- Email: [email protected]
Less is more: artificial intelligence and gene-expression arrays
Comment
fluorescence, and at the same maximise detection sensitivity?9 Other issues, such as how to develop efficient chemistry to produce quantum dot-conjugated biomolecules, and how quantum dot-labelled biomolecules (and cells or tissues) differ from their native counterparts, will also have to be addressed.10 However, the new findings indicate the feasibility of using dot-based imaging techniques to tackle one of the long-standing quests in cancer research: to find a way to precisely monitor and evaluate metastatic tumour cells in a patient.
2
Grace Y J Chen, *Shao Q Yao
7
Departments of Chemistry and Biological Sciences, National University of Singapore, Singapore 117543, Republic of Singapore. [email protected]
8
3
4
5
6
9
We declare that we have no conflict of interest 1
Voura EB, Jaiswal JK, Mattoussi H, Simon SM. Tracking metastatic tumor cell extravasation with quantum dot nanocrystals and fluorescence emission-scanning microscopy. Nature Med 2004; 10: 993–98.
10
Murray CB, Norris DJ, Bawendi, MG. Synthesis and characterization of nearly monodisperse CDE (E=S, SE, TE) semiconductor nanocrystallites. J Am Chem Soc 1993; 115: 8706–15. Dubertert B, Skourides P, Norris DJ, Noireaux V, Brivanlou AH, Libchaber A. In vivo imaging of quantum dots encapsulated in phospholipid micelles. Science 2002; 298: 1759–62. Akerman ME, Chan WC, Laakkonen P, Bhatia SN, Ruoslahti E. Nanocrystal targeting in vivo. Proc Natl Acad Sci USA 2002; 99: 12617–21. Wu X, Liu H, Liu J, et al. Immunofluorescent labeling of cancer marker Her2 and other cellular targets with semiconductor quantum dots. Nat Biotechnol 2003; 21: 41–46. Jaiswai JK, Mattoussi H, Mauro JM, Simon SM. Long-term multiple color imaging of live cells using quantum dot bioconjugates. Nat Biotechnol 2003; 21: 47–51. Larson DR, Zipfel WR, Williams RM, et al. Water-soluble quantum dots for multiphoton fluorescence imaging in vivo. Science 2003; 300: 1434–36. Gao X, Cui Y, Levenson RM, Chung LW, Nie S. In vivo cancer targeting and imaging with semiconductor quantum dots. Nat Biotechnol 2004; 22: 969–76. Kim S, Lim YT, Soltesz EG, et al. Near-infrared fluorescent type II quantum dots for sentinel lymph node mapping. Nat Biotechnol 2004; 22: 93–97. Srinivasan R, Yao SQ, Yeo DS. Chemical approaches for live cell bioimaging. Comb Chem High Throughput Screen 2004; 7: 597–604.
Less is more: artificial intelligence and gene-expression arrays
www.thelancet.com Vol 364 December 4, 2004
of gene expression, and Scherf et al3 studied 84 000 experiments (60 cell lines tested for 1400 agents) with 48 000 gene measurements (672 million outcomes measured). Therefore, improved bioinformatics are paramount in the interrogation of array data. So far, most array reports have used statistical classification methods, such as hierarchical clustering. An alternative approach was recently taken by Jun Wei and colleagues,4 using artificial neural networks to process array measurements of neuroblastoma. Such networks are computational models designed to replicate the neuronal networks of the human brain.5 With multiple layers of interconnected nodes, the network learns the importance of various data in predicting outcomes. Once a network is trained, it can analyse new data and reach high levels of predictive accuracy.
Rights were not granted to include this image in electronic media. Please refer to the printed journal. Science Photo Library
The development of gene-expression microarrays in the 1990s heralded a great future for molecular medicine. Single experiments that could simultaneously reveal the expression of thousands of genes in a cell promised to unravel the molecular pathways of numerous medical conditions, including cancer. Encouraging reports have used arrays to classify human cancers according to histology and to distinguish cancers that had been histologically inseparable.1 Examples are numerous, such as Alizadeh et al,2 who used an 18 000-gene array to classify diffuse large B-cell lymphoma into two types with different prognoses. Furthermore, the genes responsible for this stratification suggested the subtypes have different sites of origin, revealing new insights into this disease’s pathogenesis. Other researchers have used gene arrays to take a behavioural, rather than histological, view of disease classification. Scherf et al3 used an 8000-gene array to classify 60 tumour cell lines according to their chemosensitivity to 1400 agents. Improved technology has dramatically increased the capacity of gene arrays. We can now purchase a whole genome array for yeast and human arrays with over 33 000 genes. It is only a matter of time before whole humangenome arrays become readily available for more comprehensive research into the molecular mechanisms of disease. Although larger is usually better, increasing array capacity also increases the risk of statistical error and the amount of data that are generated and require analysis. For example, if one calculates the number of experiments described in previous reports, which used modest array sizes, the results are staggering. Alizadeh et al2 made 1·8 million measurements
Computer artwork of DNA microarray
2003
Comment
Although such networks appear more accurate than traditional statistical methods,6,7 their inner workings are hidden within a black box. The black box has hindered widespread acceptance of artifical neural networks, because it prevents validation and network interrogation. This lack of acceptance and hidden inner workings might be overcome with an alternative form of artificial intelligence based on fuzzy logic.7 Wei and colleagues4 studied 49 patients with neuroblastoma, the most common solid extracranial malignancy in childhood. To guide their treatment, patients were conventionally stratified into high, intermediate, and low risk, by several factors, including age at diagnosis, histological features, tumour stage, and amplification of the MYCN oncogene.8 As with most clinical classifications, this system is imperfect. About 30% of high-risk patients have prolonged survival, and 15% of low-risk patients progress despite treatment.4,8 Wei tried to improve the accuracy of this stratification by using artificial neural networks to model gene-array data. Having analysed 25 000 genes in each sample, the researchers used networks to identify the most relevant predictive genes. By repeating the model while varying the input for each individual gene, the researchers identified the genes that most altered the predictive accuracy of the model (19 genes). This selection was successfully validated in their analysis. A new model trained with these 19 genes more accurately predicted survival than both a previous model with 25 000 genes and the current clinical classification. Furthermore, when only high-risk patients (according to conventional criteria) were analysed, a 19-gene network was able to identify the small proportion of patients that would respond to treatment. It is the remarkable ease with which this approach identifies and validates important genes that highlights its very benefit. Of the 19 genes identified, seven are unknown (so-called expressed sequence tags; RNA sequences produced by cells whose gene identity is yet to be established), and two genes are currently used for neuroblastoma prediction; MYCN and CD44. Wei and colleagues were therefore able to identify ten new prognostic genes, seven of which are expressed in neural tissue. However, no single marker alone was able to accurately classify patients. Although the scientific community anticipated that single molecular markers, such as TP53, could have sufficient
power to accurately classify tumours according to behaviour and response to treatment, this approach is now recognised as unrealistic because of the complexity and interdependence of molecular mechanisms leading to the behaviour of diseases. The future of molecular diagnostics and prognostics undoubtedly lies in validation of panels of selected markers, which will complement conventional pathological classifications. Wei and colleagues4 elegantly showed that as few as 19 genes can predict outcomes for neuroblastoma, and other reports have produced similar results in other cancers.9 These novel experimental and analytical approaches are invaluable, and they require urgent attention and investment. It is likely they will pave the way to translate molecular knowledge from the bench to the bedside, a much awaited result of the concerted efforts of scientists and clinicians in years to come—an outcome that will benefit our patients.
*Freddie C Hamdy, James W F Catto Academic Urology Unit, University of Sheffield, Royal Hallamshire Hospital, Sheffield S10 2JF, UK f.c.hamdy@sheffield.ac.uk We declare that we have no conflict of interest. 1
2
3 4
5 6
7
8
9
Golub TR, Slonim DK, Tamayo P, et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 1999; 286: 531–37. Alizadeh AA, Eisen MB, Davis RE, et al. Distinct types of diffuse large -cell lymphoma identified by gene expression profiling. Nature 2000; 403: 503–11. Scherf U, Ross DT, Waltham M, et al. A gene expression database for the molecular pharmacology of cancer. Nat Genet 2000; 24: 236–44. Wei JS, Greer BT, Westermann F, et al. Prediction of clinical outcome using gene expression profiling and artificial neural networks for patients with neuroblastoma. Cancer Res 2004; 64: 6883–91. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995; 346: 1075–79. Burke HB, Goodman PH, Rosen DB, et al. Artificial neural networks improve the accuracy of cancer survival prediction. Cancer 1997; 79: 857–62. Catto JW, Linkens DA, Abbod MF, et al. Artificial intelligence in predicting bladder cancer outcome: a comparison of neuro-fuzzy modeling and artificial neural networks. Clin Cancer Res 2003; 9: 4172–77. Shimada H, Stram DO, Chatten J, et al. Identification of subsets of neuroblastomas by combined histopathologic and N-myc analysis. J Natl Cancer Inst 1995; 87: 1470–76. Dyrskjot L, Thykjaer T, Kruhoffer M, et al. Identifying distinct classes of bladder carcinoma using microarrays. Nat Genet 2003; 33: 90–96.
Climate change, health, and development goals Recently, a consortium of development and environmental non-governmental organisations published Up in Smoke.1 This report acknowledges the fundamental interdependence of development and the environment. It recognises that humanity already faces protracted, mostly hazardous, environmental and social consequences of climate change—despite whatever immediate action we might take to mitigate emissions of greenhouse gases. 2004
Up In Smoke explores the effects of climate change on key development processes in poor countries: freshwater supplies, agricultural yields, physical safety, human settlements and infrastructure, health, and economic and gender equity. Overall, the omens are not good. Climate change will amplify existing hazards, deficits, and inequities, jeopardising the already low status of population health and wellbeing of disadvantaged populations. This bleak outlook www.thelancet.com Vol 364 December 4, 2004
fluorescence, and at the same maximise detection sensitivity?9 Other issues, such as how to develop efficient chemistry to produce quantum dot-conjugated biomolecules, and how quantum dot-labelled biomolecules (and cells or tissues) differ from their native counterparts, will also have to be addressed.10 However, the new findings indicate the feasibility of using dot-based imaging techniques to tackle one of the long-standing quests in cancer research: to find a way to precisely monitor and evaluate metastatic tumour cells in a patient.
2
Grace Y J Chen, *Shao Q Yao
7
Departments of Chemistry and Biological Sciences, National University of Singapore, Singapore 117543, Republic of Singapore. [email protected]
8
3
4
5
6
9
We declare that we have no conflict of interest 1
Voura EB, Jaiswal JK, Mattoussi H, Simon SM. Tracking metastatic tumor cell extravasation with quantum dot nanocrystals and fluorescence emission-scanning microscopy. Nature Med 2004; 10: 993–98.
10
Murray CB, Norris DJ, Bawendi, MG. Synthesis and characterization of nearly monodisperse CDE (E=S, SE, TE) semiconductor nanocrystallites. J Am Chem Soc 1993; 115: 8706–15. Dubertert B, Skourides P, Norris DJ, Noireaux V, Brivanlou AH, Libchaber A. In vivo imaging of quantum dots encapsulated in phospholipid micelles. Science 2002; 298: 1759–62. Akerman ME, Chan WC, Laakkonen P, Bhatia SN, Ruoslahti E. Nanocrystal targeting in vivo. Proc Natl Acad Sci USA 2002; 99: 12617–21. Wu X, Liu H, Liu J, et al. Immunofluorescent labeling of cancer marker Her2 and other cellular targets with semiconductor quantum dots. Nat Biotechnol 2003; 21: 41–46. Jaiswai JK, Mattoussi H, Mauro JM, Simon SM. Long-term multiple color imaging of live cells using quantum dot bioconjugates. Nat Biotechnol 2003; 21: 47–51. Larson DR, Zipfel WR, Williams RM, et al. Water-soluble quantum dots for multiphoton fluorescence imaging in vivo. Science 2003; 300: 1434–36. Gao X, Cui Y, Levenson RM, Chung LW, Nie S. In vivo cancer targeting and imaging with semiconductor quantum dots. Nat Biotechnol 2004; 22: 969–76. Kim S, Lim YT, Soltesz EG, et al. Near-infrared fluorescent type II quantum dots for sentinel lymph node mapping. Nat Biotechnol 2004; 22: 93–97. Srinivasan R, Yao SQ, Yeo DS. Chemical approaches for live cell bioimaging. Comb Chem High Throughput Screen 2004; 7: 597–604.
Less is more: artificial intelligence and gene-expression arrays
www.thelancet.com Vol 364 December 4, 2004
of gene expression, and Scherf et al3 studied 84 000 experiments (60 cell lines tested for 1400 agents) with 48 000 gene measurements (672 million outcomes measured). Therefore, improved bioinformatics are paramount in the interrogation of array data. So far, most array reports have used statistical classification methods, such as hierarchical clustering. An alternative approach was recently taken by Jun Wei and colleagues,4 using artificial neural networks to process array measurements of neuroblastoma. Such networks are computational models designed to replicate the neuronal networks of the human brain.5 With multiple layers of interconnected nodes, the network learns the importance of various data in predicting outcomes. Once a network is trained, it can analyse new data and reach high levels of predictive accuracy.
Rights were not granted to include this image in electronic media. Please refer to the printed journal. Science Photo Library
The development of gene-expression microarrays in the 1990s heralded a great future for molecular medicine. Single experiments that could simultaneously reveal the expression of thousands of genes in a cell promised to unravel the molecular pathways of numerous medical conditions, including cancer. Encouraging reports have used arrays to classify human cancers according to histology and to distinguish cancers that had been histologically inseparable.1 Examples are numerous, such as Alizadeh et al,2 who used an 18 000-gene array to classify diffuse large B-cell lymphoma into two types with different prognoses. Furthermore, the genes responsible for this stratification suggested the subtypes have different sites of origin, revealing new insights into this disease’s pathogenesis. Other researchers have used gene arrays to take a behavioural, rather than histological, view of disease classification. Scherf et al3 used an 8000-gene array to classify 60 tumour cell lines according to their chemosensitivity to 1400 agents. Improved technology has dramatically increased the capacity of gene arrays. We can now purchase a whole genome array for yeast and human arrays with over 33 000 genes. It is only a matter of time before whole humangenome arrays become readily available for more comprehensive research into the molecular mechanisms of disease. Although larger is usually better, increasing array capacity also increases the risk of statistical error and the amount of data that are generated and require analysis. For example, if one calculates the number of experiments described in previous reports, which used modest array sizes, the results are staggering. Alizadeh et al2 made 1·8 million measurements
Computer artwork of DNA microarray
2003
Comment
Although such networks appear more accurate than traditional statistical methods,6,7 their inner workings are hidden within a black box. The black box has hindered widespread acceptance of artifical neural networks, because it prevents validation and network interrogation. This lack of acceptance and hidden inner workings might be overcome with an alternative form of artificial intelligence based on fuzzy logic.7 Wei and colleagues4 studied 49 patients with neuroblastoma, the most common solid extracranial malignancy in childhood. To guide their treatment, patients were conventionally stratified into high, intermediate, and low risk, by several factors, including age at diagnosis, histological features, tumour stage, and amplification of the MYCN oncogene.8 As with most clinical classifications, this system is imperfect. About 30% of high-risk patients have prolonged survival, and 15% of low-risk patients progress despite treatment.4,8 Wei tried to improve the accuracy of this stratification by using artificial neural networks to model gene-array data. Having analysed 25 000 genes in each sample, the researchers used networks to identify the most relevant predictive genes. By repeating the model while varying the input for each individual gene, the researchers identified the genes that most altered the predictive accuracy of the model (19 genes). This selection was successfully validated in their analysis. A new model trained with these 19 genes more accurately predicted survival than both a previous model with 25 000 genes and the current clinical classification. Furthermore, when only high-risk patients (according to conventional criteria) were analysed, a 19-gene network was able to identify the small proportion of patients that would respond to treatment. It is the remarkable ease with which this approach identifies and validates important genes that highlights its very benefit. Of the 19 genes identified, seven are unknown (so-called expressed sequence tags; RNA sequences produced by cells whose gene identity is yet to be established), and two genes are currently used for neuroblastoma prediction; MYCN and CD44. Wei and colleagues were therefore able to identify ten new prognostic genes, seven of which are expressed in neural tissue. However, no single marker alone was able to accurately classify patients. Although the scientific community anticipated that single molecular markers, such as TP53, could have sufficient
power to accurately classify tumours according to behaviour and response to treatment, this approach is now recognised as unrealistic because of the complexity and interdependence of molecular mechanisms leading to the behaviour of diseases. The future of molecular diagnostics and prognostics undoubtedly lies in validation of panels of selected markers, which will complement conventional pathological classifications. Wei and colleagues4 elegantly showed that as few as 19 genes can predict outcomes for neuroblastoma, and other reports have produced similar results in other cancers.9 These novel experimental and analytical approaches are invaluable, and they require urgent attention and investment. It is likely they will pave the way to translate molecular knowledge from the bench to the bedside, a much awaited result of the concerted efforts of scientists and clinicians in years to come—an outcome that will benefit our patients.
*Freddie C Hamdy, James W F Catto Academic Urology Unit, University of Sheffield, Royal Hallamshire Hospital, Sheffield S10 2JF, UK f.c.hamdy@sheffield.ac.uk We declare that we have no conflict of interest. 1
2
3 4
5 6
7
8
9
Golub TR, Slonim DK, Tamayo P, et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science 1999; 286: 531–37. Alizadeh AA, Eisen MB, Davis RE, et al. Distinct types of diffuse large -cell lymphoma identified by gene expression profiling. Nature 2000; 403: 503–11. Scherf U, Ross DT, Waltham M, et al. A gene expression database for the molecular pharmacology of cancer. Nat Genet 2000; 24: 236–44. Wei JS, Greer BT, Westermann F, et al. Prediction of clinical outcome using gene expression profiling and artificial neural networks for patients with neuroblastoma. Cancer Res 2004; 64: 6883–91. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995; 346: 1075–79. Burke HB, Goodman PH, Rosen DB, et al. Artificial neural networks improve the accuracy of cancer survival prediction. Cancer 1997; 79: 857–62. Catto JW, Linkens DA, Abbod MF, et al. Artificial intelligence in predicting bladder cancer outcome: a comparison of neuro-fuzzy modeling and artificial neural networks. Clin Cancer Res 2003; 9: 4172–77. Shimada H, Stram DO, Chatten J, et al. Identification of subsets of neuroblastomas by combined histopathologic and N-myc analysis. J Natl Cancer Inst 1995; 87: 1470–76. Dyrskjot L, Thykjaer T, Kruhoffer M, et al. Identifying distinct classes of bladder carcinoma using microarrays. Nat Genet 2003; 33: 90–96.
Climate change, health, and development goals Recently, a consortium of development and environmental non-governmental organisations published Up in Smoke.1 This report acknowledges the fundamental interdependence of development and the environment. It recognises that humanity already faces protracted, mostly hazardous, environmental and social consequences of climate change—despite whatever immediate action we might take to mitigate emissions of greenhouse gases. 2004
Up In Smoke explores the effects of climate change on key development processes in poor countries: freshwater supplies, agricultural yields, physical safety, human settlements and infrastructure, health, and economic and gender equity. Overall, the omens are not good. Climate change will amplify existing hazards, deficits, and inequities, jeopardising the already low status of population health and wellbeing of disadvantaged populations. This bleak outlook www.thelancet.com Vol 364 December 4, 2004