The work, led by Shivangi Gupta with Vineetha Menon and Jerome Baudry, confronts a problem that has quietly plagued computational drug discovery for years. Molecular dynamics simulations can now generate more than a million protein conformations and protein-ligand complexes for a single target, along with associated interaction energies. Yet only a tiny fraction of those conformations correspond to true binding-competent states. Machine learning models trained on such data face an extreme class imbalance: non-binding conformations vastly outnumber binding ones, sometimes by ratios as severe as 41 to 1. Traditional classifiers simply learn to ignore the rare binding states — precisely the ones drug hunters care about most.

The team’s insight was to stop looking at protein motion as a raw time series and start looking at it as a spectrum. Molecular dynamics trajectories sample protein conformations at uniform time intervals over spans of hundreds of nanoseconds to microseconds, producing subtle but meaningful structural transitions that are hard to detect in the temporal domain. The researchers applied two spectral transforms to their protein feature data: the Fast Fourier Transform, which converts signals into pure frequency components, and the Discrete Wavelet Transform, which provides simultaneous time and frequency resolution. FFT-based features effectively captured high-frequency conformational signatures, while DWT-based features identified localized, high-energy patterns associated with ligand-binding events.

The framework was evaluated on three G protein-coupled receptors — ADORA2A, OPRD1, and OPRK1 — using previously established datasets of roughly 3,000 protein conformations each, described by 50 physicochemical descriptors extracted with the Molecular Operating Environment platform. Class imbalance was substantial across all three: 3 to 1 for ADORA2A, 21 to 1 for OPRK1, and 41 to 1 for OPRD1. Models were trained on just 30 percent of each dataset and evaluated on the remaining 70 percent, a deliberately conservative strategy designed to test whether strong predictions could be achieved with limited training data while reducing overfitting risk.

Feature selection came first, using four complementary machine learning methods — Analysis of Variance, Mutual Information, Recurrence Quantification Analysis, and Spearman correlation — to identify the most informative protein descriptors. A unified ranking score integrated the selections from all four methods, and only features chosen by every method received the maximum consensus score of 4. Remarkably, just three descriptors survived this rigorous filtering for all three targets: water accessible surface area, hydrophobic surface area, and hydrophobicity moment. Their consistent performance across multiple targets suggests they capture key characteristics that distinguish binding from non-binding conformations.

Data rebalancing followed, using the spectral domain to guide both under-sampling and oversampling. High-energy wavelet or Fourier coefficients were identified, sorted by total energy or amplitude, and mapped back to the time domain. Non-binding conformations at unidentified indexes were reduced, while generative adversarial networks synthesized new minority-class binding conformations at identified spectral indexes. Unlike traditional oversampling techniques that duplicate existing samples and increase overfitting risk, GANs learn the underlying data distribution and generate diverse synthetic conformations that better capture the variability of binding states. The synthesized spectral-domain samples were subsequently mapped back to the temporal domain to ensure the generated conformations retained biologically meaningful structural characteristics.

Three classifiers were then trained on the rebalanced data: Gaussian Naive Bayes, Extreme Gradient Boosting, and a convolutional neural network with ReLU activation and binary cross-entropy loss. A rank-based decision-fusion strategy combined their predictions, evaluating each classifier on both specificity and sensitivity and assigning separate ranks from 0 to 2 for each metric. Rather than relying on a fixed ensemble whose behavior may be biased toward the original training distribution, the framework dynamically selects the highest-performing models as new data becomes available. A prediction is accepted only when the aggregated probability reaches at least 66.67 percent, ensuring stronger agreement among the ensemble classifiers.

The results were striking. For ADORA2A, WaveGAN-Fusion achieved 64.72 percent accuracy, 95.20 percent sensitivity, an AUC of 0.73, and an MCC of 0.44 — substantially better than any individual base learner. For OPRD1, FFT-GAN-Fusion reached 97.14 percent accuracy, 94.44 percent sensitivity, an AUC of 0.95, and an MCC of 0.66 under the most extreme 41-to-1 imbalance. For OPRK1, WaveGAN-Fusion delivered an AUC of 0.81 and an MCC of 0.30, while FFT-GAN-Fusion attained the highest overall accuracy at 95.18 percent. Notably, several individual base learners produced negative MCC values under extreme imbalance, indicating performance worse than random guessing — a limitation the fusion mechanism effectively compensated for by combining complementary decision boundaries.

Enrichment ratios provided a biologically grounded validation of the framework’s practical value. For ADORA2A, the fusion models achieved a maximum enrichment of 13.91, substantially higher than the baseline enrichment of 9 obtained without any AI-based conformation selection. For OPRD1, the framework reached an enrichment of 40 — twenty times higher than the baseline of 2. For OPRK1, the maximum enrichment was 26.51, well above the baseline of 5.5. These numbers demonstrate that the spectral-based models not only improve predictive accuracy but also meaningfully concentrate biologically relevant binding conformations at the top of ranked predictions, which is exactly what downstream molecular docking and experimental validation require.

The broader implications reach into one of the most expensive steps in structure-based drug discovery: ensemble docking. Traditional workflows require evaluating thousands of MD-derived conformations, many of which are non-binding and contribute minimally to screening accuracy. Prior approaches — clustering-based selection, RMSD filtering, energy-based pruning — have long struggled with redundancy and computational cost. Recent CASP assessments similarly highlight the persistent difficulty of capturing alternative protein conformations, while emerging tools like Microsoft’s BioEmu can emulate what would take 100,000 GPU hours of atomistic MD in minutes, though they inherit limitations such as memorization bias. By using spectral-domain signatures to identify binding-relevant conformations early, this framework offers a practical, scalable filter that can be integrated into the earliest stages of the drug-discovery pipeline — automatically labeling unlabeled MD conformations, filtering out non-binding states, and selectively docking predicted binding conformations to confirm their relevance. The team plans to validate predicted conformations through molecular docking and extend the framework to additional therapeutic protein targets, establishing a feedback loop in which docking outcomes iteratively refine the spectral-domain classifiers and progressively improve predictive accuracy.

Subject of Research: Computational Biology and Drug Discovery

Article Title: Spectral analysis and AI/ML-based data-driven approaches for enhanced protein conformation selection and prediction in drug discovery

Article References: Gupta, S., Menon, V., & Baudry, J. (2026). Spectral analysis and AI/ML-based data-driven approaches for enhanced protein conformation selection and prediction in drug discovery. Frontiers in Drug Discovery, 6, Article 1805731. https://doi.org/10.3389/fddsv.2026.1805731

Image Credits: AI Generated

DOI: 10.3389/fddsv.2026.1805731

Keywords: artificial intelligence, drug discovery, molecular dynamics, machine learning, protein conformation, spectral analysis, GPCR, generative adversarial networks, Spectral, analysis, ML-based, data-driven