Geometric deep learning-guided Suzuki reaction conditions assessment for applications in medicinal chemistry.

Suzuki cross-coupling reactions are considered a valuable tool for constructing carbon-carbon bonds in small molecule drug discovery. However, the synthesis of chemical matter often represents a time-consuming and labour-intensive bottleneck. We demonstrate how machine learning methods trained on high-throughput experimentation (HTE) data can be leveraged to enable fast reaction condition selection for novel coupling partners. We show that the trained models support chemists in determining suitable catalyst-solvent-base combinations for individual transformations including an evaluation of the need for HTE screening. We introduce an algorithm for designing 96-well plates optimized towards reaction yields and discuss the model performance of zero- and few-shot machine learning. The best-performing machine learning model achieved a three-category classification accuracy of 76.3% (±0.2%) and an F 1-score for a binary classification of 79.1% (±0.9%). Validation on eight reactions revealed a receiver operating characteristic (ROC) curve (AUC) value of 0.82 (±0.07) for few-shot machine learning. On the other hand, zero-shot machine learning models achieved a mean ROC-AUC value of 0.63 (±0.16). This study positively advocates the application of few-shot machine learning-guided reaction condition selection for HTE campaigns in medicinal chemistry and highlights practical applications as well as challenges associated with zero-shot machine learning.

Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning.

Late-stage functionalization is an economical approach to optimize the properties of drug candidates. However, the chemical complexity of drug molecules often makes late-stage diversification challenging. To address this problem, a late-stage functionalization platform based on geometric deep learning and high-throughput reaction screening was developed. Considering borylation as a critical step in late-stage functionalization, the computational model predicted reaction yields for diverse reaction conditions with a mean absolute error margin of 4-5%, while the reactivity of novel reactions with known and unknown substrates was classified with a balanced accuracy of 92% and 67%, respectively. The regioselectivity of the major products was accurately captured with a classifier F-score of 67%. When applied to 23 diverse commercial drug molecules, the platform successfully identified numerous opportunities for structural diversification. The influence of steric and electronic information on model performance was quantified, and a comprehensive simple user-friendly reaction format was introduced that proved to be a key enabler for seamlessly integrating deep learning and high-throughput experimentation for late-stage functionalization.

Identifying opportunities for late-stage C-H alkylation with high-throughput experimentation and in silico reaction screening.

Enhancing the properties of advanced drug candidates is aided by the direct incorporation of specific chemical groups, avoiding the need to construct the entire compound from the ground up. Nevertheless, their chemical intricacy often poses challenges in predicting reactivity for C-H activation reactions and planning their synthesis. We adopted a reaction screening approach that combines high-throughput experimentation (HTE) at a nanomolar scale with computational graph neural networks (GNNs). This approach aims to identify suitable substrates for late-stage C-H alkylation using Minisci-type chemistry. GNNs were trained using experimentally generated reactions derived from in-house HTE and literature data. These trained models were then used to predict, in a forward-looking manner, the coupling of 3180 advanced heterocyclic building blocks with a diverse set of sp3-rich carboxylic acids. This predictive approach aimed to explore the substrate landscape for Minisci-type alkylations. Promising candidates were chosen, their production was scaled up, and they were subsequently isolated and characterized. This process led to the creation of 30 novel, functionally modified molecules that hold potential for further refinement. These results positively advocate the application of HTE-based machine learning to virtual reaction screening.

Morphing of Amphipathic Helices to Explore the Activity and Selectivity of Membranolytic Antimicrobial Peptides.

Naturally occurring membranolytic antimicrobial peptides (AMPs) are rarely cell-type selective and highly potent at the same time. Template-based peptide design can be used to generate AMPs with improved properties de novo. Following this approach, 18 linear peptides were obtained by computationally morphing the natural AMP Aurein 2.2d2 GLFDIVKKVVGALG into the synthetic model AMP KLLKLLKKLLKLLK. Eleven of the 18 chimeric designs inhibited the growth of Staphylococcus aureus, and six peptides were tested and found to be active against one resistant pathogenic strain or more. One of the peptides was broadly active against bacterial and fungal pathogens without exhibiting toxicity to certain human cell lines. Solution nuclear magnetic resonance and molecular dynamics simulation suggested an oblique-oriented membrane insertion mechanism of this helical de novo peptide. Temperature-resolved circular dichroism spectroscopy pointed to conformational flexibility as an essential feature of cell-type selective AMPs.

In silico design and optimization of selective membranolytic anticancer peptides.

Membranolytic anticancer peptides represent a potential strategy in the fight against cancer. However, our understanding of the underlying structure-activity relationships and the mechanisms driving their cell selectivity is still limited. We developed a computational approach as a step towards the rational design of potent and selective anticancer peptides. This machine learning model distinguishes between peptides with and without anticancer activity. This classifier was experimentally validated by synthesizing and testing a selection of 12 computationally generated peptides. In total, 83% of these predictions were correct. We then utilized an evolutionary molecular design algorithm to improve the peptide selectivity for cancer cells. This simulated molecular evolution process led to a five-fold selectivity increase with regard to human dermal microvascular endothelial cells and more than ten-fold improvement towards human erythrocytes. The results of the present study advocate for the applicability of machine learning models and evolutionary algorithms to design and optimize novel synthetic anticancer peptides with reduced hemolytic liability and increased cell-type selectivity.

Designing Anticancer Peptides by Constructive Machine Learning.

Constructive (generative) machine learning enables the automated generation of novel chemical structures without the need for explicit molecular design rules. This study presents the experimental application of such a deep machine learning model to design membranolytic anticancer peptides (ACPs) de novo. A recurrent neural network with long short-term memory cells was trained on α-helical cationic amphipathic peptide sequences and then fine-tuned with 26 known ACPs by transfer learning. This optimized model was used to generate unique and novel amino acid sequences. Twelve of the peptides were synthesized and tested for their activity on MCF7 human breast adenocarcinoma cells and selectivity against human erythrocytes. Ten of these peptides were active against cancer cells. Six of the active peptides killed MCF7 cancer cells without affecting human erythrocytes with at least threefold selectivity. These results advocate constructive machine learning for the automated design of peptides with desired biological activities.

Recurrent Neural Network Model for Constructive Peptide Design.

We present a generative long short-term memory (LSTM) recurrent neural network (RNN) for combinatorial de novo peptide design. RNN models capture patterns in sequential data and generate new data instances from the learned context. Amino acid sequences represent a suitable input for these machine-learning models. Generative models trained on peptide sequences could therefore facilitate the design of bespoke peptide libraries. We trained RNNs with LSTM units on pattern recognition of helical antimicrobial peptides and used the resulting model for de novo sequence generation. Of these sequences, 82% were predicted to be active antimicrobial peptides compared to 65% of randomly sampled sequences with the same amino acid distribution as the training set. The generated sequences also lie closer to the training data than manually designed amphipathic helices. The results of this study showcase the ability of LSTM RNNs to construct new amino acid sequences within the applicability domain of the model and motivate their prospective application to peptide and protein design without the need for the exhaustive enumeration of sequence libraries.

Generative Recurrent Networks for De Novo Drug Design.

Generative artificial intelligence models present a fresh approach to chemogenomics and de novo drug design, as they provide researchers with the ability to narrow down their search of the chemical space and focus on regions of interest. We present a method for molecular de novo design that utilizes generative recurrent neural networks (RNN) containing long short-term memory (LSTM) cells. This computational model captured the syntax of molecular representation in terms of SMILES strings with close to perfect accuracy. The learned pattern probabilities can be used for de novo SMILES generation. This molecular design concept eliminates the need for virtual compound library enumeration. By employing transfer learning, we fine-tuned the RNN's predictions for specific molecular targets. This approach enables virtual compound design without requiring secondary or external activity prediction, which could introduce error or unwanted bias. The results obtained advocate this generative RNN-LSTM system for high-impact use cases, such as low-data drug discovery, fragment based molecular design, and hit-to-lead optimization for diverse drug targets.

Rational Design of Membrane-Pore-Forming Peptides.

Specific interactions of peptides with lipid membranes are essential for cellular communication and constitute a central aspect of the innate host defense against pathogens. A computational method for generating innovative membrane-pore-forming peptides inspired by natural templates is presented. Peptide representation in terms of sequence- and topology-dependent hydrophobic moments is introduced. This design concept proves to be appropriate for the de novo generation of first-in-class membrane-active peptides with the anticipated mode of action. The designed peptides outperform the natural template in terms of their antibacterial activity. They form a kinked helical structure and self-assemble in the membrane by an entropy-driven mechanism to form dynamically growing pores that are dependent on the lipid composition. The results of this study demonstrate the unique potential of natural template-based peptide design for chemical biology and medicinal chemistry.

Peptide-Membrane Interaction between Targeting and Lysis.

Certain cationic peptides interact with biological membranes. These often-complex interactions can result in peptide targeting to the membrane, or in membrane permeation, rupture, and cell lysis. We investigated the relationship between the structural features of membrane-active peptides and these effects, to better understand these processes. To this end, we employed a computational method for morphing a membranolytic antimicrobial peptide into a nonmembranolytic mitochondrial targeting peptide by "directed simulated evolution." The results obtained demonstrate that superficially subtle sequence modifications can strongly affect the peptides' membranolytic and membrane-targeting abilities. Spectroscopic and computational analyses suggest that N- and C-terminal structural flexibility plays a crucial role in determining the mode of peptide-membrane interaction.

modlAMP: Python for antimicrobial peptides.

SUMMARY: We have implemented the lecular esign aboratory's nti icrobial eptides package ( ), a Python-based software package for the design, classification and visual representation of peptide data. modlAMP offers functions for molecular descriptor calculation and the retrieval of amino acid sequences from public or local sequence databases, and provides instant access to precompiled datasets for machine learning. The package also contains methods for the analysis and representation of circular dichroism spectra. AVAILABILITY AND IMPLEMENTATION: The modlAMP Python package is available under the BSD license from URL http://doi.org/10.5905/ethz-1007-72 or via pip from the Python Package Index (PyPI). CONTACT: gisbert.schneider@pharma.ethz.ch. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Hybrid Network Model for "Deep Learning" of Chemical Data: Application to Antimicrobial Peptides.

We present a "deep" network architecture for chemical data analysis and classification together with a prospective proof-of-concept application. The model features a self-organizing map (SOM) as the input layer of a feedforward neural network. The SOM converts molecular descriptors to a two-dimensional image for further processing. We implemented lateral neuron inhibition for contrast enhancement. The model achieved improved classification accuracy and predictive robustness compared to feedforward network classifiers lacking the SOM layer. By nonlinear dimensionality reduction the networks extracted meaningful chemical features from the data and outperformed linear principal component analysis (PCA). The learning machine was trained on the sequence-length independent recognition of antibacterial peptides and correctly predicted the killing activity of a synthetic test peptide against Staphylococcus aureus in an in vitro experiment.

Sparse Neural Network Models of Antimicrobial Peptide-Activity Relationships.

We present an adaptive neural network model for chemical data classification. The method uses an evolutionary algorithm for optimizing the network structure by seeking sparsely connected architectures. The number of hidden layers, the number of neurons in each layer and their connectivity are free variables of the system. We used the method for predicting antimicrobial peptide activity from the amino acid sequence. Visualization of the evolved sparse network structures suggested a high charge density and a low aggregation potential in solution as beneficial for antimicrobial activity. However, different training data sets and peptide representations resulted in greatly varying network structures. Overall, the sparse network models turned out to be less accurate than fully-connected networks. In a prospective application, we synthesized and tested 10 de novo generated peptides that were predicted to either possess antimicrobial activity, or to be inactive. Two of the predicted antibacterial peptides showed cosiderable bacteriostatic effects against both Staphylococcus aureus and Escherichia coli. None of the predicted inactive peptides possessed antibacterial properties. Molecular dynamics simulations of selected peptide structures in water and TFE suggest a pronounced peptide helicity in a hydrophobic environment. The results of this study underscore the applicability of neural networks for guiding the computer-assisted design of new peptides with desired properties.