prop01

=Crowdsourcing Chemistry and Modeling using Open Notebook Science= Jean-Claude Bradley, Rajarshi Guha and Antony Williams submitted to [|NSF CDI] December 8, 2008 [|PDF version]

This project aims to synergize the following tools and processes: Open Notebook Science, crowdsourcing, laboratory automation, a chemical structure repository, cheminformatics and QSPR modeling. Crowdsourcing is a term introduced by Jeff Howe to describe the solution of problems through a distributed network of people.1 Although there are several examples of crowdsourcing in science, most are not open. The best known example is Innocentive, where scientific problems are made public with prizes for the solvers.2 However, the proposed solutions are not made public and accepted solutions are intended to be proprietary to the company funding the solution. Innocentive is also collaborating with the Rockefeller Institute for help to combat third world and orphan diseases.3 [AJW1]  non-commercial examples in science include Stardust@home4 and GalaxyZoo,5 initiatives designed to identify astronomical objects. Some recent examples of crowdsourcing in chemistry are Chemmunity,6 the Synaptic Leap,7 OrgList,8 Chemists Without Borders,9 ChemUnPub10 and NineSigma.11 The term "Open Notebook Science" was coined to represent a form of Open Science where the laboratory notebook is made public in as close to real time as possible.12,13 The Bradley lab has demonstrated the feasibility of carrying out Open Notebook Science since the summer of 2005 with the UsefulChem project.14 Benefitting from the use of monthly prizes and open to students from the US and UK, the Open Notebook Science Challenge, presently underway, is another example.15 Experiments are stored on wiki pages, much like in a paper notebook, but with hyperlinks to all the raw data collected. There is also a blog to report on the overall progress of the scientific work.16 The UsefulChem blog and wiki receive about 200 hits per day. Over time, collaborations with other scientists have evolved. Rajarshi Guha at Indiana University and Tsu-Soo Tan from Nanyang Polytechnic in Singapore have invested significant amounts of time in running docking calculations for UsefulChem virtual libraries and reporting their results openly, in near real-time. Dan Zaharevitz, from the National Cancer Institute has contributed by testing compounds for potential anti-tumor activity. Phil Rosenthal, from UCSF, has process of testing some compounds for anti-malarial activity. Other individuals from around the world have contributed by engaging in open discussions in the blogosphere. The UsefulChem project has also been cited in the peer-reviewed literature,17-18 including an article in a peer-reviewed journal citing UsefulChem lab notebook pages as primary references.19 These examples of spontaneous collaboration clearly indicate that there is a huge potential for doing science under open conditions. However, in order to take full advantage of this potential, additional elements must come into play. Currently, experimental results on the wiki pages are written in a format amenable to human interpretation. The system could be made much more powerful by representing the information in a format understandable by machines as well, so as to contribute to the growing semantic web. As a step in this direction, each experiment page has a “tag” section at the end where more machine-friendly representations of molecules used in the experiment are listed. As an example, InChI codes are alphanumeric representations of molecules using standards supported by IUPAC. Since these InChIs are indexed by search engines, such as Google, it permits an unambiguous means of retrieving relevant information about each chemical. Other means of representing chemicals (SMILES, chemical names) suffer from having non-unique representations. This tagging system works for simple retrieval of an exact chemical structure but fails to allow more sophisticated queries, such as, identifying chemical substructure, similarity of structure, whether a chemical is a reagent or a product, the reaction conditions, and details regarding characterization of the materials, etc. ChemSpider20 is a free access online database working to build a structure community for chemists and offers some of the necessary capabilities to facilitate this crowdsourcing project. The system has been built specifically with the intention of hosting chemical structures and related information such as analytical data. The system presently hosts over 20 million unique chemical structures and is the first online system to facilitate public depositions of single structures or collections, property data, unstructured annotations and analytical data. ChemSpider (CS) has been designed with the needs of the chemist in mind and over 5000 chemists per day use the system for the purpose of searching for data and information associated with chemical entities. The system has already started to deliver on its promise to facilitate connectivities via the semantic web by providing open web services to allow other platforms to integrate and derive value from the information available online. CS has already committed to two directions of related interest to this project : Development of a platform for collaboration between chemists21 and the development of a structure-based encyclopedia for chemists, WiChempedia.22 These developments will provide core functionality to support our Open Notebook Science efforts. Additional capabilities will be introduced on an as-needed basis to facilitate our efforts in terms of the support of experimental data capture, control and reporting as outlined below. Another component of this partnership is the automation of the execution of some of the experiments. This will allow us to take scientific crowdsourcing to the next level, where the online community can not only provide feedback but also directly carry out simple experiments. We propose to incorporate a Mettler-Toledo Miniblock system with a MiniMapper automated liquid handler. This system has already proven effective in optimizing the Ugi reaction.19 The MiniMapper software provides an XML log file and we have demonstrated a simple way to transfer reaction conditions from an open Google Spreadsheet to program the robot.19 This set-up is ideal for crowdsourcing since anyone can request experiments to be performed without having to grasp the controls for the MiniMapper. The final component is the use of computational methods to prioritize molecules for experimental study, avoiding the time and effort involved in testing all available compounds. One approach is the use of Quantitative Structure Property Relationship (QSPR) models. A QSPR model employs statistical methods to correlate molecular structural features of tested molecules to a property (such as solubility) and then predict that property for untested molecules. More importantly, the development of QSPR models is an iterative process, whereby one experimentally tests a set of molecules, develops a predictive model, obtains predictions for new molecules and then experimentally tests them for confirmation. The results of the confirmatory experiments can be fed back into the modeling process to enhance the predictive performance. In addition to QSPR models, one can employ diversity analysis23 to identify classes of molecular structures to focus on. This is useful in identifying molecules that have not been well-studied, allowing resources to be focused on them. Common techniques include principal components analysis (PCA) and k-nearest neighbor methods. In combination these methods allow crowd-source collaborations to be much more time- and cost-efficient. Given the distributed nature of crowdsourcing, it is vital that the computational methods also be available in a distributed fashion. Web services have been employed in a variety of fields, and the availability of SOAP based cheminformatics24 and statistical25 services, allows the integration of of raw exprimental data (from Wiki's and Google Docs), structure information (from repositories such as ChemSpider) with computational models and tools. The following scientific objectives will be initiated as projects: Since the Ugi reaction is of value to the generation of product libraries in multiple applications, the ability to predict the crystallization event should be of significant value to the chemistry community. By generating models to accurately predict precipitation, virtual libraries being screened for a desired activity can be further screened for ease of preparation. A better understanding of the solubility of Ugi products in different solvents using more sophisticated models may also recommend different reaction conditions for those products that do not crystallize from the standard protocol using methanol. The simplicity of this initial project is well suited for crowdsourcing. The results (precipitate/no precipitate) are simple and easy to tabulate. Anyone can propose a model and suggest combinations of starting materials to most efficiently test it. The experiments are simple to execute, either manually or with automation assistance,19 and observe. In the case of precipitates, all products will be characterized by standard techniques (1H NMR, 13C NMR, IR, MS, HRMS). Surprisingly, the solubilities of readily available aldehydes, carboxylic acids, amines and isonitriles in common non-aqueous solvents are not widely available, either from the peer-reviewed literature or databases, whether freely accessible or not. Clearly, making these measurements is the first step in understanding how to control the course of the Ugi reaction by changing solvents. But by making our measurements publicly available in real time, we are providing valuable information for all chemists interested in selecting solvents for other reactions with these compounds. Thus, measuring solubilities has been the core focus of the Open Notebook Science challenge.15 So far, students from Drexel and Southampton Universities have participated. Storing results in a Google Spreasheet document has enabled Rajarshi Guha to create a web query via the Google Visualization API. This has provided an extremely convenient way to collect and search for information. Thus measurements of the same compound in the same solvent can be viewed together. A link to the lab notebook page where the measurement was detailed is available for each number.
 * Jean-Claude Bradley (Drexel University, PI)** is an Associate Professor of Chemistry and E-Learning Coordinator for the College of Arts and Sciences at Drexel University. He leads the UsefulChem project, an initiative started in the summer of 2005 to make the scientific process as transparent as possible by publishing all research work in real time to a collection of public blogs, wikis and other web pages. Jean-Claude coined the term Open Notebook Science (ONS) to distinguish this approach from other more restricted forms of Open Science. He established the ONS Challenge in the fall of 2008, open to students in the US and the UK for the measurement of solubilities in non-aqueous solvents, with prizes through sponsorship from Submeta, Nature and Sigma-Aldrich. The main chemistry objective of the UsefulChem project is currently the synthesis and testing of novel anti-malarial agents and the measurement of non-aqueous solubility for readily available organic compounds. The cheminformatics component aims to interface as much of the research work as possible with autonomous agents to automate the scientific process in novel ways. Jean-Claude teaches undergraduate organic chemistry and chemical information retrieval courses with most content freely available on public blogs, wikis, games and audio and video podcasts. Openness in research meshes well with openness in teaching. Real data from the laboratory can be used in assignments to practice concepts learned in class. Jean-Claude has a Ph.D. in organic chemistry and has published articles and obtained patents in the areas of synthetic and mechanistic chemistry, gene therapy, nanotechnology and scientific knowledge management.
 * Rajarshi Guha (Indiana University, co-PI)** is a visiting Assistant Professor in the School of Informatics at Indiana University. His area of research focus on the development of cheminformatics methodologies to enhance various aspects of the drug discovery process. Much of his research focuses on the development of novel methods for Quantitative Structure-Activity Relationship (QSAR) models that enhance their reliability and interpretability. This research has been applied to a number of biological systems including artemisinin analogs, PDGFR inhibitors and cell proliferation agents. More recently he has developed methods to characterize structure-activity relationship (SAR) data using a novel network paradigm. In addtion to QSAR modeling, he is also addressing efficient ways to characterize and explore chemical spaces of arbitrary dimensionality. These methods make use of nearest neighbor algorithms and have been employed in the analysis of combinatorial libraries and determination of cluster cardinality in an a priori fashion. Along with methodology development he has contributed extensively to the software implementations in cheminformatics. Examples include signifcant contributions to the Chemistry Development Kit (CDK ) and Java toolkit for cheminformatics, development of a cheminformatics web service infrastructure and the development of a statistical web service ifnrastructure (based on R) that supports access to specific machine learning routines as well deployment of prebuilt statistical models. He has a Ph.D. in chemistry and has published extensively in peer reviewed journals.
 * Antony Williams (ChemSpider, co-PI)** is a Senior Fellow at the National Institute of Statistical Sciences, is the President of ChemZoo Inc., and is the host of ChemSpider, a Structure Centric Community for Chemists. Antony has worked in academia, in industry (Eastman-Kodak) and in the commercial cheminformatics sector in the role of Chief Science Officer. He has authored and co-authored over 100 scientific publications and multiple invited book chapters. He has two issued patents and has one pending. His formal training is as an NMR Spectroscopist with a focus on small molecule structure elucidation and analytical data processing algorithms. the last decade of his career has been focused on the development of integrated sample, structure and analytical data management systems at the desktop and at the enterprise level utilizing web-based technologies. In the past year he has established a free access website, ChemSpider, to provide access to chemistry-related information for almost 20 million chemical entities and their associated data. He is interested in all aspects of cheminformatics, with special interest in chemical structure handling, nomenclature and computer-assisted structure elucidation. He conducts research in the area of data processing techniques for spectroscopy and development of the semantic web.
 * Justification of Intellectual Partnership **
 * Research program **
 * 1) Determining optimal conditions for the precipitation of Ugi products.** The Ugi reaction provides a rapid access to large combinatorial spaces by bringing together an amine, an aldehyde, a carboxylic acid and an isonitrile.26-28 In addition, reaction conditions are generally very convenient: methanol at room temperature is a common protocol. The Bradley group has used the Ugi reaction to synthesize potential anti-malarial compounds.29 We observed that sometimes the products simply crystallized from the reaction mixture. Although NMR monitoring indicated that the reaction proceeded smoothly in most cases, the products did not always precipitate. The ability to purify products by a simple filtering translates into a significant advantage compared against the alternatives, which would usually consist in chromatography, a costly and inefficient process that severely limits scaling up.
 * 2) The determination of solubility.** The Ugi reaction has been widely used to prepare diverse libraries of compounds.26-28 Based on this success we hypothesize that the Ugi reaction generally proceeds in high yields under a variety of conditions but does not usually provide the product in pure form because the solubility conditions have not been optimized. If that is true, it may be possible to generally design reaction conditions fully optimized to generate maximum precipitate by searching for the lowest solubility of the Ugi product while still solubilizing all the starting materials at a sufficiently high concentration (generally above 2 molar prior to mixing).19

This level of detailed is far beyond that provided in the experimental section of peer-reviewed articles or even Supplementary Data sections. The access to the laboratory notebook allows a reviewer to understand what has and has not been measured and reported. This has proved very effective in including or excluding measurements done by different students using different methods.30 There are currently five judges for the ONSchallenge, ranging from professors and postdocs in academia to a company president. The ongoing feedback provided by the judges ensures reasonable quality control, effectively generating peer-reviewed Open Notebooks. We will evaluate molecular descriptors for the currently tested molecules, using the Chemistry Development Kit (CDK).32 Due to the large number of descriptors available, we will employ a genetic algorithm to perform feature selection33 for each type of model. This will allow us to focus on a small subset of molecular descriptors, thus avoiding one source of overfitting. All models will undergo cross-validation to ensure that they are not over-fit. Finally, we will define a domain of applicability34 for the models allowing users to obtain a measure of confidence in the predictions. The final predicted solubility will be taken as the mean of the individual predictions. Since one of the aims of the ONSsolubility challenge is to measure solubility in multiple solvents, we will develop panels of models, covering all solvents as well as single, monolithic models that incorporate both solvent and solute information. In addition to predicting solubility, we will employ PCA based diversity analysis to suggest classes of structural features that deserve further consideration. We will employ pre-defined subsets of physicochemical descriptors and visualize the distribution of molecules in the space defined by the first two or three principal components. This will allow us to evaluate which regions of the space (i.e., types of structural features) are yet to be explored and which regions are over-represented. Both graduate and undergraduate students will acquire the following skills: 1) Howe, J. The Rise of Crowdsourcing, Wired, June 2006. (http://www.wired.com/wired/archive/14.06/crowds.html) 2) [|http://www.innocentive.com] 3) http://www.innocentive.com/servlets/project/Pavilion.po?p=Rockefeller%20Foundation 4) http://stardustathome.ssl.berkeley.edu/ 5) [|http://www.galaxyzoo.org] 6) [|http://www.chemmunity.com] 7) [|http://www.thesynapticleap.org] 8) [|http://www.orglist.net] 9) [|http://www.chemistswithoutborders.org] 10) [|http://www.chemunpub.it] 11) http://www.ninesigma.com 12) http://drexel-coas-elearning.blogspot.com/2006/09/open-notebook-science.html 13) http://en.wikipedia.org/wiki/Open_Notebook_Science 14) http://usefulchem.wikispaces.com/All+Reactions 15) |http://onschallenge.wikispaces.com 16) [|http://usefulchem.blogspot.com] 17) Todd, M. Open Access and Open Source in Chemistry, Chemistry Central Journal 2007, 1:3. (doi:10.1186/1752-153X-1-3) 18) Lancashire, R. The JSpecView Project: an Open Source Java viewer and converter for JCAMP-DX, and XML spectral data files, Chemistry Central Journal 2007, **1:**31. (doi:10.1186/1752-153X-1-31) 19) Bradley, Jean-Claude; Mirza, Khalid; Owens, Kevin; Osborne, Tom and Williams, Antony (November 2008). "[|Optimization of the Ugi reaction using parallel synthesis and automated liquid handling]". //Journal of Visualized Experiments//. [|doi]:[|10.3791/942], 20) [|http://www.chemspider.com] 21) Williams, A. J. http://www.chemspider.com/blog/the-chemspider-team-chooses-our-future-platform-for-collaboration-microsoft-sharepoint.html 22) Williams, A.J. http://www.chemspider.com/blog/wichempedia-is-now-on-its-way.html). 23) Martin, Y. C.; "Diverse Viewpoints on Computational Aspects of Molecular Diversity", J. Comb.Chem., 2001, 3, 231-250 24) Dong, X., Gilbert, K., Guha, R., Heiland, R., Kim, J., Pierce, M., Fox, G., and Wild, D. J.; "A Web Service Infrastructure for Chemoinformatics", J.~Chem.~Inf.~Model., 2007, 47, 1303--1307 25) Guha, R.; "A Flexible Web Service Infrastructure for the Development and Deployment of Predictive Models", J. Chem. Inf. Model., 2008, 48, 456--464 26) Marcaccini, S and Torroba, T. [|The use of the Ugi four-component condensation], Nature Protocols 2(3), 632 (2007). 27) Domling, A. and Ugi, I. [|Multicomponent reactions with isocyanides] Angew. Chem. Int. Eng. Ed. 39, 3168 (2000). 28) Domling, A. [|Recent Developments in Isocyanide Based Multi-Component Reactions in Applied Chemistry], Chem. Rev. 106, 17 (2006). 29) Bradley, Jean-Claude UsefulChem blog, January 25, 2007 (http://usefulchem.blogspot.com/2008/01/we-have-anti-malarial-activity.html) 30) Bradley, Jean-Claude UsefulChem blog, November 21, 2008. (http://usefulchem.blogspot.com/2008/11/what-is-solubility-of-vanillin-in.html) 31) Johnson, S. R., Chen, X. Q., Murphy, D., and Gudmundsson, O.; "A Computational Model for the Prediction of Aqueous Solubility That Includes Crystal Packing, Intrinsic Solubility, and Ionization Effects ", Mol.Pharmaceutics, 2007, 4, 513--523 32) Steinbeck, C., Han, Y. Q., Kuhn, S., Horlacher, O., Luttmann, E., and Willighagen, E.; "The Chemistry Development Kit ({CDK}): An Open-Source {J}ava Library for Chemo- and Bioinformatics", J. Chem. Inf. Comput. Sci., 2003, 43, 493—500 33) Guha, R. and Jurs, P. C.; "Development of Linear, Ensemble, and Nonlinear Models for the Prediction and Interpretation of the Biological Activity of a Set of {PDGFR} Inhibitors.", J. Chem. Inf. Comput.Sci., 2004, 44, 2179--2189 34) Benigni, R. and Bossa, C.; "Predictivity of {QSAR}", J. Chem. Inf. Model., 2008, 48, 971--980
 * 3) Modeling of solubility and prioritizing compounds.** We will develop statistical models of solubility based on experimental data made available as part of the ONS solubility Challenge. While previous work has considered factors such as crystal packing and thermodynamic factors31 //ab initio// computation of these properties is time consuming and requires crystal structures, which we will not in general, have access to. Thus, we will employ a statistical approach using and ensemble of linear regression and neural network models, based on molecular descriptors.
 * 4) The development of web services.** Initially, computational results will be reported on the project wiki as they are obtained. Direct, distributed access to solubility models and diversity analysis will be achieved by deploying them in a cheminformatics and statistical web service infrastructure.24-25 The infrastructure will be enhanced to support automated model development, allowing decentralization of the model development process.We will develop exemplar applications using these services such as real-time exploration of the chemical space of tested compounds, linking the web service deployed diversity analysis method to the live data in the Google Docs. It should be noted that these deployed models and services are not restricted to our project. Since they are all SOAP based services, any client located on the Internet can freely access them.
 * Timeline **
 * Year 1:** During the first year, the main focus will be on measuring solubility and finding collaborators from the online community. Experiments will be performed manually until the automatic reactor has been installed and students trained to operate it. Informatics tools to represent experimental workflows and reactions through their various entity-state-increment-transition relationships will be developed. In a reaction workflow at any point in time chemical entities are involved and are at a particular state (mixed, unmixed, heated, stirred) for a particular time increment and then pass through a transition to another state (analytical data acquired, new material added etc.). These relationships can be described in a logical sequence and described in a manner allowing the data to be mined throughout a workflow sequence. Informatics tools to describe, capture, manage and datamine will be developed. As experimental measurements are deposited we will develop QSPR models and report results on the wiki. We will also perform PCA based diversity analysis to suggest classes of molecules for testing. Results will be reported manually on the wiki.
 * Year 2:** The automated reactor will be fully implemented to run Ugi reactions in parallel and made available for direct operation by the crowdsourcing community under the oversight of the PI and co-PIs. Enhanced workflow management tools to generate workflow documents capable of driving the automation systems will be developed. Data will be generated in a standard manner to allow data capture and management on the ChemSpider platform. The web service infrastructure to support automated model development and PCA analysis will be developed. A web application to show results of diversity analysis in real time will be implemented.
 * Year 3:** Projects suggested by the crowdsourcing community will be evaluated and resources will be allocated to execute the experiments, under supervision of the PI and co-PI. As sub-projects get completed, work will be submitted for traditional peer reviewed publication. Arguments in the papers will be supported by links to the original experiments in the online open notebook wiki, giving unprecedented systematic access to experimental raw data to be re-analyzed or re-purposed by anyone. Co-authorships will be based on documented contributions from members in the community who sufficiently participated. More applications based on web service infrastructure will be developed that integrate raw experimental data and the structure repository.
 * Educational program**
 * 1) Organic chemistry skills.** Students working in the Bradley laboratory will learn standard organic chemistry techniques, including executing experiments then isolating and fully characterizing products with conventional spectroscopic analysis (NMR, IR, MS, HRMS).
 * 2) Networking.** Students will be required to record their experiments using an Open Notebook on a wiki and will receive and respond to comments from the online world. Maintaining a public laboratory notebook can be a very efficient way to learn about the proper way to document an experiment because the adviser and other interested parties can provide immediate and ongoing feedback, which is impossible with a conventional closed paper notebook. They will also be encouraged to engage in conversation about their project on various social networks, including mailing lists (e.g. OrgList, UsefulChem), our collaborators' blogs and wikis, Facebook, Nature Networks, SciVee, Flickr, FriendFeed, etc. Not only will interacting with peers and mentors be valuable as a learning experience, the contacts formed may be helpful for the progress of the student's career after graduation.
 * 3) Automatic reactor programming.** Students working in the Bradley laboratory will be charged with setting up, running and maintaining the automated chemical synthesis system. Coupled with their work doing conventional manual organic synthesis, these students will be competitively trained to enter the chemistry workforce.
 * 4) An understanding of cheminformatics.** In order for the experiments to be properly indexed by ChemSpider and other automatic aggregators, students will have to learn how to generate and use representations of molecules using modern cheminformatics techniques (e.g. use structure drawing tools for the generation of SMILES, InChI, InChIKey). They will also need to become familiar with tools for the manipulation of analytical data, image management tools etc and will learn to order their data management processes and behaviors in a logical manner in order to capture, document and mine data in an electronic environment.
 * Desired Project Outcomes**
 * 1)** From the larger perspective, a key outcome is the establishment of a precedent and model to demonstrate how crowdsourcing can be used to solve scientific problems. As the project evolves, progress will be reported on social software networks. This type of dissemination has proved effective to identify collaborators in the online world for the UsefulChem project, primarily through the use of blogs. Another key advantage is that reports of what appears to work or not work are generally followed quickly with helpful discussions in the blogosphere, in addition to being helpful to others with related projects. The demonstration that an automatic reactor can be effectively added to the scientific crowdsourcing toolkit is also a key outcome.
 * 2)** At a more basic level, meeting the scientific objectives outlined in this proposal will benefit the scientific community in a more direct way. The non-aqueous solubility data alone is expected to be used for various applications in the chemistry community. The Ugi reaction is already used by many groups and further understanding of how to prepare these products using a much simpler purification protocol can be quite valuable in many areas. For example, several of the Ugi products from UsefulChem project have shown anti-malarial activity.
 * References**