Machine-Learning Approach to Label-free Detection of new Bacterial Pathogens

描述(由申请人提供)：我们感谢所有评审者所花费的时间和努力，并感谢他们提供的有用的意见和建议。我们仔细审查了这些批评意见，我们高兴地看到，小组接受了我们的建议。审查人员在总结说明中表达了三个主要问题：(1)虽然调查小组很有资格，但我们的合作历史很短；(2)缺乏关于Bardot系统实际限制的细节；(3)项目中使用的机器学习技术被认为是相当标准的。下面，我们将简要讨论审查者的意见，并说明我们如何更改修订后的应用程序以应对批评。(1)Dundar博士于2008年秋季从工业界转到学术界，在这一点上，Rajwa博士(BarDot的发明者之一)和Dundar博士开始合作，寻找新的方法来解决表型筛选中非详尽定义的类别问题。这种科学伙伴关系立即产生了有趣的结果，在提交原始申请时，Dundar博士和Rajwa博士的第一份手稿正在审查中。对原始提案中提出的方法进行了测试，并将结果提交给ACM第15届年度SIGKDD知识发现和数据挖掘国际会议(KDD‘09)，这是该领域规模最大、最受尊敬的会议之一。在经过全面同行评审后，该手稿被接受为从551篇提交的论文中挑选出来的50篇定期论文之一[20]。在提交提案后，研究工作继续进行，并产生了另一种方法来解决本赠款申请中描述的问题。初步发现在一份新的手稿中报告，该手稿目前正在审查中[4]。(2)我们重写了我们提案中的背景和研究方法部分，以包括评价者要求的关于Bardot系统实际方面的信息，如准确性问题(D.3.2节)、遇到新的未知类别的频率(B.3.1节)和验证(D.3.1节)。(3)细菌的表型筛选和分类问题可以在穷尽(标准)或非穷尽的学习框架内定义。尽管我们同意，对BARDOT实施详尽的分类方法确实只需要相当标准的工具，但培训图书馆的非详尽性质的问题不能通过直接使用任何教科书级别的技术来解决。事实上，非穷尽定义的类集合的存在违反了大多数监督学习系统的基本假设。非穷尽定义类的问题是机器学习在表型分析中应用的主要障碍，因为可能的表型数量可能是无限的。在我们最初的提案中，我们认为使用非详尽定义的类集合进行学习仍然是一个非常具有挑战性的问题，并提供了证据表明，标准技术的简单扩展不能提供可接受的解决方案。随后，我们提出了一种基于类的贝叶斯模拟的新方法，并表明初步结果优于基准技术[4]。尽管这些初步结果看起来很有希望，但我们并不认为所描述的初步算法是最终和最终的，我们不相信在这一点上我们能够为这个复杂的问题提供准确的算法解决方案。如果我们能够做到这一点，这将意味着我们已经实现了所有的赠款目标。拟议研究的本质是找到已定义问题的答案，而答案将一直是未知的，直到工作完成。然而，积极的评价和KDD‘09会议评委对我们工作的接受，告诉我们我们正朝着正确的方向前进。在本申请的修正版本中，我们提出了一种基于Wishart先验的修正贝叶斯方法(D.2.3节)。该算法动态地创建新的类，并用更新后的类集评估最大似然，逐步提高对未来样本的检测精度。我们相信，这比以前的方法有了很大的改进。因此，C部分的初步结果进行了更新，以反映我们的进展情况。由于改进的技术允许使用相同的算法对非穷举和穷举集进行分类，因此我们在修订的申请中将先前的特定目标3和5合并为一个。

DESCRIPTION (provided by applicant): We appreciate the time and effort spent by all the reviewers, and we are grateful for the useful comments and provided suggestions. We have carefully reviewed the critiques and we are happy to see that the panel was receptive to our proposal. The reviewers expressed three major concerns in the summary statement: (1) although the investigating team is well qualified our history of collaboration is short; (2) details regarding the practical constraints of the BARDOT system are lacking; (3) the machine learning techniques employed in the project are considered fairly standard. Below we briefly discuss the reviewers comments and indicate how we have changed our revised application to address the critique. (1) Dr. Dundar moved from industry to academia in the fall of 2008, at which point Dr. Rajwa (one of the original inventors of BARDOT) and Dr. Dundar began their collaboration on new approaches to the problem of non-exhaustively defined classes in phenotypic screening. This scientific partnership immediately produced interesting results, and at the time of submission of the original application, Dr. Dundar and Dr. Rajwa had their first manuscript under review. The approach presented in the original proposal was tested and the results were submitted to the ACM 15th Annual SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD'09), which is the largest and one of the most respected conferences in this field. The manuscript was accepted after a full peer review as one of the 50 regular papers selected from 551 submissions [20]. Following the proposal submission, research efforts continued and produced yet another approach to the problem described in this grant application. The preliminary findings are reported in a new manuscript which is currently under review [4]. (2) We rewrote the background and research methods sections of our proposal to include information re- quested by the reviewers regarding practical aspects of the BARDOT system, such as accuracy issues (Section D.3.2), frequency of encountering new, unknown classes (Section B.3.1), and validation (Section D.3.1). (3) The problem of phenotypic screening and classification of bacteria can be defined within exhaustive (stan- dard) or non-exhaustive learning frameworks. Although we agree that the implementation of an exhaustive clas- sification approach for BARDOT does require only fairly standard tools, the problem of the non-exhaustive nature of training libraries cannot be addressed by straightforward use of any textbook-level technique. In fact, the presence of non-exhaustively defined set of classes violates basic assumptions for most supervised learning systems. The issue of non-exhaustively defined classes is the major obstacle for application of machine learning in phenotypic analysis since the number of possible phenotypes may be infinite. In our original proposal we argued that learning with a non-exhaustively defined set of classes remains a very challenging problem, and presented evidence demonstrating that simple extensions of standard techniques cannot provide an acceptable solution. Subsequently, we proposed a new approach based on Bayesian simulation of classes and showed that preliminary results outperformed benchmark techniques [4]. Although these initial results looked promising, we did not consider the described preliminary algorithms final and definitive, and we do not believe that at this point we are able to provide an exact algorithmic solution to this complex problem. If we were able to do that, it would mean that we had already accomplished all the grant goals. The very essence of the proposed research is finding the answer to the defined problem, and the answer will remain unknown until after the work has been done. However, positive reviews and an acceptance of our work by KDD'09 conference judges, tell us that we are heading in the right direction. In the amended version of this application we propose a modified Bayesian approach based on Wishart priors (Section D.2.3). The algorithm creates new classes on the fly and evaluates maximum likelihood with the updated set of classes, gradually improving detection accuracy for future samples. We believe that this offers a substantial improvement over the previous method. Consequently, the preliminary results in Section C are updated to reflect our progress. Since the modified technique allows for classification with non-exhaustive and exhaustive sets using the same algorithm, we consolidated the previous specific aims 3 and 5 into one in the revised application.