Using Domain-Specific Fingerprints Generated Through Neural Networks to Enhance Ligand-Based Virtual Screening

Abstract

Molecular fingerprints are essential for different cheminformatics approaches like similarity-based virtual screening. In this work, the concept of neural (network) fingerprints in the context of similarity search is introduced in which the activation of the last hidden layer of a trained neural network represents the molecular fingerprint. The neural fingerprint performance of five different neural network architectures was analyzed and compared to the well-established Extended Connectivity Fingerprint (ECFP) and an autoencoder-based fingerprint. This is done using a published compound dataset with known bioactivity on 160 different kinase targets. We expect neural networks to combine information about the molecular space of already known bioactive compounds together with the information on the molecular structure of the query and by doing so enrich the fingerprint. The results show that indeed neural fingerprints can greatly improve the performance of similarity searches. Most importantly, it could be shown that the neural fingerprint performs well even for kinase targets that were not included in the training. Surprisingly, while Graph Neural Networks (GNNs) are thought to offer an advantageous alternative, the best performing neural fingerprints were based on traditional fully connected layers using the ECFP4 as input. The best performing kinase-specific neural fingerprint will be provided for public use.

Natural Product Scores and Fingerprints Extracted from Artificial Neural Networks.

“…However, on all task the GCN performed worse than the other architectures. This falls in line with earlier research of ours [12], and is not uncommon to occur. For example, Jiang and colleagues [34] showed that the Attention FP [35], a graph neural network providing state of the art results on many benchmark sets, performs only on par with descriptor-based models.…”

“…The fingerprint is built upon fragments frequently found in natural products. An alternative strategy could be derived from our recent work, using the activations of trained neural networks as a novel molecular fingerprint [12]. Given that a network is trained to predict the properties of molecules, one should expect that the activations of molecules with similar properties have similar activations in the deeper layers of the network.…”

“…Given that a network is trained to predict the properties of molecules, one should expect that the activations of molecules with similar properties have similar activations in the deeper layers of the network. In our previous work, we showed that these neural networks can also be used to generate target-specific fingerprints implicitly incorporating target-relevant information [12]. A similar approach was chosen by Stojanovic ´and colleagues [13] who trained a Graph Neural Network to predict the binding of molecules to a specific protein target.…”

Dual-view Molecule Pre-training

“…The MACCS (Molecular ACCess System) keys [12] are another commonly used structural descriptors, in which each bit is associated with a predefined SMARTS pattern. As the development of deep learning, molecule fingerprints are learned by neural networks [38,13]. After the emergence and success of pre-training in NLP and CV, people introduce it to molecule representation.…”

GenUI: Interactive and Extensible Open Source Software Platform for De Novo Molecular Generation

“…DNNs found many use cases in molecular virtual screening and de novo compound generation (Figure 1) [19]. This rapidly evolving class of algorithms has been influencing modern drug discovery by building more accurate QSAR models [12,23], creating better molecular representations [24][25][26], predicting 3D protein structure with impressive accuracy [27] or achieving other promising results in many medicinal and clinical applications [3,12,17,21,[28][29][30]. Depending on the architecture, the network is trained either as a bioactivity predictor (e.g.…”