Researchers have developed an artificial intelligence framework that pairs a transformer neural network with a convolutional neural network to predict binding affinities between small drug molecules and biological targets, demonstrating its utility by identifying candidate therapeutics against a key protein implicated in amyotrophic lateral sclerosis. The study, published in Discover Pharmaceutical Sciences, demonstrates how machine learning tools can screen extensive pharmaceutical libraries in silico to rapidly highlight compounds that exhibit strong binding interactions. By evaluating potential therapeutic agents computationally prior to physical testing, this computational strategy is designed to streamline early drug discovery pipelines, lower financial barriers, and reduce the high failure rates that historically impede de novo compound development.

Developing an entirely new therapeutic compound through traditional de novo design represents an exceptionally demanding undertaking that averages 1.24 billion dollars, spans 10 to 17 years, and carries an estimated 45 percent failure rate. To alleviate these significant constraints, pharmaceutical researchers frequently turn to drug repurposing, an approach that systematically investigates whether established medications approved or evaluated for one disease can produce beneficial effects in another. Repurposing existing molecules can diminish developmental constraints by roughly 60 percent while substantially lowering failure risks. In molecular pharmacology, a biological target is commonly a protein tightly linked to a pathological process whose function can be modulated by a small molecule to achieve a beneficial therapeutic response.

Standard laboratory workflows for drug repurposing typically rely on high-throughput screening, an experimental methodology where automated robotic systems apply thousands of candidate compounds across extensive assay panels containing target solutions. While high-throughput screening has successfully yielded important clinical medications, such as maraviroc for the management of HIV following the evaluation of 500,000 compounds against the CC-chemokine receptor 5 target, the approach requires multimillion-dollar investments in physical chemical libraries, specialized robotic machinery, and highly trained personnel. Virtual screening offers an alternative by conducting early evaluations computationally rather than in vitro. Although physical simulation approaches like molecular docking predict how compounds interact within a protein binding site, docking remains computationally intensive, demands high-resolution three-dimensional crystal structures, and exhibits inherent run-to-run variability that limits its throughput across vast compound repositories.

To overcome these physical and computational bottlenecks, the investigators utilized DeepPurpose, an open-source deep learning framework designed to facilitate end-to-end drug-target interaction modeling. Rather than depending on three-dimensional spatial coordinates, the team engineered an encoder-decoder architecture that operates directly on sequence-based representations of drugs and proteins. Chemical structures were encoded from their Simplified Molecular Input Line Entry System notation, a chemical format that describes molecular graphs using ASCII character sequences, while target proteins were represented through their primary amino acid sequences. This sequence-to-sequence structure allows the neural network to formulate binding predictions without requiring crystallized structural conformations or predefined active site coordinates.

The training corpus for the predictive model was retrieved from BindingDB, a major public repository containing over 2.8 million experimentally measured interactions between small molecules and proteins. To ensure biological relevance and model integrity, the raw dataset underwent rigorous preprocessing steps. The researchers restricted entries exclusively to human proteins from Homo sapiens, eliminated records lacking quantitative half-maximal inhibitory concentration annotations, discarded inconsistent or duplicate measurements, removed invalid molecular strings, and standardized chemical structures into their canonical forms. Because experimental binding affinities spanned seven orders of magnitude, all values were converted into base 10 logarithmic space, and entries outside the range of 0.1 to 1,000,000 nanomolar were excluded, leaving a curated dataset of 605,180 unique interactions comprising 450,214 drugs and 1,672 proteins.

The neural network architecture consists of dedicated encoder branches that merge into a dense prediction network. The drug-processing branch employs an eight-layer transformer encoder with an embedding layer that maps a vocabulary of 2,586 chemical tokens alongside 50 positional indices into 128-dimensional vectors regularized by LayerNorm and dropout. Multi-head self-attention mechanisms and two-layer feed-forward networks within each encoder layer generate a unified 128-dimensional molecular representation. Concurrently, the protein-processing branch utilizes a three-layer one-dimensional convolutional neural network with increasing filter channels and kernel sizes to process amino acid strings, projecting the learned sequence features into a 256-dimensional vector through a fully connected layer. These representations are concatenated into a 384-dimensional joint vector and processed through a four-layer multilayer perceptron to predict the scalar binding affinity.

The complete dataset was randomly partitioned into 70 percent training, 10 percent validation, and 20 percent hold-out test sets, ensuring the evaluation tested chemical generalization across novel compounds. Optimization was conducted using the mean squared error loss function, a batch size of 512, and a learning rate of 0.0005 over 30 epochs, requiring eight hours of total execution time on an NVIDIA Tesla P100 graphics processing unit. On the unseen test split, the model achieved a mean squared error of 0.71, corresponding to a root mean squared error of approximately 0.84 log10 units, indicating that the typical prediction variance remained within roughly a factor of seven. When evaluated as a classification tool using an established threshold where half-maximal inhibitory concentrations of 1,000 nanomolar or lower denote strong binders, the architecture exhibited an 86 percent precision and an 89 percent recall.

To rigorously verify generalizability beyond the original training domain, the investigators conducted an external validation against an independent subset of the ChEMBL database comprising 214,634 curated interactions. Although performance metrics on this external set were slightly lower than on the internal test split, which the authors attributed to distribution variations across repositories, the framework maintained a classification precision of 76 percent. Because the model was trained across 1,672 distinct human proteins, it operates in a target-agnostic manner, enabling researchers to predict binding affinities for biological targets that possess little to no dedicated interaction data without requiring architecture retuning or specialized retraining procedures.

The research team applied their trained model to screen the Drug Repurposing Hub, a curated library containing 3,556 approved and investigational molecules, against superoxide dismutase 1, a 153-amino acid homodimeric metalloenzyme heavily implicated in familial amyotrophic lateral sclerosis. In the training database, only 22 recorded interactions involved this protein, and all exhibited weak binding values above 40,000 nanomolar, which the model correctly identified as negative interactions. When screening the repurposing library against the target sequence, the framework identified 21 compounds with predicted affinities below 10 nanomolar, 298 compounds between 10 and 100 nanomolar, and 586 between 100 and 1,000 nanomolar, indicating that filtering by a 1,000 nanomolar threshold eliminates 75 percent of unpromising candidates from physical assay pipelines.

Bibliographic verification of the top 46 predicted binders revealed that 10 molecules had established published connections to the target protein or neurodegenerative pathology. Top candidates included monepantel, an autophagy-inducing compound with a predicted affinity of 3.9 nanomolar previously evaluated in phase 1 trials for amyotrophic lateral sclerosis; the c-Abl inhibitor dasatinib, which had a predicted value of 5.9 nanomolar and improved survival in transgenic mouse models; and the investigational agent thymalfasin, predicted at 4.9 nanomolar, which modulates ubiquitin load in animal disease models. Additional high-ranking molecules included barasertib, milademetan, regorafenib, fostamatinib, tetrandrine, cangrelor, and mangafodipir, providing cross-validation that the sequence-driven deep learning framework successfully enriches libraries for biologically active compounds.

The authors noted specific operational limitations, including the network’s tendency to predict affinity values primarily below 10,000 nanomolar due to the scarcity of weak-binding data points in public repositories, making the architecture best suited for binary ranking rather than exact kinetic measurement. Training evaluations also confirmed that scaling dataset volume directly improved predictive accuracy, with mean squared errors dropping from 0.85 when utilizing 30 percent of the corpus to 0.70 at full data utilization. Future enhancements could incorporate message passing neural networks to capture intricate molecular topologies, provided high-performance computational infrastructure is available to support larger architectures.

Subject of Research: Deep learning model for drug-target affinity prediction and virtual drug screening

Article Title: A deep learning-based model for drug-target affinity prediction with application to drug repurposing and virtual drug screening

Article References: Papikinos, T., Kalykakis, G., Krokidis, M., Vrahatis, A., Vlamos, P., & Exarchos, T. P. (2025). A deep learning-based model for drug-target affinity prediction with application to drug repurposing and virtual drug screening. Discover Pharmaceutical Sciences, 1(1), Article 16. https://doi.org/10.1007/s44395-025-00013-5

Image Credits: AI Generated

DOI: 10.1007/s44395-025-00013-5

Keywords: Deep Learning, Drug Discovery, Drug Repurposing, Virtual Screening, Computational Biology, Amyotrophic Lateral Sclerosis, Superoxide Dismutase 1, Neural Networks, deep, learning-based, model, drug-target