SMICLR: Contrastive Learning on Multiple Molecular Representations for Semisupervised and Unsupervised Representation Learning.

The accurate prediction of suitable chiral stationary phases (CSPs) for resolving the enantiomers of a given compound poses a significant challenge in chiral chromatography. Previous attempts at developing machine learning models for structure-based CSP prediction have primarily relied on 1D SMILES strings [the simplified molecular-input line-entry system (SMILES) is a specification in the form of a line notation for describing the structure of chemical species using short ASCII strings] or 2D graphical representations of molecular structures and have met with only limited success. In this study, we apply the recently developed 3D molecular conformation representation learning algorithm, which uses rapid conformational analysis and point clouds of atom positions in the 3D space, enabling efficient chemical structure-based machine learning. By harnessing the power of the rapid 3D molecular representation learning and a data set comprising over 300,000 chromatographic enantioseparation records sourced from the literature, our models afford notable improvements for the chemical structure-based choice of appropriate CSP for enantioseparation, paving the way for more efficient and informed decision-making in the field of chiral chromatography.

Great and continuous efforts have been made to discover high-performance engineering plastics with specific properties to replace traditional engineering materials in many fields. The utilization of machine learning (ML) has brought more opportunities for the discovery of high-performing engineering plastics. However, hindered by either the relatively small database or a lack of accurate structure descriptors with clear physical and chemical meanings relating to polymer properties, the current ML studies show some flaws in the accuracy and efficiency in polymer development. Herein, we collected a dataset of 878 polyimides (PI), one of the best engineering plastics, with experimentally measured glass-transition temperature (Tg) values, and developed a rapid and accurate ML approach to design PI candidates with the desired Tg value. After the conversion from PI structures into "mechanically identifiable" SMILES (Simplified molecular input line entry system) language, the eight most critical descriptors were ultimately obtained by multiple analysis methods. The physiochemical meaning of the key descriptors was further analyzed carefully to translate the implicit "machine language" to chemical knowledge. The artificial neural network (ANN)-based model gave the most accurate results with a root-mean-square error of ∼11 K among the studied ML methods. More importantly, three potential PI candidates with desired Tg (DPIs) were designed according to the chemical insight of the key descriptors, which were then verified by experiments. The experimental and predicted Tg values of DPIs have an acceptable average deviation of ca. 3.66%. This accuracy has reached the level of the traditional molecular simulation, but the time consumption and hold-up computing resource are tremendously reduced. Furthermore, the current ML approach could offer a scalable and adaptable framework in future engineer plastics innovation.

Deep learning models for the quantitative structure-property relationship (QSPR) have traditionally encountered challenges related to limited interpretability and generalizability. In this study, we present the simplified molecular input line entry system (SMILES) token additivity (STA) model for accurately predicting fuel properties, which takes SMILES as input and employs stacked multihead self-attention encoders to extract molecular structural information. This model provides insights into the structure-property relationships by quantifying the contributions of individual tokens to target properties. Furthermore, since the STA model operates without handcrafted molecular fingerprints, it is capable of generalizing to a broad spectrum of structure-related properties. To validate the model's efficacy, seven critical fuel properties of standard enthalpy of formation (ΔfH°), entropy (S), isobaric heat capacity (Cp), cetane number (CN), boiling point (BP), melting point (MP), and flash point (FP) were tested. The 10-fold cross-validation demonstrated outstanding predictive accuracy, with mean absolute errors of 1.86 kcal/mol (ΔfH°), 0.62 kcal/mol/K (S), and 1.82 kcal/mol/K (Cp), alongside root-mean-square errors (RMSE) of 4.90 (CN), 11.27 °C (BP), 14.09 °C (MP), and 9.47 °C (FP). All properties achieved R2 values exceeding 0.95. The results demonstrate that it achieves predictive accuracy comparable to conventional machine learning models relying on sophisticated feature engineering while also identifying the effect of key tokens on ΔfH° and CN.

In real-world applications, clustering or classification can usually be improved by fusing information from different views. Therefore, unsupervised representation learning on multi-view data becomes a compelling topic in machine learning. In this paper, we propose a novel and flexible unsupervised multi-view representation learning model termed Collaborative Multi-View Information Bottleneck Networks (CMIB-Nets), which comprehensively explores the common latent structure and the view-specific intrinsic information, and discards the superfluous information in the data significantly improving the generalization capability of the model. Specifically, our proposed model relies on the information bottleneck principle to integrate the shared representation among different views and the view-specific representation of each view, prompting the multi-view complete representation and flexibly balancing the complementarity and consistency among multiple views. We conduct extensive experiments (including clustering analysis, robustness experiment, and ablation study) on real-world datasets, which empirically show promising generalization ability and robustness compared to state-of-the-arts.

Local discriminative representation is needed in many medical image analysis tasks such as identifying sub-types of lesion or segmenting detailed components of anatomical structures. However, the commonly applied supervised representation learning methods require a large amount of annotated data, and unsupervised discriminative representation learning distinguishes different images by learning a global feature, both of which are not suitable for localized medical image analysis tasks. In order to avoid the limitations of these two methods, we introduce local discrimination into unsupervised representation learning in this work. The model contains two branches: one is an embedding branch which learns an embedding function to disperse dissimilar pixels over a low-dimensional hypersphere; and the other is a clustering branch which learns a clustering function to classify similar pixels into the same cluster. These two branches are trained simultaneously in a mutually beneficial pattern, and the learnt local discriminative representations are able to well measure the similarity of local image regions. These representations can be transferred to enhance various downstream tasks. Meanwhile, they can also be applied to cluster anatomical structures from unlabeled medical images under the guidance of topological priors from simulation or other structures with similar topological characteristics. The effectiveness and usefulness of the proposed method are demonstrated by enhancing various downstream tasks and clustering anatomical structures in retinal images and chest X-ray images.KeywordsUnsupervised representation learningLocal discriminationTopological priors

We report a novel framework for achieving fragment-based molecular design using pixel convolutional neural network (PixelCNN) combined with the simplified molecular input line entry system (SMILES) as molecular representation. While a widely used recurrent neural network (RNN) assumes monotonically decaying correlations in strings, PixelCNN captures a periodicity among characters of SMILES. Thus, PixelCNN provides us with a novel solution for the analysis of chemical space by extracting the periodicity of molecular structures that will be buried in SMILES. Moreover, this characteristic enables us to generate molecules by combining several simple building blocks, such as a benzene ring and side-chain structures, which contributes to the effective exploration of chemical space by step-by-step searching for molecules from a target fragment. In conclusion, PixelCNN could be a powerful approach focusing on the periodicity of molecules to explore chemical space for the fragment-based molecular design.

High-performanceliquid chromatography (HPLC) remainsthe goldstandard for analyzing and purifying molecular components in solutions.However, developing HPLC methods is material- and time-consuming,so computer-aided shortcuts are highly desirable. In line with thedigitalization of process development and the growth of HPLC databases,we propose a data-driven methodology to predict molecule retentionfactors as a function of mobile phase composition without the needfor any new experiments, solely relying on molecular descriptors (MDs)obtained via simplified molecular input line entry system (SMILES)string representations of molecules. This new approach combines: (a)quantitative structure–property relationships (QSPR) usingMDs to predict solute-dependent parameters in (b) linear solvationenergy relationships (LSER) and (c) linear solvent strength (LSS)theory. We demonstrate the potential of this computational methodologyusing experimental data for retention factors of small molecules madeavailable by the research community for which the MDs were obtainedvia SMILES string representations determined by the structural formulasof the molecules. This method can be adopted directly to predict elutiontimes of molecular components; however, in combination with first-principle-basedmechanistic transport models, the method can also be employed to optimizeHPLC methods in-silico. Both options can reduce the experimental loadand accelerate HPLC method development significantly, lowering thetime and cost of the drug manufacturing cycle and reducing the timeto market. Given the growing number and quality of HPLC databases,the predictive power of this methodology will only increase in thecoming years.

Hydrogenated amorphous carbons (HACs) are complex, disordered forms of carbon that are of interest in various scientific fields, such as the study of air pollution from soot particles and astrochemistry. A new stochastic, structurally guided algorithm is presented for the large-scale generation of atomistic models of HACs. It consists of a two-step procedure: (i) the randomized generation of 2D structures using the SMILES (simplified molecular input line entry system) description, respecting user-predefined chemical constraints; (ii) the subsequent generation of 3D structures, making use of a stochastic sampling algorithm combined with local optimizations at the DFTB (density functional-based tight binding) level. The method was used to generate the ARMAGNHAC database (web site: https://armagnhac.laas.fr/) which provides structural (Cartesian coordinates, functional group ratios, Hill-Wheeler parameters, and aromaticity descriptors), energetic (HOMO-LUMO gap, ionization energy, and electronic affinity), and spectroscopic properties of 4366 HACs. Correlation plots between these descriptors can be generated on the web site, as well as IR spectra, possibly including their evolution as a function of a given property. Several illustrations are given, such as the dependence of the ionization potentials and electronic affinities on the size of the largest aromatic island of the HACs. The database (structures and properties) can be downloaded by users. This work paves the way for future studies aiming to derive relationships between structure, energetic, and spectral properties of HACs, as required to interpret spectral observations and experimental measurements.

The use of in silico toxicity prediction methods plays an important role in the selection of lead compounds and in ADMET studies since in vitro and in vivo methods are often limited by ethics, time, budget and other resources. In this context, we present our new web tool VenomPred, a user-friendly platform for evaluating the potential mutagenic, hepatotoxic, carcinogenic and estrogenic effects of small molecules. VenomPred platform employs several in-house Machine Learning (ML) models developed with datasets derived from VEGA QSAR, a software that includes a comprehensive collection of different toxicity models and has been used as a reference for building and evaluating our ML models. The results showed that our models achieved equal or better performance than those obtained with the reference models included in VEGA QSAR. In order to improve the predictive performance of our platform, we adopted a consensus approach combining the results of different ML models, which was able to predict chemical toxicity better than the single models. This improved method was thus implemented in the VenomPred platform, a freely accessible webserver that takes the SMILES (Simplified Molecular-Input Line-Entry System) strings of the compounds as input and sends the prediction results providing a probability score about their potential toxicity.

Abstract—Predicting and understanding the potential side effects of pharmaceutical drugs is a formidable challenge in the realm of medical science. The conventional process of assessing drug safety, primarily reliant on manual clinical testing [3] and post-market surveillance [4], [5], is not only arduous but also time-consuming, making it an impediment to rapid drug development and patient well-being. This paper introduces an innovative approach that harnesses the power of advanced machine learning techniques to address this challenge [6], [7], [8], [10], [9]. Our research delves into the intricate web of drug interactions within the human body and the complex factors contributing to adverse reactions. It is a realm where the exact mechanisms that trigger side effects often remain elusive, and the prevalence of rare and severe reactions complicates the task. Our approach encompasses the integration of diverse data sources, including drug characteristics, generic names, molecular structures (SMILES) [11], [12], and drug category [13]. Thereby providing a comprehensive understanding of the complex relationships between these factors and adverse reactions, unifying these disparate pieces of information, we aim to unveil hidden patterns and relationships that can significantly enhance the accuracy of drug side effect predictions. In this study, we employ the Random Forest classification model, known for its robustness and interpretability, to make predictions that not only incorporate the broad spectrum of drug-related factors [14] but also ensure the inclusion of rare and severe side effects in the assessment [15]. To address the challenge of handling non-numerical data, we employ methodologies to convert features such as SMILES structures, Drug generic names, and others into meaningful numerical descriptors, enabling their seamless integration into the predictive model. Our research does not only seek to enhance drug safety evaluation but also endeavors to bridge the chasm between the unpredictability of drug side effects and the need for a more efficient and informed drug development process. With the potential to revolutionize drug safety practices, this research has far-reaching implications in the domain of patient care and drug industry decision-making, promising a safer and more efficient future for the pharmaceutical research industry. Index Terms—SMILES (Simplified Molecular Input Line En- try System), Indication or Drug Target Diseases, Drug Side Effect, Random Forest, decision tree.

Atmospheric aerosols are complex mixtures of highly functionalized organic compounds, water, inorganic electrolytes, metals, and carbonaceous species. The surface properties of atmospheric aerosol particles can influence several of their chemical and physical impacts, including their hygroscopic growth, aerosol-cloud interactions, and heterogeneous chemical reactions. The effects of the various compounds within a particle on its surface tension depend in part on the pure-component surface tensions. For many of the myriad of organic compounds of interest, experimental pure-component surface tension data at tropospheric temperatures are lacking, thus, requiring the development and application of property estimation methods. In this work, a compiled database of experimental pure-component surface tension data, covering a wide range of organic compound classes and temperatures, is used to train four different types of machine learning models to predict the temperature-dependent pure-component surface tensions of atmospherically relevant organic compounds. The trained models process input information about the temperature and the molecular structure of an organic compound, initially in the form of a Simplified Molecular Input Line Entry System (SMILES) string, to enable predictions. Our quantitative model assessment shows that extreme gradient-boosted descent along with Molecular ACCess System (MACCS) key descriptors of molecular structure provided the best balance of derived input complexity and model performance, resulting in a root-mean-square error (RMSE) of ∼1 mJ m-2 in pure-component surface tension. Additionally, a simplified model based on molar mass, elemental ratios, and temperature as inputs was developed for use in applications for which molecular structure information is incomplete (RMSE of ∼2 mJ m-2). We demonstrate that including predicted pure-component surface tension values in thermodynamically rigorous bulk-surface partitioning calculations may substantially modify the critical supersaturations necessary for aerosol activation into cloud droplets.

A new paradigm combining quantum‐mechanical calculations with machine learning (ML) to rationally design compounds with specific properties from extremely large chemical space is emerging and developing at a rapid pace. In this context, appropriately describing molecules and efficiently extracting patterns from electronic‐structure calculations are the core challenges for the success of the quantum‐mechanics‐based ML approaches. Here, MolNet‐3D is introduced, a strong deep learning architecture capable of mapping from a flexible and universal 3D topography descriptor to quantum‐mechanical observables of molecules of arbitrary shape. The model can learn an invariant representation without the need for the transformation of atom coordinates into interatomic distances, thus preserving the intrinsic 3D topography information of molecules. The capabilities of MolNet‐3D are shown by accurately predicting the various density functional theory calculated properties for molecules, including energetic, electronic, and thermodynamic properties. Compared with the previously proposed ML methods in the MoleculeNet benchmarks, our model generally offers the best performance in those quantum‐mechanical tasks, elucidating the importance of intrinsic topography information in molecular representation learning. This work may provide new insight into the construction of molecular ML models from 3D topography recognition perspectives.

CURL: Image Classification using co-training and Unsupervised Representation Learning

Advances in small molecule representations and AI-driven drug research: bridging the gap between theory and application.

The conventional method for determining whether a drug is carcinogenic involves subjecting rodents to a 2-year bioassay, but this approach is both time-consuming and expensive, not to mention unethical. Consequently, machine learning techniques have emerged as a popular alternative. One such technique is ensemble learning, which aims to create more accurate and robust models. In this particular study, the LightGBM model was utilized to predict the carcinogenicity of chemicals using its fingerprints. Molecular fingerprints were generated from the Simplified Molecular-Input Line-Entry System (SMILES) of 1003 chemicals from the Carcinogenic Potency Database (CPDB) dataset. The performance of the LightGBM model was found to be superior to other machine learning models reported in previous research. To further validate the model, it was tested on a database related to humans from the International Agency for Research on Cancer (IARC), as well as on chemicals that were withdrawn from the market between 1950 and 2014. The results showed that the LightGBM model was effective in identifying carcinogenic chemicals, suggesting that this approach could potentially replace traditional methods of carcinogenicity testing in the future.