Abstract

Chemical reaction prediction, encompassing forward synthesis and retrosynthesis, stands as a fundamental challenge in organic synthesis. A widely adopted computational approach frames synthesis prediction as a sequence-to-sequence translation task, using the common SMILES representation for molecules. Current evaluation of machine learning methods for retrosynthesis assume perfect training data, overlooking imperfections in reaction equations in popular datasets, such as missing reactants, products, other physical and practical constraints such as temperature and cost, primarily driven by a focus on the target molecule. This limitation leads to an incomplete representation of viable synthetic routes, especially when multiple sets of reactants can yield a given desired product. In response to these shortcomings, this study examines the prevailing evaluation methods and introduces comprehensive metrics designed to address imperfections in the dataset. Our novel metrics not only assess absolute accuracy by comparing predicted outputs with ground truth but also introduce a nuanced evaluation approach. We provide scores for partial correctness and compute adjusted accuracy through graph matching, acknowledging the inherent complexities of retrosynthetic pathways. Additionally, we explore the impact of small molecular augmentations while preserving chemical properties and employ similarity matching to enhance the assessment of prediction quality. We introduce SynFormer, a sequence-to-sequence model tailored for SMILES representation. It incorporates architectural enhancements to the original transformer, effectively tackling the challenges of chemical reaction prediction. SynFormer achieves a top-1 accuracy of 53.2% on the USPTO-50k dataset, demonstrating an improvement over previous state-of-the-art language models while being more efficient and eliminating the need for pre-training.

Abstract

RNA molecules play a significant role in many biological pathways and have diverse functional roles, which is a result of their structural flexibility to fold into diverse conformations. This structural flexibility makes it challenging to obtain the structures of RNAs experimentally. Deep learning can be used to predict the secondary structures of RNA and other properties such as the backbone torsion angles, to be used as restraints for the computational optimization of the tertiary structures of RNA. TorRNA is a transformer encoder-decoder model, that takes an input RNA sequence and predicts the (pseudo)torsion angles of each nucleotide with a pre-trained RNA-FM model as the encoder. TorRNA is able to achieve a performance boost of 2% − 16% over the previous (pseudo)torsion angle prediction method for RNAs. We also demonstrate that TorRNA can used as a tool for model quality assessment of candidate RNA structures

Abstract

Computing binding affinities is of great importance in drug discovery pipeline and its prediction using advanced machine learning methods still remains a major challenge as the existing datasets and models do not consider the dynamic features of protein-ligand interactions. To this end, we have developed PLAS-20k dataset, an extension of previously developed PLAS-5k, with 97,500 independent simulations on a total of 19,500 different protein-ligand complexes. Our results show good correlation with the available experimental values, performing better than docking scores. This holds true even for a subset of ligands that follows Lipinski’s rule, and for diverse clusters of complex structures, thereby highlighting the importance of PLAS-20k dataset in developing new ML models. Along with this, our dataset is also beneficial in classifying strong and weak binders compared to docking. Further, OnionNet model has been retrained on PLAS-20k dataset and is provided as a baseline for the prediction of binding affinities. We believe that large-scale MD-based datasets along with trajectories will form new synergy, paving the way for accelerating drug discovery.

Abstract

Molecular spectroscopy studies the interaction of molecules with electromagnetic radiation, and interpreting the resultant spectra is invaluable for deducing the molecular structures. However, predicting the molecular structure from spectroscopic data is a strenuous task that requires highly specific domain knowledge. DeepSPInN is a deep reinforcement learning method that predicts the molecular structure when given Infrared and 13C Nuclear magnetic resonance spectra by formulating the molecular structure prediction problem as a Markov decision process (MDP) and employs Monte-Carlo tree search to explore and choose the actions in the formulated MDP. On the QM9 dataset, DeepSPInN is able to predict the correct molecular structure for 91.5% of the input spectra in an average time of 77 seconds for molecules with less than 10 heavy atoms. This study is the first of its kind that uses only infrared and 13C nuclear magnetic resonance spectra for molecular structure prediction without referring to any pre-existing spectral databases or molecular fragment knowledge bases, and is a leap forward in automated molecular spectral analysis.

Abstract

With Cardiovascular Diseases on the rise around the world, Electrocardiograms (ECGs) play a crucial role in their diagnosis owing to their non-invasive nature and simplicity. Medical professionals typically use 12-lead ECGs for medical analysis but gathering 12-lead ECG data is an arduous task outside clinical setting. Modern wearables can collect an ECG with fewer leads than the standard 12 leads. However, medical professionals and conventional ECG analysis software find this reduced lead set data challenging to interpret. By using the reduced lead set data to create standard 12-lead ECG data, ECG reconstruction can solve this issue. This paper proposes a novel single-lead to multi-lead ECG reconstruction solution using a modified Attention U-net framework. Using only the lead II of ECG, our model is able to reproduce the other 11 leads of conventional 12-lead ECG with a Pearson correlation, Mean square error and R-squared value of 0.805, 0.0122 and 0.639, respectively. Further, a single combined model is used to reconstruct all 11 leads simultaneously, improving performance and simultaneously reducing the computational resources needed for training compared to current literature in the field. In comparison to previous works, which only reconstruct small ECG segments, our model is trained to reconstruct longer 10-second ECG signals. We demonstrate our model’s ability for real-life utilisation using a cardiovascular disease classification task. A deep learning model was trained for multi-disease classification on actual 12-lead ECG data and was tested on both original and reconstructed 12-lead ECG signals. The classification accuracies for the original and reconstructed signals were comparable, portraying that our reconstruction model can preserve diagnostically relevant artefacts in its reconstructed signals. This work provides a new promising solution in the field of single-lead ECG reconstruction, taking us a step closer to bridging the divide between reduced lead set data and existing 12-lead ECG end users like clinicians and automatic ECG classifiers.

Abstract

In this work, we describe the creation of our machine- learning-based solution for coma prognosis after cardiac arrest using longitudinal EEG and ECG recordings for the ”Predicting Neurological Recovery from Coma After Car- diac Arrest: The George B. Moody PhysioNet Challenge 2023”. Our team, “ComaToast”, had its best submission ranked 28 out of 36 teams selected worldwide, with a chal- lenge score of 0.381 on the official leaderboard for the hid- den test set. We use a combination of age and signal fea- tures from EEG and ECG recordings. Frequency domain features, specifically mean power spectral density from 4 different bands of frequencies (Delta, Theta, Alpha and Beta) and mean Burst Suppression Ratio, were extracted from pre-processed EEG recordings from the first and last available recording for a given patient. Features like mean and standard deviations were extracted along channels for ECG recordings. After imputing missing values, these fea- tures are fed to an XGBoost classifier for the final binary classification of the outcome prediction task. The features are fed to a random forest regressor to predict the CPC outcome for every patient. A solution like ours, which uses a simple model and training technique, may be more viable than deep-learning solutions in general use cases. In our final model, our approach achieved a 5-fold cross- validation score of 0.34 on the public train set.

Abstract

The discovery of potential therapeutic agents for life‑threatening diseases has become a significant problem. There is a requirement for fast and accurate methods to identify drug‑like molecules that can be used as potential candidates for novel targets. Existing techniques like high‑throughput screening and virtual screening are time‑consuming and inefficient. Traditional molecule generation pipelines are more efficient than virtual screening but use time‑consuming docking software. Such docking functions can be emulated using Machine Learning models with comparable accuracy and faster execution times. However, we find that when pre‑trained machine learning models are employed in generative pipelines as oracles, they suffer from model degradation in areas where data is scarce. In this study, we propose an active learning‑based model that can be added as a supplement to enhanced molecule generation architectures. The proposed method uses uncertainty sampling on the molecules created by the generator model and dynamically learns as the generator samples molecules from different regions of the chemical space. The proposed framework can generate molecules with high binding affinity with ∼a 70% improvement in runtime compared to the baseline model by labeling only ∼30% of molecules compared to the baseline oracle.

Abstract

In this paper, we propose MolOpt, the first attempt of its kind to use Multi-Agent Reinforcement Learning (MARL) for autonomous molecular geometry optimization (MGO). Typically MGO algorithms are hand-designed, but MolOpt uses MARL to learn a learned optimizer (policy) that can perform MGO without depending on other hand-designed optimizers. We cast MGO as a MARL problem, where each agent corresponds to a single atom in the molecule. MolOpt performs MGO by minimizing the forces on each atom in the molecule. Our experiments demonstrate the generalizing ability of MolOpt for MGO of Propane, Pentane, Heptane, Hexane, and Octane when trained on Ethane, Butane, and Isobutane. In terms of performance, MolOpt outperforms the MDMin optimizer and demonstrates similar performance to the FIRE optimizer. However, it does not surpass the BFGS optimizer. The results demonstrate that MolOpt has the potential to introduce innovative advancements in MGO by providing a novel approach using reinforcement learning (RL), which may open up new research directions for MGO. Overall, this work serves as a proof-of-concept for the potential of MARL in MGO.

Abstract

The discovery of potential therapeutic agents for life-threatening diseases has become a significant problem. There is a requirement for fast and accurate methods to identify drug-like molecules that can be used as potential candidates for novel targets. Existing techniques like high-throughput screening and virtual screening are time-consuming and inefficient. Traditional molecule generation pipelines are more efficient than virtual screening but use time-consuming docking software. Such docking functions can be emulated using Machine Learning models with comparable accuracy and faster execution times. However, we find that when pre-trained machine learning models are employed in generative pipelines as oracles, they suffer from model degradation in areas where data is scarce. In this study, we propose an active learning-based model that can be added as a supplement to enhanced molecule generation architectures. The proposed method uses uncertainty sampling on the molecules created by the generator model and dynamically learns as the generator samples molecules from different regions of the chemical space. The proposed framework can generate molecules with high binding affinity with ∼ a 70% improvement in runtime compared to the baseline model by labeling only ∼ 30% of molecules compared to the baseline oracle.

Abstract

In drug discovery applications, high throughput virtual screening exercises are rou- tinely performed to determine an initial set of candidate molecules referred to as “hits”. In such an experiment, each molecule from large small-molecule drug library is evalu- ated for physical property such as the binding affinity (docking score) against a target receptor. In real-life drug discovery experiments, the drug libraries are extremely large but still a minor representation of the essentially infinite chemical space , and evaluation