Enhancing Retrosynthetic Reaction Prediction with Deep Learning Using Multiscale Reaction Classification.

Chemical synthesis planning is a key aspect in many fields of chemistry, especially drug discovery. Recent implementations of machine learning and artificial intelligence techniques for retrosynthetic analysis have shown great potential to improve computational methods for synthesis planning. Herein, we present a multiscale, data-driven approach for retrosynthetic analysis with deep highway networks (DHN). We automatically extracted reaction rules (i.e., ways in which a molecule is produced) from a data set consisting of chemical reactions derived from U.S. patents. We performed the retrosynthetic reaction prediction task in two steps: first, we built a DHN model to predict which group of reactions (consisting of chemically similar reaction rules) was employed to produce a molecule. Once a reaction group was identified, a DHN trained on the subset of reactions within the identified reaction group, was employed to predict the transformation rule used to produce a molecule. To validate our approach, we predicted the first retrosynthetic reaction step for 40 approved drugs using our multiscale model and compared its predictive performance with a conventional model trained on all machine-extracted reaction rules employed as a control. Our multiscale approach showed a success rate of 82.9% at generating valid reactants from retrosynthetic reaction predictions. Comparatively, the control model trained on all machine-extracted reaction rules yielded a success rate of 58.5% on the validation set of 40 pharmaceutical molecules, indicating a significant statistical improvement with our approach to match known first synthetic reaction of the tested drugs in this study. While our multiscale approach was unable to outperform state-of-the-art rule-based systems curated by expert chemists, multiscale classification represents a marked enhancement in retrosynthetic analysis and can be easily adapted for use in a range of artificial intelligence strategies.

Refining the impact of TCF7L2 gene variants on type 2 diabetes and adaptive evolution.

We recently described an association between risk of type 2diabetes and variants in the transcription factor 7-like 2 gene (TCF7L2; formerly TCF4), with a population attributable risk (PAR) of 17%-28% in three populations of European ancestry. Here, we refine the definition of the TCF7L2 type 2diabetes risk variant, HapB(T2D), to the ancestral T allele of a SNP, rs7903146, through replication in West African and Danish type 2 diabetes case-control studies and an expanded Icelandic study. We also identify another variant of the same gene, HapA, that shows evidence of positive selection in East Asian, European and West African populations. Notably, HapA shows a suggestive association with body mass index and altered concentrations of the hunger-satiety hormones ghrelin and leptin in males, indicating that the selective advantage of HapA may have been mediated through effects on energy metabolism.

A variant in CDKAL1 influences insulin response and risk of type 2 diabetes.

We conducted a genome-wide association study for type 2 diabetes (T2D) in Icelandic cases and controls, and we found that a previously described variant in the transcription factor 7-like 2 gene (TCF7L2) gene conferred the most significant risk. In addition to confirming two recently identified risk variants, we identified a variant in the CDKAL1 gene that was associated with T2D in individuals of European ancestry (allele-specific odds ratio (OR) = 1.20 (95% confidence interval, 1.13-1.27), P = 7.7 x 10(-9)) and individuals from Hong Kong of Han Chinese ancestry (OR = 1.25 (1.11-1.40), P = 0.00018). The genotype OR of this variant suggested that the effect was substantially stronger in homozygous carriers than in heterozygous carriers. The ORs for homozygotes were 1.50 (1.31-1.72) and 1.55 (1.23-1.95) in the European and Hong Kong groups, respectively. The insulin response for homozygotes was approximately 20% lower than for heterozygotes or noncarriers, suggesting that this variant confers risk of T2D through reduced insulin secretion.

Genome-wide association study identifies a second prostate cancer susceptibility variant at 8q24.

Prostate cancer is the most prevalent noncutaneous cancer in males in developed regions, with African American men having among the highest worldwide incidence and mortality rates. Here we report a second genetic variant in the 8q24 region that, in conjunction with another variant we recently discovered, accounts for about 11%-13% of prostate cancer cases in individuals of European descent and 31% of cases in African Americans. We made the current discovery through a genome-wide association scan of 1,453 affected Icelandic individuals and 3,064 controls using the Illumina HumanHap300 BeadChip followed by four replication studies. A key step in the discovery was the construction of a 14-SNP haplotype that efficiently tags a relatively uncommon (2%-4%) susceptibility variant in individuals of European descent that happens to be very common (approximately 42%) in African Americans. The newly identified variant shows a stronger association with affected individuals who have an earlier age at diagnosis.

Common variants on chromosomes 2q35 and 16q12 confer susceptibility to estrogen receptor-positive breast cancer.

Familial clustering studies indicate that breast cancer risk has a substantial genetic component. To identify new breast cancer risk variants, we genotyped approximately 300,000 SNPs in 1,600 Icelandic individuals with breast cancer and 11,563 controls using the Illumina Hap300 platform. We then tested selected SNPs in five replication sample sets. Overall, we studied 4,554 affected individuals and 17,577 controls. Two SNPs consistently associated with breast cancer: approximately 25% of individuals of European descent are homozygous for allele A of rs13387042 on chromosome 2q35 and have an estimated 1.44-fold greater risk than noncarriers, and for allele T of rs3803662 on 16q12, about 7% are homozygous and have a 1.64-fold greater risk. Risk from both alleles was confined to estrogen receptor-positive tumors. At present, no genes have been identified in the linkage disequilibrium block containing rs13387042. rs3803662 is near the 5' end of TNRC9 , a high mobility group chromatin-associated protein whose expression is implicated in breast cancer metastasis to bone.

Common variants conferring risk of schizophrenia.

Schizophrenia is a complex disorder, caused by both genetic and environmental factors and their interactions. Research on pathogenesis has traditionally focused on neurotransmitter systems in the brain, particularly those involving dopamine. Schizophrenia has been considered a separate disease for over a century, but in the absence of clear biological markers, diagnosis has historically been based on signs and symptoms. A fundamental message emerging from genome-wide association studies of copy number variations (CNVs) associated with the disease is that its genetic basis does not necessarily conform to classical nosological disease boundaries. Certain CNVs confer not only high relative risk of schizophrenia but also of other psychiatric disorders. The structural variations associated with schizophrenia can involve several genes and the phenotypic syndromes, or the 'genomic disorders', have not yet been characterized. Single nucleotide polymorphism (SNP)-based genome-wide association studies with the potential to implicate individual genes in complex diseases may reveal underlying biological pathways. Here we combined SNP data from several large genome-wide scans and followed up the most significant association signals. We found significant association with several markers spanning the major histocompatibility complex (MHC) region on chromosome 6p21.3-22.1, a marker located upstream of the neurogranin gene (NRGN) on 11q24.2 and a marker in intron four of transcription factor 4 (TCF4) on 18q21.2. Our findings implicating the MHC region are consistent with an immune component to schizophrenia risk, whereas the association with NRGN and TCF4 points to perturbation of pathways involved in brain development, memory and cognition.

Variant of transcription factor 7-like 2 (TCF7L2) gene confers risk of type 2 diabetes.

We have previously reported suggestive linkage of type 2 diabetes mellitus to chromosome 10q. We genotyped 228 microsatellite markers in Icelandic individuals with type 2 diabetes and controls throughout a 10.5-Mb interval on 10q. A microsatellite, DG10S478, within intron 3 of the transcription factor 7-like 2 gene (TCF7L2; formerly TCF4) was associated with type 2 diabetes (P = 2.1 x 10(-9)). This was replicated in a Danish cohort (P = 4.8 x 10(-3)) and in a US cohort (P = 3.3 x 10(-9)). Compared with non-carriers, heterozygous and homozygous carriers of the at-risk alleles (38% and 7% of the population, respectively) have relative risks of 1.45 and 2.41. This corresponds to a population attributable risk of 21%. The TCF7L2 gene product is a high mobility group box-containing transcription factor previously implicated in blood glucose homeostasis. It is thought to act through regulation of proglucagon gene expression in enteroendocrine cells via the Wnt signaling pathway.

A variant of the gene encoding leukotriene A4 hydrolase confers ethnicity-specific risk of myocardial infarction.

Variants of the gene ALOX5AP (also known as FLAP) encoding arachidonate 5-lipoxygenase activating protein are known to be associated with risk of myocardial infarction. Here we show that a haplotype (HapK) spanning the LTA4H gene encoding leukotriene A4 hydrolase, a protein in the same biochemical pathway as ALOX5AP, confers modest risk of myocardial infarction in an Icelandic cohort. Measurements of leukotriene B4 (LTB4) production suggest that this risk is mediated through upregulation of the leukotriene pathway. Three cohorts from the United States also show that HapK confers a modest relative risk (1.16) in European Americans, but it confers a threefold larger risk in African Americans. About 27% of the European American controls carried at least one copy of HapK, as compared with only 6% of African American controls. Our analyses indicate that HapK is very rare in Africa and that its occurrence in African Americans is due to European admixture. Interactions with other genetic or environmental risk factors that are more common in African Americans are likely to account for the greater relative risk conferred by HapK in this group.

A common variant associated with prostate cancer in European and African populations.

With the increasing incidence of prostate cancer, identifying common genetic variants that confer risk of the disease is important. Here we report such a variant on chromosome 8q24, a region initially identified through a study of Icelandic families. Allele -8 of the microsatellite DG8S737 was associated with prostate cancer in three case-control series of European ancestry from Iceland, Sweden and the US. The estimated odds ratio (OR) of the allele is 1.62 (P = 2.7 x 10(-11)). About 19% of affected men and 13% of the general population carry at least one copy, yielding a population attributable risk (PAR) of approximately 8%. The association was also replicated in an African American case-control group with a similar OR, in which 41% of affected individuals and 30% of the population are carriers. This leads to a greater estimated PAR (16%) that may contribute to higher incidence of prostate cancer in African American men than in men of European ancestry.

Genetics of gene expression and its effect on disease.

Common human diseases result from the interplay of many genes and environmental factors. Therefore, a more integrative biology approach is needed to unravel the complexity and causes of such diseases. To elucidate the complexity of common human diseases such as obesity, we have analysed the expression of 23,720 transcripts in large population-based blood and adipose tissue cohorts comprehensively assessed for various phenotypes, including traits related to clinical obesity. In contrast to the blood expression profiles, we observed a marked correlation between gene expression in adipose tissue and obesity-related traits. Genome-wide linkage and association mapping revealed a highly significant genetic component to gene expression traits, including a strong genetic effect of proximal (cis) signals, with 50% of the cis signals overlapping between the two tissues profiled. Here we demonstrate an extensive transcriptional network constructed from the human adipose data that exhibits significant overlap with similar network modules constructed from mouse adipose data. A core network module in humans and mice was identified that is enriched for genes involved in the inflammatory and immune response and has been found to be causally associated to obesity-related traits.

A common inversion under selection in Europeans.

A refined physical map of chromosome 17q21.31 uncovered a 900-kb inversion polymorphism. Chromosomes with the inverted segment in different orientations represent two distinct lineages, H1 and H2, that have diverged for as much as 3 million years and show no evidence of having recombined. The H2 lineage is rare in Africans, almost absent in East Asians but found at a frequency of 20% in Europeans, in whom the haplotype structure is indicative of a history of positive selection. Here we show that the H2 lineage is undergoing positive selection in the Icelandic population, such that carrier females have more children and have higher recombination rates than noncarriers.

Variants conferring risk of atrial fibrillation on chromosome 4q25.

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia in humans and is characterized by chaotic electrical activity of the atria. It affects one in ten individuals over the age of 80 years, causes significant morbidity and is an independent predictor of mortality. Recent studies have provided evidence of a genetic contribution to AF. Mutations in potassium-channel genes have been associated with familial AF but account for only a small fraction of all cases of AF. We have performed a genome-wide association scan, followed by replication studies in three populations of European descent and a Chinese population from Hong Kong and find a strong association between two sequence variants on chromosome 4q25 and AF. Here we show that about 35% of individuals of European descent have at least one of the variants and that the risk of AF increases by 1.72 and 1.39 per copy. The association with the stronger variant is replicated in the Chinese population, where it is carried by 75% of individuals and the risk of AF is increased by 1.42 per copy. A stronger association was observed in individuals with typical atrial flutter. Both variants are adjacent to PITX2, which is known to have a critical function in left-right asymmetry of the heart.

Ancestry-shift refinement mapping of the C6orf97-ESR1 breast cancer susceptibility locus.

We used an approach that we term ancestry-shift refinement mapping to investigate an association, originally discovered in a GWAS of a Chinese population, between rs2046210[T] and breast cancer susceptibility. The locus is on 6q25.1 in proximity to the C6orf97 and estrogen receptor alpha (ESR1) genes. We identified a panel of SNPs that are correlated with rs2046210 in Chinese, but not necessarily so in other ancestral populations, and genotyped them in breast cancer case:control samples of Asian, European, and African origin, a total of 10,176 cases and 13,286 controls. We found that rs2046210[T] does not confer substantial risk of breast cancer in Europeans and Africans (OR = 1.04, P = 0.099, and OR = 0.98, P = 0.77, respectively). Rather, in those ancestries, an association signal arises from a group of less common SNPs typified by rs9397435. The rs9397435[G] allele was found to confer risk of breast cancer in European (OR = 1.15, P = 1.2 x 10(-3)), African (OR = 1.35, P = 0.014), and Asian (OR = 1.23, P = 2.9 x 10(-4)) population samples. Combined over all ancestries, the OR was 1.19 (P = 3.9 x 10(-7)), was without significant heterogeneity between ancestries (P(het) = 0.36) and the SNP fully accounted for the association signal in each ancestry. Haplotypes bearing rs9397435[G] are well tagged by rs2046210[T] only in Asians. The rs9397435[G] allele showed associations with both estrogen receptor positive and estrogen receptor negative breast cancer. Using early-draft data from the 1,000 Genomes project, we found that the risk allele of a novel SNP (rs77275268), which is closely correlated with rs9397435, disrupts a partially methylated CpG sequence within a known CTCF binding site. These studies demonstrate that shifting the analysis among ancestral populations can provide valuable resolution in association mapping.

Recombination rate and reproductive success in humans.

Intergenerational mixing of DNA through meiotic recombinations of homologous chromosomes during gametogenesis is a major event that generates diversity in the eukaryotic genome. We examined genome-wide microsatellite data for 23,066 individuals, providing information on recombination events of 14,140 maternal and paternal meioses each, and found a positive correlation between maternal recombination counts of an offspring and maternal age. We postulated that the recombination rate of eggs does not increase with maternal age, but that the apparent increase is the consequence of selection. Specifically, a high recombination count increased the chance of a gamete becoming a live birth, and this effect became more pronounced with advancing maternal age. Further support for this hypothesis came from our observation that mothers with high oocyte recombination rate tend to have more children. Hence, not only do recombinations have a role in evolution by yielding diverse combinations of gene variants for natural selection, but they are also under selection themselves.

A high-resolution recombination map of the human genome.

Determination of recombination rates across the human genome has been constrained by the limited resolution and accuracy of existing genetic maps and the draft genome sequence. We have genotyped 5,136 microsatellite markers for 146 families, with a total of 1,257 meiotic events, to build a high-resolution genetic map meant to: (i) improve the genetic order of polymorphic markers; (ii) improve the precision of estimates of genetic distances; (iii) correct portions of the sequence assembly and SNP map of the human genome; and (iv) build a map of recombination rates. Recombination rates are significantly correlated with both cytogenetic structures (staining intensity of G bands) and sequence (GC content, CpG motifs and poly(A)/poly(T) stretches). Maternal and paternal chromosomes show many differences in locations of recombination maxima. We detected systematic differences in recombination rates between mothers and between gametes from the same mother, suggesting that there is some underlying component determined by both genetic and environmental factors that affects maternal recombination rates.

The gene encoding 5-lipoxygenase activating protein confers risk of myocardial infarction and stroke.

We mapped a gene predisposing to myocardial infarction to a locus on chromosome 13q12-13. A four-marker single-nucleotide polymorphism (SNP) haplotype in this locus spanning the gene ALOX5AP encoding 5-lipoxygenase activating protein (FLAP) is associated with a two times greater risk of myocardial infarction in Iceland. This haplotype also confers almost two times greater risk of stroke. Another ALOX5AP haplotype is associated with myocardial infarction in individuals from the UK. Stimulated neutrophils from individuals with myocardial infarction produce more leukotriene B4, a key product in the 5-lipoxygenase pathway, than do neutrophils from controls, and this difference is largely attributed to cells from males who carry the at-risk haplotype. We conclude that variants of ALOX5AP are involved in the pathogenesis of both myocardial infarction and stroke by increasing leukotriene production and inflammation in the arterial wall.

The gene encoding phosphodiesterase 4D confers risk of ischemic stroke.

We previously mapped susceptibility to stroke to chromosome 5q12. Here we finely mapped this locus and tested it for association with stroke. We found the strongest association in the gene encoding phosphodiesterase 4D (PDE4D), especially for carotid and cardiogenic stroke, the forms of stroke related to atherosclerosis. Notably, we found that haplotypes can be classified into three distinct groups: wild-type, at-risk and protective. We also observed a substantial disregulation of multiple PDE4D isoforms in affected individuals. We propose that PDE4D is involved in the pathogenesis of stroke, possibly through atherosclerosis, which is the primary pathological process underlying ischemic stroke.

Multiple genetic loci for bone mineral density and fractures.

BACKGROUND: Bone mineral density influences the risk of osteoporosis later in life and is useful in the evaluation of the risk of fracture. We aimed to identify sequence variants associated with bone mineral density and fracture. METHODS: We performed a quantitative trait analysis of data from 5861 Icelandic subjects (the discovery set), testing for an association between 301,019 single-nucleotide polymorphisms (SNPs) and bone mineral density of the hip and lumbar spine. We then tested for an association between 74 SNPs (most of which were implicated in the discovery set) at 32 loci in replication sets of Icelandic, Danish, and Australian subjects (4165, 2269, and 1491 subjects, respectively). RESULTS: Sequence variants in five genomic regions were significantly associated with bone mineral density in the discovery set and were confirmed in the replication sets (combined P values, 1.2x10(-7) to 2.0x10(-21)). Three regions are close to or within genes previously shown to be important to the biologic characteristics of bone: the receptor activator of nuclear factor-kappaB ligand gene (RANKL) (chromosomal location, 13q14), the osteoprotegerin gene (OPG) (8q24), and the estrogen receptor 1 gene (ESR1) (6q25). The two other regions are close to the zinc finger and BTB domain containing 40 gene (ZBTB40) (1p36) and the major histocompatibility complex region (6p21). The 1p36, 8q24, and 6p21 loci were also associated with osteoporotic fractures, as were loci at 18q21, close to the receptor activator of the nuclear factor-kappaB gene (RANK), and loci at 2p16 and 11p11. CONCLUSIONS: We have discovered common sequence variants that are consistently associated with bone mineral density and with low-trauma fractures in three populations of European descent. Although these variants alone are not clinically useful in the prediction of risk to the individual person, they provide insight into the biochemical pathways underlying osteoporosis.

Lessons from the past: familial aggregation analysis of fatal pandemic influenza (Spanish flu) in Iceland in 1918.

The pandemic influenza of 1918 (Spanish flu) killed 21-50 million people globally, including in Iceland, where the characteristics and spread of the epidemic were well documented. It has been postulated that genetic host factors may have contributed to this high mortality. We identified 455 individuals who died of the Spanish flu in Iceland during a 6-week period during the winter of 1918, representing >92% of all fatal domestic cases mentioned by historical accounts. The highest case fatality proportion was 2.8%, and peak excess mortality was 162/100,000/week. Fatality proportions were highest among infants, young adults, and the elderly. A genealogical database was used to study relatedness and relative risk (RR) of the fatal influenza victims and relatives of their unaffected mates. The significance of these RR computations was assessed by drawing samples randomly from the genealogical database matched for age, sex, and geographical distribution. Familial aggregation of fatalities was seen, with RRs for death ranging from 3.75 for first-degree relatives (P < 0.0001) to 1.82 (P = 0.005), 1.12 (P = 0.252), and 1.47 (P = 0.0001) for second- to fourth-degree relatives of fatal influenza victims, respectively. The RRs within the families of unaffected mates of fatal influenza victims were 2.95 (P < 0.0001), 1.27 (P = 0.267), 1.35 (P = 0.04), and 1.42 (P = 0.001), for first- to fourth-degree relatives, respectively. In conclusion, the risk of death from the Spanish flu was similar within families of patients who succumbed to the illness and within families of their mates who survived. Our data do not provide conclusive evidence for the role of genetic factors in susceptibility to the Spanish flu.

The chromosome 9p21 risk locus is associated with angiographic severity and progression of coronary artery disease.

AIMS: we tested the hypothesis that the 9p21 risk locus promotes atherosclerosis by examining the association between rs10757278 and coronary artery disease (CAD) severity and progression determined by semi-quantitative angiographic scores. METHODS AND RESULTS: the rs10757278 single nucleotide polymorphism (SNP) was genotyped as the marker for the 9p21 locus in 2334 Caucasian patients undergoing cardiac catheterization (mean age 63, male 67%). Angiographic CAD was assessed using two semi-quantitative scoring systems with one estimating severity (Gensini) and the other extent (Sullivan). A subset of 308 patients who underwent two or more coronary angiograms at least 6 months apart were examined for net change in Gensini and Sullivan scores over time to determine the rate of CAD progression by genotype and were further classified as 'progressors' or 'non-progressors' based on absolute change per year in angiographic severity score. We replicated the association between the rs10757278 SNP and myocardial infarction and binary (presence/absence) angiographic classifications of CAD. Furthermore, we observed a significant additive association with this SNP, and both severity and extent of CAD using angiographic scores, after adjustment for age, gender, body mass index, traditional cardiovascular risk factors, myocardial infarction, and statin use (Gensini P = 0.016, Sullivan P = 0.005). In addition, there was a significant linear association with CAD progression before and after adjustment for covariates (Gensini P = 0.023, Sullivan P = 0.003) with homozygotes for the risk variant having three-fold greater odds of CAD progression compared with the referent group. CONCLUSION: the 9p21 risk locus is associated with angiographically defined severity, extent, and progression of CAD, suggesting a role for this locus in influencing atherosclerosis and its progression.