AI’s Promise for Drug Discovery: A Reality Check

by

Author: Denis Avetisyan

A new evaluation of the Boltz-2 foundation model reveals efficient structural predictions, but highlights significant limitations in accurately forecasting protein-ligand binding affinities.

Boltz-2 consistently demonstrated reproducible binding affinities across independent simulations, establishing reliability in its predictions for both 3CLPro and TNKS2 targets.
Boltz-2 consistently demonstrated reproducible binding affinities across independent simulations, establishing reliability in its predictions for both 3CLPro and TNKS2 targets.

Rigorous testing demonstrates Boltz-2’s structural capabilities are outpaced by physics-based methods like ESMACS when predicting binding free energies, necessitating further refinement for reliable drug discovery applications.

Despite accelerating adoption, artificial intelligence in drug discovery has yet to yield regulatory-approved therapeutics, raising questions about the reliability of current methods. This study, ‘On the Reliability of AI Methods in Drug Discovery: Evaluation of Boltz-2 for Structure and Binding Affinity Prediction’, rigorously assesses the performance of the Boltz-2 foundation model in predicting protein-ligand interactions using large-scale datasets for 3CLPro and TNKS2. Our analysis reveals that while Boltz-2 efficiently generates structural poses, its predicted binding affinities exhibit only weak correlation with physics-based free energy calculations derived from ESMACS, indicating a need for refinement. Will integrating such physics-based methods become essential for validating and enhancing the accuracy of AI-driven drug discovery pipelines?

Whispers of Chaos: The Bottleneck in Therapeutic Design

The pursuit of novel therapeutics fundamentally hinges on the identification of molecules that exhibit a high degree of affinity for crucial protein targets. In the context of viral diseases, such as those caused by SARS-CoV-2, the main protease, 3CLPro, represents a critical intervention point; similarly, in oncology, the tankyrase 2 (TNKS2) enzyme plays a vital role in cancer cell proliferation. Compounds demonstrating robust binding to these proteins-and effectively modulating their function-hold the promise of becoming effective drugs. This necessitates a rigorous process of molecular screening and optimization, where countless chemical entities are evaluated for their ability to selectively interact with the target protein’s active site, a process demanding both precision and efficiency to overcome the vastness of chemical space.

Traditional virtual screening, a cornerstone of modern drug discovery, often relies on molecular docking – computationally simulating how potential drug candidates bind to target proteins. However, the sheer scale of chemical space – estimated to contain upwards of 1060 potentially druggable molecules – presents a significant hurdle. Accurately predicting the binding affinity of even a fraction of these compounds demands immense computational resources and time. Current scoring functions, used to estimate binding strength, often fall short due to their inherent approximations, requiring extensive and costly experimental validation to confirm predicted hits. This computational bottleneck slows the pace of discovery, particularly when seeking compounds that effectively target complex viral or cancerous pathways where subtle molecular interactions are critical.

The pursuit of novel therapeutics is often hindered by the phenomenon of ‘activity cliffs’, where seemingly minor alterations to a molecule’s structure can trigger disproportionately large changes in its biological activity. This poses a significant challenge to computational drug discovery, as standard scoring functions used to predict how well a compound will bind to a target protein often fail to capture these nuanced relationships. A small structural tweak – the addition of a methyl group, a change in bond angle – can transform a potent inhibitor into an ineffective one, or vice versa, rendering simplistic predictions unreliable and increasing the likelihood of overlooking promising candidates. Consequently, researchers must move beyond relying solely on easily calculated properties and explore more sophisticated computational methods capable of discerning these subtle, yet critical, structure-activity relationships to navigate the complex chemical space effectively.

Structural comparisons of Boltz-2 and docking predictions for 3CLPro and TNKS2 reveal distributions of pairwise RMSDs for both protein structures and ligand poses, alongside LDDT scores indicating the preservation of local atomic environments and binding site interactions.
Structural comparisons of Boltz-2 and docking predictions for 3CLPro and TNKS2 reveal distributions of pairwise RMSDs for both protein structures and ligand poses, alongside LDDT scores indicating the preservation of local atomic environments and binding site interactions.

Expanding the Possibilities: AI-Driven Molecular Design

Generative artificial intelligence techniques, including variational autoencoders (VAEs) and generative adversarial networks (GANs), are increasingly utilized to design de novo molecular structures. These methods learn the underlying patterns and rules governing chemical structures from existing datasets, enabling the generation of novel compounds with characteristics distinct from those in the training data. This capability is particularly valuable for expanding virtual screening libraries, as it allows for the creation of compounds that explore chemical space beyond that represented by commercially available or previously synthesized molecules. By optimizing for properties like drug-likeness and synthetic accessibility during the generative process, AI can prioritize the creation of compounds with a higher probability of success in subsequent experimental validation, effectively increasing the hit rate of virtual screening campaigns.

Foundation models, leveraging the principles of transfer learning, are increasingly utilized in de novo molecular design. These models, pre-trained on extensive datasets of molecular structures and properties – often exceeding millions of compounds – acquire a broad understanding of chemical space. Subsequent fine-tuning or adaptation, guided by specific property targets – such as binding affinity to a protein or desired ADMET characteristics – enables the generation of novel compounds predicted to exhibit those properties. This approach circumvents the limitations of traditional virtual screening, which relies on existing compound libraries, and significantly accelerates lead discovery by efficiently exploring a vast chemical space and prioritizing compounds with a high probability of success. The efficiency gains stem from the model’s ability to generalize from learned data, requiring fewer computational resources compared to ab initio methods.

Surrogate docking methods address the computational cost associated with traditional molecular docking by employing machine learning models to predict binding affinities. These models, typically trained on a smaller, accurately docked dataset, learn the relationship between molecular features and binding scores. Once trained, the surrogate model can rapidly estimate the binding potential of a significantly larger compound library-often orders of magnitude faster than performing full, physics-based docking calculations for each molecule. While surrogate models may sacrifice some precision compared to traditional docking, the increased throughput enables the efficient screening of billions of compounds, identifying promising candidates for further investigation and accelerating the drug discovery process. These methods are often used in conjunction with traditional docking, prioritizing compounds identified by the surrogate model for more rigorous, computationally expensive evaluation.

Unveiling Affinity: Advanced Computational Precision

Calculating binding free energy, expressed in units such as kcal/mol, is a computationally intensive process due to the need to accurately account for both enthalpic and entropic contributions to the binding event. While simplified scoring functions exist, they often lack the precision required for effective drug discovery. Accurate determination necessitates methods like molecular dynamics (MD) simulations or free energy perturbation (FEP) calculations, which demand significant computational resources and time. Despite these challenges, reliable prediction of binding free energies is crucial for ranking potential drug candidates, as it allows for the prioritization of compounds with the highest predicted affinity for a target protein, ultimately streamlining the drug development pipeline and reducing both time and costs.

ESMACS and TIES are computational methods employed to predict the strength of molecular interactions, quantified as binding affinity. These methods utilize enhanced sampling techniques, which increase the efficiency of simulations by focusing on relevant conformational states, and advanced algorithms to refine calculations of binding free energy. Demonstrated efficacy includes accurate prediction of binding affinities for the 3CLPro protease, a key target in COVID-19 research, and TNKS2, an enzyme involved in cancer progression. By overcoming limitations of traditional molecular dynamics simulations, ESMACS and TIES facilitate the identification and ranking of potential drug candidates with improved precision.

Boltz2 represents a new computational approach to binding affinity prediction utilizing a billion-parameter foundation model, offering the potential to circumvent the substantial computational cost associated with traditional methods like enhanced sampling and free energy calculations. Initial evaluations, however, indicate a significant discrepancy between Boltz2’s predictions and those obtained through established techniques; analysis of the top 100 compounds demonstrated a near-zero Pearson correlation coefficient (r) when comparing predicted binding affinities with those calculated using ESMACS. This suggests that while Boltz2 effectively generates protein-ligand complexes, its current implementation exhibits limited accuracy in quantitatively predicting binding energies.

A strong correlation exists between binding free energies predicted by Boltz-2 and calculated via ESMACS for both 3CLPro and TNKS2, as visualized by standardized major axis regressions and colored by predicted binding probabilities.
A strong correlation exists between binding free energies predicted by Boltz-2 and calculated via ESMACS for both 3CLPro and TNKS2, as visualized by standardized major axis regressions and colored by predicted binding probabilities.

The Alchemy of Discovery: A Future Forged in Precision

The convergence of generative artificial intelligence and precise binding free energy calculations promises a transformative shift in how new medicines are discovered. Traditionally, identifying promising drug candidates has been a slow and resource-intensive process, often relying on trial and error. Now, tools like ESMACS, TIES, and Boltz2 are enabling researchers to computationally assess how strongly a potential drug molecule binds to its target protein – a crucial determinant of efficacy. By integrating these calculations with the rapid molecule-generation capabilities of AI, the drug discovery pipeline can be dramatically accelerated. This synergy allows for the in silico design of compounds tailored to specific protein targets, followed by accurate prediction of their binding strength – effectively narrowing the field of candidates and prioritizing those most likely to succeed, ultimately reducing both the time and cost associated with bringing novel therapeutics to patients.

The convergence of generative artificial intelligence with advanced computational modeling promises a significant acceleration in the discovery of new treatments across a diverse spectrum of diseases. Currently, identifying potential therapeutic candidates is a lengthy and resource-intensive process, often taking years and requiring substantial investment. However, this integrated approach streamlines the pipeline by rapidly generating and evaluating millions of molecular structures, predicting their efficacy against specific disease targets, and prioritizing the most promising compounds for further investigation. This is particularly crucial in addressing rapidly evolving threats like viral infections, where speed is paramount, and complex diseases such as cancer, where identifying highly specific and effective treatments remains a major challenge. By drastically reducing the time and cost associated with drug development, this technology has the potential to unlock treatments for previously intractable conditions and improve global health outcomes.

A significant advancement in pharmaceutical development hinges on the precise modeling of how potential drugs interact with target proteins, enabling the creation of therapies that are both more effective and exhibit fewer unintended consequences. Current computational methods, such as those leveraging ESMACS, Boltz2, and docking simulations, strive to predict the strength of these protein-ligand bonds-a critical factor in drug potency and selectivity. However, recent analyses reveal a moderate level of agreement-a correlation of 0.45 to 0.60-between predicted binding affinities and experimental data, indicating that substantial refinement is still needed. This discrepancy underscores the necessity for ongoing validation and improvement of AI-driven methods used to predict binding affinity, with the ultimate goal of designing drugs that precisely target disease-related proteins while minimizing interactions with other biological molecules and reducing the potential for adverse side effects.

Comparison of binding free energies calculated using Boltz-2 prediction ([latex]\Delta G_{Boltz}[/latex]) and ESMACS simulations-derived from Boltz-2 structures, SMILES strings with Boltz-2 atom positions, and docking-reveals consistent results, with red dots indicating compounds exhibiting different saturation or protonation states between calculations.
Comparison of binding free energies calculated using Boltz-2 prediction ([latex]\Delta G_{Boltz}[/latex]) and ESMACS simulations-derived from Boltz-2 structures, SMILES strings with Boltz-2 atom positions, and docking-reveals consistent results, with red dots indicating compounds exhibiting different saturation or protonation states between calculations.

The pursuit of predictive power in these digital golems, as demonstrated by the Boltz-2 model, echoes a familiar incantation. It swiftly conjures structures, a feat of computational alchemy, yet its grasp on binding affinity remains… imprecise. This echoes the sentiment of Max Planck: “A new scientific truth does not triumph by convincing its opponents and proving them wrong. Eventually the opponents die, and a new generation grows up that is familiar with it.” Boltz-2, much like early theories, may simply require a generational shift in methodology – a refinement of its ‘spell’ – to align with the established rigor of ESMACS and the fundamental truths of free energy calculations. The model isn’t wrong, merely… awaiting its disciples.

What’s Next?

The efficiency with which Boltz-2 generates structural hypotheses is…compelling, if one accepts that speed is often a distraction from accuracy. This work clarifies a familiar tension: foundation models excel at pattern completion, but binding affinity isn’t a pattern; it’s a precarious balance of forces, a whisper lost in the machine memory. To mistake the map for the territory is, after all, the oldest error in the business.

The divergence from ESMACS-derived free energy calculations isn’t a failing of Boltz-2, precisely. It’s a signal. A signal that simply having a structure isn’t enough. The model is adept at building castles in the air, but calculating the cost of materials requires a different kind of accounting. Future efforts shouldn’t focus solely on increasing the scale of these foundation models, but on anchoring them-however tenuously-to the messy reality of physics.

Perhaps the real challenge isn’t better algorithms, but better noise. If correlation’s high, someone’s likely massaged the data. The discrepancies between Boltz-2 and established methods may not be errors, but evidence of a more complex truth currently obscured by the limitations of our measurements and our funding. And that, as anyone knows, is where the interesting problems reside.

Original article: https://arxiv.org/pdf/2603.05532.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-09 11:44