SBIR Phase I: Knowledge Extraction from Support Vector Machines Models of Non-linear Drug Data Sets

这个小企业创新研究（SBIR）第一阶段项目将研究从支持向量机（SVM）非线性模型中提取知识的可行性及其对药物数据集的结果。SVM是一种用于数据挖掘的统计机器学习方法。该项目参考了支持向量机在预测类药物化合物的活性和性质方面表现非常好的研究论文。非线性模型的可解释性的缺乏已经阻止了化学和制药行业对SVM的成功商业采用。该项目描述了使用药物数据集测试不同非线性特征发现方法的实验，以确定其商业化潜力。这项活动对改善非线性SVM模型的知识提取的更广泛影响包括从改善石油勘探到信用欺诈检测的广泛应用。该公司的目标是为制药行业提供创新的药物发现信息学软件研究工具。它将向工业界展示从这笔赠款中发现的创新能力。改进药物发现的研究工具的需求是巨大的。2002年，十大制药公司在药物发现研究上花费了330亿美元。数据挖掘方法的改进将有助于制药公司找到治疗艾滋病等致命疾病的新方法。预测建模可以用于虚拟筛选，优先排序，然后决定哪些化合物将在实验室中进行测试。这可以减少测试化合物的数量并提高其成功率。通过提供知识发现的创新方法，可以更深入地了解为什么化合物表现出治疗某些疾病所必需的活性和性质，从而改进衍生物。这一发现将帮助科学家更快地发现更好的药物。

This Small Business Innovation Research (SBIR) Phase I project will investigate the feasibility of knowledge extraction from Support Vector Machine (SVM) non-linear models and their results for drug data sets. SVMs are an accurate method of statistical machine learning for data mining. This project references research papers where SVMs have performed very well at predicting activity and properties for drug-like compounds. The lack of interpretability of non-linear models has prevented successful commercial adoption of SVMs by the chemical and pharmaceutical industry. This project describes experiments to test different methods of non-linear feature discovery using drug data sets in order to determine their potential for commercialization. The broader impacts of this activity for improving knowledge extraction of non-linear SVM models include a broad range of applications from improved oil exploration to credit fraud detection. The company's goal is to provide innovative drug discovery informatics software research tools to the pharmaceutical industry. It will present the innovative capabilities discovered from this grant to industry. The need for improved research tools for drug discovery is great. This is demonstrated by the fact that the top ten largest pharmaceutical companies spent over 33 billion dollars in 2002 on drug discovery research. Improvements to data mining methods will assist pharmaceutical companies in finding new cures to deadly diseases such as HIV. Predictive modeling can be employed to virtually screen, prioritize and then decide which of these compounds will be tested in the lab. This can reduce the number of compounds tested and increase their success rate. By providing innovative methods of knowledge discovery, deeper insight can be gained into why compounds exhibit activity and properties necessary to treat certain diseases, leading to improved derivatives. This insight will help scientists discover better drugs faster.