Florian Nigsch

Publications of Florian Nigsch

• Rethinking molecular similarity: comparing compounds based on biological activity.

  Authors: Paula M Petrone, Benjamin Simms, Florian Nigsch, Eugen Lounkine, Peter Kutchukian, Allen Cornett, Zhan Deng, John W Davies, Jeremy Jenkins, Meir Glick

  ACS chemical biology. 05/2012;

  Since the advent of High Throughput Screening (HTS), there has been an urgent need for methods that facilitate the interrogation of large-scale chemical biology data to build a mode of action (MoA)Since the advent of High Throughput Screening (HTS), there has been an urgent need for methods that facilitate the interrogation of large-scale chemical biology data to build a mode of action (MoA) hypothesis. This can be done either prior to the HTS by subset design of compounds with known MoA or post HTS by data annotation and mining. To enable this process, we develop a tool that compares compounds solely based on their bioactivity -the chemical biological descriptors "High-Throughput Screening Fingerprints" (HTS-FP). In the current embodiment, data are aggregated from 195 biochemical and cell-based assays developed at Novartis and can be used to identify bioactivity relationships among the in-house collection comprising ~1.5 million compounds. We demonstrate the value of the HTS-FP for virtual screening and in particular scaffold hopping. HTS-FP outperforms state of the art methods in several aspects, retrieving bioactive compounds with remarkable chemical dissimilarity to a probe structure. We also apply HTS-FP for the design of screening subsets in HTS. Using retrospective data, we show that a biodiverse selection of plates performs significantly better than a chemically diverse selection of plates, both in terms of number of hits and diversity of chemotypes retrieved. This is also true in the case of hit expansion predictions using HTS-FP similarity. Sets of compounds clustered with HTS-FP are biologically meaningful, in a sense that these clusters enrich for genes and gene ontology (GO) terms, showing that compounds that are bioactively similar also tend to target proteins that operate together in the cell. HTS-FP are not only valuable because of their predictive power, but mainly because they relate compounds solely based on bioactivity, harnessing the accumulated knowledge of a high-throughput screening facility towards the understanding of how compounds interact with the proteome.

• Determination of minimal transcriptional signatures of compounds for target prediction.

  Authors: Florian Nigsch, Janna Hutz, Ben Cornett, Douglas W Selinger, Gregory McAllister, Somnath Bandyopadhyay, Joseph Loureiro, Jeremy L Jenkins

  EURASIP journal on bioinformatics & systems biology. 05/2012; 2012(1):2.

  ABSTRACT: The identification of molecular target and mechanism of action of compounds is a key hurdle in drug discovery. Multiplexed techniques for bead-based expression profiling allow theABSTRACT: The identification of molecular target and mechanism of action of compounds is a key hurdle in drug discovery. Multiplexed techniques for bead-based expression profiling allow the measurement of transcriptional signatures of compound-treated cells in high-throughput mode. Such profiles can be used to gain insight into compounds' mode of action and the protein targets they are modulating. Through the proxy of target prediction from such gene signatures we explored important aspects of the use of transcriptional profiles to capture biological variability of perturbed cellular assays. We found that signatures derived from expression data and signatures derived from biological interaction networks performed equally well, and we showed that gene signatures can be optimised using a genetic algorithm. Gene signatures of approximately 128 genes seemed to be most generic, capturing a maximum of the perturbation inflicted on cells through compound treatment. Moreover, we found evidence for oxidative phosphorylation to be one of the most general ways to capture compound perturbation.

• Predicting the mechanism of phospholipidosis.

  Authors: Robert Lowe, Hamse Y Mussa, Florian Nigsch, Robert C Glen, John Bo Mitchell

  Journal of cheminformatics. 01/2012; 4:2.

  ABSTRACT: The mechanism of phospholipidosis is still not well understood. Numerous different mechanisms have been proposed, varying from direct inhibition of the breakdown of phospholipids to theABSTRACT: The mechanism of phospholipidosis is still not well understood. Numerous different mechanisms have been proposed, varying from direct inhibition of the breakdown of phospholipids to the binding of a drug compound to the phospholipid, preventing breakdown. We have used a probabilistic method, the Parzen-Rosenblatt Window approach, to build a model from the ChEMBL dataset which can predict from a compound's structure both its primary pharmaceutical target and other targets with which it forms off-target, usually weaker, interactions. Using a small dataset of 182 phospholipidosis-inducing and non-inducing compounds, we predict their off-target activity against targets which could relate to phospholipidosis as a side-effect of a drug. We link these targets to specific mechanisms of inducing this lysosomal build-up of phospholipids in cells. Thus, we show that the induction of phospholipidosis is likely to occur by separate mechanisms when triggered by different cationic amphiphilic drugs. We find that both inhibition of phospholipase activity and enhanced cholesterol biosynthesis are likely to be important mechanisms. Furthermore, we provide evidence suggesting four specific protein targets. Sphingomyelin phosphodiesterase, phospholipase A2 and lysosomal phospholipase A1 are shown to be likely targets for the induction of phospholipidosis by inhibition of phospholipase activity, while lanosterol synthase is predicted to be associated with phospholipidosis being induced by enhanced cholesterol biosynthesis. This analysis provides the impetus for further experimental tests of these hypotheses.

• Activity-aware clustering of high throughput screening data and elucidation of orthogonal structure-activity relationships.

  Authors: Eugen Lounkine, Florian Nigsch, Jeremy L Jenkins, Meir Glick

  Journal of chemical information and modeling. 11/2011; 51(12):3158-68.

  From a medicinal chemistry point of view, one of the primary goals of high throughput screening (HTS) hit list assessment is the identification of chemotypes with an informative structure-activityFrom a medicinal chemistry point of view, one of the primary goals of high throughput screening (HTS) hit list assessment is the identification of chemotypes with an informative structure-activity relationship (SAR). Such chemotypes may enable optimization of the primary potency, as well as selectivity and phamacokinetic properties. A common way to prioritize them is molecular clustering of the hits. Typical clustering techniques, however, rely on a general notion of chemical similarity or standard rules of scaffold decomposition and are thus insensitive to molecular features that are enriched in biologically active compounds. This hinders SAR analysis, because compounds sharing the same pharmacophore might not end up in the same cluster and thus are not directly compared to each other by the medicinal chemist. Similarly, common chemotypes that are not related to activity may contaminate clusters, distracting from important chemical motifs. We combined molecular similarity and Bayesian models and introduce (I) a robust, activity-aware clustering approach and (II) a feature mapping method for the elucidation of distinct SAR determinants in polypharmacologic compounds. We evaluated the method on 462 dose-response assays from the Pubchem Bioassay repository. Activity-aware clustering grouped compounds sharing molecular cores that were specific for the target or pathway at hand, rather than grouping inactive scaffolds commonly found in compound series. Many of these core structures we also found in literature that discussed SARs of the respective targets. A numerical comparison of cores allowed for identification of the structural prerequisites for polypharmacology, i.e., distinct bioactive regions within a single compound, and pointed toward selectivity-conferring medchem strategies. The method presented here is generally applicable to any type of activity data and may help bridge the gap between hit list assessment and designing a medchem strategy.

• Computational methods for early predictive safety assessment from biological and chemical data.

  Authors: Florian Nigsch, Eugen Lounkine, Patrick McCarren, Ben Cornett, Meir Glick, Kamal Azzaoui, Laszlo Urban, Philippe Marc, Arne Müller, Florian Hahne, David J Heard, Jeremy L Jenkins

  Expert opinion on drug metabolism & toxicology. 11/2011; 7(12):1497-511.

  INTRODUCTION: The goal of early predictive safety assessment (PSA) is to keep compounds with detectable liabilities from progressing further in the pipeline. Such compounds jeopardize the core ofINTRODUCTION: The goal of early predictive safety assessment (PSA) is to keep compounds with detectable liabilities from progressing further in the pipeline. Such compounds jeopardize the core of pharmaceutical research and development and limit the timely delivery of innovative therapeutics to the patient. Computational methods are increasingly used to help understand observed data, generate new testable hypotheses of relevance to safety pharmacology, and supplement and replace costly and time-consuming experimental procedures. AREAS COVERED: The authors survey methods operating on different scales of both physical extension and complexity. After discussing methods used to predict liabilities associated with structures of individual compounds, the article reviews the use of adverse event data and safety profiling panels. Finally, the authors examine the complexities of toxicology data from animal experiments and how these data can be mined. EXPERT OPINION: A significant obstacle for data-driven safety assessment is the absence of integrated data sets due to a lack of sharing of data and of using standard ontologies for data relevant to safety assessment. Informed decisions to derive focused sets of compounds can help to avoid compound liabilities in screening campaigns, and improved hit assessment of such campaigns can benefit the early termination of undesirable compounds.

• Online chemical modeling environment (OCHEM): web platform for data storage, model development and publishing of chemical information.

  Authors: Iurii Sushko, Sergii Novotarskyi, Robert Körner, Anil Kumar Pandey, Matthias Rupp, Wolfram Teetz, Stefan Brandmaier, Ahmed Abdelaziz, Volodymyr V Prokopenko, Vsevolod Y Tanchuk [......] Dmitriy Chekmarev, Artem Cherkasov, Joao Aires-de-Sousa, Qing-You Zhang, Andreas Bender, Florian Nigsch, Luc Patiny, Antony Williams, Valery Tkachenko, Igor V Tetko

  Journal of computer-aided molecular design. 06/2011; 25(6):533-54.

  The Online Chemical Modeling Environment is a web-based platform that aims to automate and simplify the typical steps required for QSAR modeling. The platform consists of two major subsystems: theThe Online Chemical Modeling Environment is a web-based platform that aims to automate and simplify the typical steps required for QSAR modeling. The platform consists of two major subsystems: the database of experimental measurements and the modeling framework. A user-contributed database contains a set of tools for easy input, search and modification of thousands of records. The OCHEM database is based on the wiki principle and focuses primarily on the quality and verifiability of the data. The database is tightly integrated with the modeling framework, which supports all the steps required to create a predictive model: data search, calculation and selection of a vast variety of molecular descriptors, application of machine learning methods, validation, analysis of the model and assessment of the applicability domain. As compared to other similar systems, OCHEM is not intended to re-implement the existing tools or models but rather to invite the original authors to contribute their results, make them publicly available, share them with other users and to become members of the growing research community. Our intention is to make OCHEM a widely used platform to perform the QSPR/QSAR studies online and share it with other users on the Web. The ultimate goal of OCHEM is collecting all possible chemoinformatics tools within one simple, reliable and user-friendly resource. The OCHEM is free for web users and it is available online at http://www.ochem.eu.

  Download

• Classifying large chemical data sets: using a regularized potential function method.

  Authors: Hamse Y Mussa, Lezan Hawizy, Florian Nigsch, Robert C Glen

  Journal of chemical information and modeling. 01/2011; 51(1):4-14.

  In recent years classifiers generated with kernel-based methods, such as support vector machines (SVM), Gaussian processes (GP), regularization networks (RN), and binary kernel discrimination (BKD)In recent years classifiers generated with kernel-based methods, such as support vector machines (SVM), Gaussian processes (GP), regularization networks (RN), and binary kernel discrimination (BKD) have been very popular in chemoinformatics data analysis. Aizerman et al. were the first to introduce the notion of employing kernel-based classifiers in the area of pattern recognition. Their original scheme, which they termed the potential function method (PFM), can basically be viewed as a kernel-based perceptron procedure and arguably subsumes the modern kernel-based algorithms. PFM can be computationally much cheaper than modern kernel-based classifiers; furthermore, PFM is far simpler conceptually and easier to implement than the SVM, GP, and RN algorithms. Unfortunately, unlike, e.g., SVM, GP, and RN, PFM is not endowed with both theoretical guarantees and practical strategies to safeguard it against generating overfitting classifiers. This is, in our opinion, the reason why this simple and elegant method has not been taken up in chemoinformatics. In this paper we empirically address this drawback: while maintaining its simplicity, we demonstrate that PFM combined with a simple regularization scheme may yield binary classifiers that can be, in practice, as efficient as classifiers obtained by employing state-of-the-art kernel-based methods. Using a realistic classification example, the augmented PFM was used to generate binary classifiers. Using a large chemical data set, the generalization ability of PFM classifiers were then compared with the prediction power of Laplacian-modified naive Bayesian (LmNB), Winnow (WN), and SVM classifiers.

• A lead discovery strategy driven by a comprehensive analysis of proteases in the peptide substrate space.

  Authors: Sai Chetan K Sukuru, Florian Nigsch, Jean Quancard, Martin Renatus, Rajiv Chopra, Natasja Brooijmans, Dmitri Mikhailov, Zhan Deng, Allen Cornett, Jeremy L Jenkins, Ulrich Hommel, John W Davies, Meir Glick

  Protein science : a publication of the Protein Society. 11/2010; 19(11):2096-109.

  We present here a comprehensive analysis of proteases in the peptide substrate space and demonstrate its applicability for lead discovery. Aligned octapeptide substrates of 498 proteases taken fromWe present here a comprehensive analysis of proteases in the peptide substrate space and demonstrate its applicability for lead discovery. Aligned octapeptide substrates of 498 proteases taken from the MEROPS peptidase database were used for the in silico analysis. A multiple-category naïve Bayes model, trained on the two-dimensional chemical features of the substrates, was able to classify the substrates of 365 (73%) proteases and elucidate statistically significant chemical features for each of their specific substrate positions. The positional awareness of the method allows us to identify the most similar substrate positions between proteases. Our analysis reveals that proteases from different families, based on the traditional classification (aspartic, cysteine, serine, and metallo), could have substrates that differ at the cleavage site (P1-P1') but are similar away from it. Caspase-3 (cysteine protease) and granzyme B (serine protease) are previously known examples of cross-family neighbors identified by this method. To assess whether peptide substrate similarity between unrelated proteases could reliably translate into the discovery of low molecular weight synthetic inhibitors, a lead discovery strategy was tested on two other cross-family neighbors--namely cathepsin L2 and matrix metallo proteinase 9, and calpain 1 and pepsin A. For both these pairs, a naïve Bayes classifier model trained on inhibitors of one protease could successfully enrich those of its neighbor from a different family and vice versa, indicating that this approach could be prospectively applied to lead discovery for a novel protease target with no known synthetic inhibitors.

• Computational toxicology: an overview of the sources of data and of modelling methods.

  Authors: Florian Nigsch, N J Maximilan Macaluso, John B O Mitchell, Donatas Zmuidinavicius

  Expert opinion on drug metabolism & toxicology. 02/2009; 5(1):1-14.

  BACKGROUND: Toxicology has the goal of ensuring the safety of humans, animals and the environment. Computational toxicology is an area of active development and great potential. There are tangibleBACKGROUND: Toxicology has the goal of ensuring the safety of humans, animals and the environment. Computational toxicology is an area of active development and great potential. There are tangible reasons for the emerging interest in this discipline from academia, industry, regulatory bodies and governments. RESULTS: Pharmaceuticals, personal health care products, nutritional ingredients and products of the chemical industries are all potential hazards and need to be assessed. Toxicological tests for these products are costly, frequently use laboratory animals and are time-consuming. This delays end-user access to improved products or, conversely, the timely withdrawal of dangerous substances from the market. The aim of computational toxicology is to accelerate the assessment of potentially dangerous substances through in silico models. CONCLUSIONS: In this review, we provide an overview of the development of models for computational toxicology. Addressing the significant divide between the experimental and computational worlds-believed to be a prime hindrance to computational toxicology-we briefly consider the fundamental issue of toxicological data and the assays they stem from. Different kinds of models that can be built using such data are presented: computational filters, models for specific toxicological endpoints and tools for the generation of testable hypotheses.

• Ligand-Target Prediction Using Winnow and Naive Bayesian Algorithms and the Implications of Overall Performance Statistics.

  Authors: Florian Nigsch, Andreas Bender, Jeremy L Jenkins, John B O Mitchell

  Journal of chemical information and modeling. 01/2009;

  We compared two algorithms for ligand-target prediction, namely, the Laplacian-modified Bayesian classifier and the Winnow algorithm. A dataset derived from the WOMBAT database, spanning 20We compared two algorithms for ligand-target prediction, namely, the Laplacian-modified Bayesian classifier and the Winnow algorithm. A dataset derived from the WOMBAT database, spanning 20 pharmaceutically relevant activity classes with 13 000 compounds, was used for performance assessment in 24 different experiments, each of which was assessed using a 15-fold Monte Carlo cross-validation. Compounds were described by different circular fingerprints, ECFP_4 and MOLPRINT 2D. A detailed analysis of the resulting approximately 2.4 million predictions led to very similar measures for overall accuracy for both classifiers, whereas we observed significant differences for individual activity classes. Moreover, we analyzed our data with respect to the numbers of compounds which are exclusively retrieved by either of the algorithmsbut never by the otheror by neither of them. This provided detailed information that can never be obtained by considering the overall performance statistics alone.

• Toxicological relationships between proteins obtained from protein target predictions of large toxicity databases.

  Authors: Florian Nigsch, John B O Mitchell

  Toxicology and applied pharmacology. 06/2008;

  The combination of models for protein target prediction with large databases containing toxicological information for individual molecules allows the derivation of "toxiclogical" profiles, i.e., toThe combination of models for protein target prediction with large databases containing toxicological information for individual molecules allows the derivation of "toxiclogical" profiles, i.e., to what extent are molecules of known toxicity predicted to interact with a set of protein targets. To predict protein targets of drug-like and toxic molecules, we built a computational multiclass model using the Winnow algorithm based on a dataset of protein targets derived from the MDL Drug Data Report. A 15-fold Monte Carlo cross-validation using 50% of each class for training, and the remaining 50% for testing, provided an assessment of the accuracy of that model. We retained the 3 top-ranking predictions and found that in 82% of all cases the correct target was predicted within these three predictions. The first prediction was the correct one in almost 70% of cases. A model built on the whole protein target dataset was then used to predict the protein targets for 150000 molecules from the MDL Toxicity Database. We analysed the frequency of the predictions across the panel of protein targets for experimentally determined toxicity classes of all molecules. This allowed us to identify clusters of proteins related by their toxicological profiles, as well as toxicities that are related. Literature-based evidence is provided for some specific clusters to show the relevance of the relationships identified.

• How to winnow actives from inactives: introducing molecular orthogonal sparse bigrams (MOSBs) and multiclass Winnow.

  Authors: Florian Nigsch, John B O Mitchell

  Journal of chemical information and modeling. 03/2008; 48(2):306-18.

  In the present paper we combine the Winnow algorithm and an advanced scheme for feature generation into a tool for multiclass classification. The Winnow algorithm, specifically designed in the lateIn the present paper we combine the Winnow algorithm and an advanced scheme for feature generation into a tool for multiclass classification. The Winnow algorithm, specifically designed in the late 1980s to work well with high-dimensional data, by design ignores most of the irrelevant features for the scoring of each single training/test case. To augment the pool of available molecular features we use the Winnow algorithm in conjunction with a process that creates additional features from a set of given ones. We adapt a technique formerly employed in text classification termed "orthogonal sparse bigrams" and extend the use of that method to the domain of cheminformatics. Using circular molecular fingerprints as initial features, we create "molecular orthogonal sparse bigrams" (MOSBs) and report their successful application to the task of classification of bioactive molecules. Additionally, we introduce a memory-efficient way of bagging individual classifiers, avoiding the need to hold the complete training data set in memory. To compare the performance of our method with published results, we use the Hert data set of 8293 active molecules in 11 classes. We compare our method to Random Forest and find that our method not only is comparable or better in classification accuracy (up to 50% higher in MCC [Matthews correlation coefficient], 98% higher in fraction of correct predictions) but also is quicker to train (by a factor between 2 and 18, depending on the feature generation), more memory efficient, and able to cope more easily with large data sets when we seeded the actives into a pool of 94290 inactive molecules. It is shown that this method can be used with different fingerprints.

• Why are some properties more difficult to predict than others? A study of QSPR models of solubility, melting point, and Log P.

  Authors: Laura D Hughes, David S Palmer, Florian Nigsch, John B O Mitchell

  Journal of chemical information and modeling. 02/2008; 48(1):220-32.

  This paper attempts to elucidate differences in QSPR models of aqueous solubility (Log S), melting point (Tm), and octanol-water partition coefficient (Log P), three properties of pharmaceuticalThis paper attempts to elucidate differences in QSPR models of aqueous solubility (Log S), melting point (Tm), and octanol-water partition coefficient (Log P), three properties of pharmaceutical interest. For all three properties, Support Vector Machine models using 2D and 3D descriptors calculated in the Molecular Operating Environment were the best models. Octanol-water partition coefficient was the easiest property to predict, as indicated by the RMSE of the external test set and the coefficient of determination (RMSE = 0.73, r2 = 0.87). Melting point prediction, on the other hand, was the most difficult (RMSE = 52.8 degrees C, r2 = 0.46), and Log S statistics were intermediate between melting point and Log P prediction (RMSE = 0.900, r2 = 0.79). The data imply that for all three properties the lack of measured values at the extremes is a significant source of error. This source, however, does not entirely explain the poor melting point prediction, and we suggest that deficiencies in descriptors used in melting point prediction contribute significantly to the prediction errors.

• In vitro models for processes involved in intestinal absorption.

  Authors: Florian Nigsch, Werner Klaffke, Silvia Miret

  Expert opinion on drug metabolism & toxicology. 09/2007; 3(4):545-56.

  The abundance of different techniques and protocols available reflects the need for reliable in vitro methods to assess intestinal absorption of potentially bioactive compounds. PhysicochemicalThe abundance of different techniques and protocols available reflects the need for reliable in vitro methods to assess intestinal absorption of potentially bioactive compounds. Physicochemical assays try to pinpoint the molecular properties contributing to the absorption process. The end points of biologically based methods, such as cell cultures and excised tissues, account for all processes undergone by a molecule that traverses a 'living' biological membrane, a cell or tissue. On top of fundamental physical processes (e.g., solubility, diffusion) such biological methods incorporate physiological responses such as active transport and metabolism. In this review, an account of in vitro methods for the assessment of molecular properties (lipophilicity, solubility, permeability) influencing intestinal absorption is given. Their advantages and limitations and the possibilities offered by this area of research are also evaluated. The combination of results from both classes of assays (physicochemical and biological) and integration with computational models will guide future developments in this field. Finally, possible future developments including stem cell research and multiple-end point assays are discussed.

• Toxicological relationships between proteins obtained from protein target predictions of large toxicity databases

  Authors: Florian Nigsch, John B.O. Mitchell

  Toxicology and Applied Pharmacology.

  The combination of models for protein target prediction with large databases containing toxicological information for individual molecules allows the derivation of “toxiclogical” profiles, i.e., toThe combination of models for protein target prediction with large databases containing toxicological information for individual molecules allows the derivation of “toxiclogical” profiles, i.e., to what extent are molecules of known toxicity predicted to interact with a set of protein targets. To predict protein targets of drug-like and toxic molecules, we built a computational multiclass model using the Winnow algorithm based on a dataset of protein targets derived from the MDL Drug Data Report. A 15-fold Monte Carlo cross-validation using 50% of each class for training, and the remaining 50% for testing, provided an assessment of the accuracy of that model. We retained the 3 top-ranking predictions and found that in 82% of all cases the correct target was predicted within these three predictions. The first prediction was the correct one in almost 70% of cases. A model built on the whole protein target dataset was then used to predict the protein targets for 150 000 molecules from the MDL Toxicity Database. We analysed the frequency of the predictions across the panel of protein targets for experimentally determined toxicity classes of all molecules. This allowed us to identify clusters of proteins related by their toxicological profiles, as well as toxicities that are related. Literature-based evidence is provided for some specific clusters to show the relevance of the relationships identified.

• Melting point prediction employing k-nearest neighbor algorithms and genetic parameter optimization.

  Authors: Florian Nigsch, Andreas Bender, Bernd van Buuren, Jos Tissen, Eduard Nigsch, John B O Mitchell

  Journal of chemical information and modeling. 46(6):2412-22.

  We have applied the k-nearest neighbor (kNN) modeling technique to the prediction of melting points. A data set of 4119 diverse organic molecules (data set 1) and an additional set of 277 drugs (dataWe have applied the k-nearest neighbor (kNN) modeling technique to the prediction of melting points. A data set of 4119 diverse organic molecules (data set 1) and an additional set of 277 drugs (data set 2) were used to compare performance in different regions of chemical space, and we investigated the influence of the number of nearest neighbors using different types of molecular descriptors. To compute the prediction on the basis of the melting temperatures of the nearest neighbors, we used four different methods (arithmetic and geometric average, inverse distance weighting, and exponential weighting), of which the exponential weighting scheme yielded the best results. We assessed our model via a 25-fold Monte Carlo cross-validation (with approximately 30% of the total data as a test set) and optimized it using a genetic algorithm. Predictions for drugs based on drugs (separate training and test sets each taken from data set 2) were found to be considerably better [root-mean-squared error (RMSE)=46.3 degrees C, r2=0.30] than those based on nondrugs (prediction of data set 2 based on the training set from data set 1, RMSE=50.3 degrees C, r2=0.20). The optimized model yields an average RMSE as low as 46.2 degrees C (r2=0.49) for data set 1, and an average RMSE of 42.2 degrees C (r2=0.42) for data set 2. It is shown that the kNN method inherently introduces a systematic error in melting point prediction. Much of the remaining error can be attributed to the lack of information about interactions in the liquid state, which are not well-captured by molecular descriptors.