Where is the knowledge we have lost in information?
Where is the wisdom we have lost in knowledge?
--T.S. Elliot, "The Rock"
Bioinformatics, the study of how information is represented and transmitted in biological systems, is a data-intensive field of research and development. It encompasses networking, databases, visualization techniques, search engine design, statistical techniques, modeling and simulation, AI and related pattern matching, and (the subject of this article) data mining. In bioinformatics, data mining is concerned with discovering how simple base pairs can be combined in different ways, many of which are unknown, to provide the form and function of the larger building blocks of life. Mastering this biological dataincluding discovering many of the underlying rules, relationships, and meaningsrequires human intelligence and intuition, leveraged by computer-based tools.
Because of automated gene sequencing machines and new worldwide activity in the field, both experimental (wet lab) and computer-generated data are increasing at an exponential rate. Consider the growth in the holdings of GenBank and Swiss-Prot, major online nucleotide sequence and protein sequence databases, respectively. As illustrated in Figure 1, about 90% of the entries in both databases have been made since 1998, when Celera Genomics entered the human genome sequencing race against the nearly decade-old government-sponsored activity. PubMed, the major online biomedical bibliographic database, has experienced similar growth.
Figure 1 Growth of the GenBank (nucleotide sequences) and Swiss-Prot (protein sequences) databases from 1980 through 2002.
The increase in the holdings of GenBank, Swiss-Prot, and PubMed mirrors the growth of the hundreds of public and private online databases that reflect the work of thousands of researchers in laboratories around the world who are engaged in mass-producing biological data. There are more data to deal with today because modern researchers are using computer-enabled, data-centric, high-throughput processes, such as automated sequencing machines and microarrays. These researchers are looking for data about the structure of the protein in order to allow, for example, the design of molecules to match key regions of the protein. In this way, designer drugs can be synthesized to catalyze or block reactions involving the protein. Similarly, an increasing proportion of the data is derived from mining and manipulating data from other databases, as opposed to direct experimental methods. For example, there are dozens of labs around the world focused on predicting protein structure from sequence data, as opposed to the traditional time-consuming method of direct observation.
Getting at the hard-won sequence and structure data in molecular biology databases and the functional data in the online biomedical literature is complicated by the size and complexity of the databases. Exhaustively searching for the raw data and performing the transformation and manipulations on the data through manual operations is often impractical. However, even when computer resources are available, the time and computational resources required to locate and manipulate the data are limiting factors. As a result, executing exhaustive, non-directed searches for potential correlations isn't possible. Without an organizing theme, the billions of data points from genomic or proetomic studies are of little value. Regardless of whether this categorization is at the base pair, chromosome, or gene level, an organizing theme is critical because it simplifies and reduces the complexity of what could otherwise be a flood of incomprehensible data. For example, the PubMed, Swiss-Prot, and GenBank databases represent generally recognizable organizational themes that facilitate use of their contents. At a higher level, our understanding of health and disease is facilitated by the organization of clinical research data by organ system, pathogen, genetic aberration, or site of trauma.
Ideally, the creator and users of the database share an understanding of the underlying organizational theme. These themes and the tools used to support them determine how easily databases created for one purpose can be used for other purposes. For example, in a relational database of gene sequences, the data may be arranged in tables, and the user may need to construct SQL (structured query language) statements to search for and retrieve data. However, if inherited diseases organize the relational database, it may not readily support an efficient search by protein sequence.
The challenge for researchers looking in the exponentially increasing quantities of microbiology data for assumed and unknown relationships can be formidable. Even simple queries may involve creating relatively complicated, computationally intensive joins in order to create views that support a given hypothesis of how data are related. In addition, even if the technology is available that allows a researcher to specify any hypothetical query, the potential for discovering new relationships in data is a function of the insights and biases imposed by the researcher. While these limitations may be problematic in relatively small databases, they may be intolerable in databases with billions of interrelated data elements.
To avoid the computational constraints imposed by these large molecular biology databases, researchers frequently turn to biological heuristics to avoid exhaustive searches or processes with a low likelihood of success. For example, in hunting for new genes, a good place to start from a statistical perspective is near sequences that tend to be found between introns and exons. However, even with heuristics, user-directed discovery is inherently limited by the time required to manually search for new data.
The aim of this article is to introduce data mining techniques as an automated means of reducing the complexity of data in large bioinformatics databases and of discovering meaningful, useful patterns and relationships in data. The following sections provide an overview of the methods, technologies, and challenges associated with data mining.
Data mining is one stage in an overall knowledge discovery process. As illustrated in Figure 2, this process involves selection and sampling of the appropriate data from the database(s); preprocessing and cleaning of the data to remove redundancies, errors, and conflicts; transforming and reducing data to a format more suitable for the data mining; data mining; evaluation of the mined data; and visualization of the evaluation results.
Figure 2 Data mining in the larger context of the knowledge discovery process.
In most cases, several iterations of the knowledge discovery process are required, each involving the design of new data queries to test new hypothesesIn addition, although the process may seem straightforward, data mining and the overall knowledge discovery process involve much more than the simple statistical analysis of data. For example, difficult-to-describe metrics, such as novelty, interestingness, and understandability, are often used to define data mining parameters for data discovery. Similarly, each phase of the knowledge discovery process has associated challenges, as outlined here.
Selection and Sampling
Because of practical computational limitations and a priori knowledge, data mining isn't simply about searching for every possible relationship in a database. In a large database or data warehouse, there may be hundreds or thousands of valueless relationships. Because there may be millions of records involved and thousands of variables, initial data mining is typically restricted to computationally tenable samples of the holding in an entire data warehouse. The evaluation of the relationships that are revealed in these samples can be used to determine which relationships in the data should be mined further using the complete data warehouse. With large complex databases, even with sampling, the computational resource requirements associated with non-directed data mining may be excessive. In this situation, researchers generally rely on their knowledge of biology to identify potentially valuable relationships, and they limit sampling based on these heuristics.
Preprocessing and Cleaning
The bulk of work associated with knowledge discovery is in preparing the data for the actual analysis associated with data mining. The major preparatory activities include the following:
Data Characterizationcreating a high-level description of the nature and the content of the data to be mined.
Consistency Analysisdetermining the statistical variability in the data, independent of the domain.
Domain Analysisvalidating the data values in the larger context of the biology.
Data Enrichmentdrawing from multiple data sources to minimize the limitations of a single data source.
Frequency and Distribution Analysisweighing values as a function of their frequency of occurrence.
Normalizationtransforming data values from one representation to another.
Missing Value Analysisdetecting, characterizing, and dealing with missing data values.
Transformation and Reduction
In the transformation and reduction phase of the knowledge discovery process, data sets are reduced to the minimum size possible through sampling or summary statistics. For example, tables of data may be replaced by descriptive statistics such as mean and standard deviation.
Data Mining Methods
The process of data mining is concerned with extracting patterns from the data by using techniques such as classification, regression, link analysis, segmentation, or deviation detection. Classification involves mapping data into one of several predefined or newly discovered classes. Regression methods involve assigning data a continuous numerical variable based on statistical methods. One goal in using regression methods is to extrapolate trends from a few samples of the data. Link analysis involves evaluating apparent connections or links between data in the database. Deviation detection identifies data values that are outside of the norm, as defined by existing models or by evaluating the ordering of observations. Segmentation identifies classes or groups of data that behave similarly, according to an established metric. These methods of data mining are typically used in combination with each other, either in parallel or as part of a sequential operation.
In the evaluation phase of knowledge discovery, the patterns identified by the data mining analysis are interpreted. Typical evaluation ranges from simple statistical analysis and complex numerical analysis of sequences and structures to determining the clinical relevance of the findings.
Visualization of evaluation results can range from simple pie charts to 3-D virtual reality displays that can be manipulated by haptic (force feedback) controllers.