (sty 0.001 0.89999998))) "A random gene (exclusive of LY96) belongs to GO:0045087 (with a low probability)" EFTA01125213 (4) AttractionRule (And (Subset A B) (Subset (Not A) B)) I- (AttractionLink A B) Make AttractionLinks for LY96 and TBK1 for each common relationship (IOW for each relationship in the intersection of the supersets). Premises: (SubsetLink (sty 1 0.99999982) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998)) (SubsetLink (sty 0.028667862 0.99999982) (NotLink (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998))) {O196} is a subset of "GO:0045087," "A random gene not in {LY96} is a subset of GO:0045087 (with a low probability)" Conclusions: (AttractionLink (sty 0.97133213 0.99999982) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998))) "GO:0045087 is a property of/pattern in (LY96)" EFTA01125214 (5)IntensionalSimilarity Direct Evaluation (And (Attraction P A) (Attraction P B) (Attraction (Q A) (Attraction (0 B) ...) I- (IntensionalSimilarity A B) Premises: (AttractionLink (sty 0.97133213 0.99999982) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998))) (AttractionLink (sty 0.97133213 0.99999982) (SetLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998))) (ConceptNode "GO:0045087" (sty 0.001 0.89999998))) "GO:0045087 is a property of {1Y96}" "GO:0045087 is a property of {TBK1}" Etc.. . . Conclusion: (IntensionalSimilarityLink (sty 0.19570713 0.99999982) (SetLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998))) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) "{TBK1} properties are similar to {LY96} properties" EFTA01125215 (6) Singleton-Similarity-Rule (Similarity {A} {B}) I- (Similarity A B) Premise: (IntensionalSimilarityLink (sty 0.19570713 0.99999982) (SetLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998))) (SetLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))) "{TBK1} properties are similar to {LY96} properties" Conclusion: (IntensionalSimilarityLink (sty 0.19570713 0.99999982) (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)) (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)) ) "TBK1 properties are similar to LY96 properties" EFTA01125216 (7) Gene-Similarity-to-Overexpression-Equivalence (Similarity (Gene A) (Gene B)) I- (Equivalence (A-overexpressed) (B-overexpressed) Premise: (IntensionalSimilarityLink (sty 0.19570713 0.99999982) (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)) (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)) ) "TBK1 properties are similar to LY96 properties" Conclusion: untensionalEquivalenceLink (sty 0.19570713 0.99999982) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)))) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))))) "Properties associated with over-expression of TBK1 are similar to properties associated with overexpression of LY96" EFTA01125217 (8) Equivalence-Transformation Rule (Equivalence A B) I- (And (Implication A B) (Implication B A)) Premise: (IntensionalEquivalenceLink (sty 0.19570713 0.99999982) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998)))) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998))))) " `Overexpression of TBK1 properties' is similar to 'overexpression of RYR1 properties' " Conclusion: (IntensionalImplicationLink (sty 0.3273496 0.99999982) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)))) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998))))) "Having properties associated with over-expression of LY96 implies having properties associated with overexpression of TBK1" EFTA01125218 (9) Implication Deduction Rule (And (Implication A B) (Implication B C) I- (Implication A C) (Part 1) Premises: (IntensionalImplicationLink (sty 0.3273496 0.99999982) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)))) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty 4.1666666e-05 0.89999998))))) "Having properties associated with overexpression of LY96, implies having properties associated with overexpression of TBK1" (IntensionalImplicationLink (sty 0.3 0.7) (ExecutionOutputLink (sty 0.2 0.7) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "TBK1" (sty .0004 0.9)))) (PredicateNode "LongLived" (sty 0.15 0.8))) ) "Having properties associated with overexpression of TBK1, implies having properties associated with longevity" EFTA01125219 (9) Implication Deduction Rule (And (Implication A B) (Implication B I- (Implication A C) (Part 2) Conclusion: (IntensionalImplicationLink (sty 0.17387806 0.69999999) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)) (PredicateNode "LongLived" (sty 0.15000001 0.80000001)) "Having properties associated with 'Overexpression of LY96' implies having properties associated with longevity" EFTA01125220 (10) Implication Conversion Rule (IntensionalImplication A B) - (Implication A B) Premise: (IntensionalImplicationLink (sty 0.17387806 0.69999999) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)))) (PredicateNode "LongLived" (sty 0.15000001 0.80000001)) "Having properties associated with `Overexpression of LY96' implies having properties associated with longevity" Conclusion: (ImplicationLink (sty 0.17387806 0.48999998) (ExecutionOutputLink (sty 0.2 0.69999999) (GroundedSchemallode "scm: make-overexpression-predicate") (ListLink (GeneNode "LY96" (sty 4.1666666e-05 0.89999998)))) (PredicateNode "LongLived" (sty 0.15000001 0.80000001)) "Overexpression of LY96 implies longevity" (Our target conclusion) Next big Al challenge here: Fully automated, scalable inference control (choice of which inference steps to take), via data-mining of inference history EFTA01125221 Broad Vision: Al Scientist • Integrated knowledge-base of all biological (+ chemical etc.) knowledge, in the Atomspace, built in semi- automated way • Knowledge comes from: datasets, databases, texts, simulations, automated use of lab equipment • MOSES, PLN and other Al methods used for hypothesis discovery and validation • Connect OpenCog w/ simulation engine, use OpenCog data/inferences to help set simulation parameters • Al to design experiments, run robotized experiments • Language generation to produce written reports • Full-on Al Scientist!! EFTA01125222