Making Generative Small Molecule Design Better

Over the past several years, generative artificial intelligence (AI) has attracted significant attention in drug discovery, often accompanied by claims that AI alone will revolutionize therapeutic design. The ability of generative models to rapidly propose novel molecular structures, explore vast chemical spaces, and optimize compounds across multiple properties simultaneously represents a genuine technological leap. AI, as a technological disruptor in drug discovery, has promised to accelerate molecular design through unprecedented scale, automation and data-driven creativity. However, as excitement around AI continues to grow, an important question remains: Is generative AI sufficient on its own to reliably produce successful drug candidates?

Experience suggests that transformative innovation rarely comes from a single technology operating in isolation. Instead, the enduring advances emerge when complementary approaches combine to overcome each other’s limitations. In therapeutic discovery, physics-based simulation and classical computational chemistry remain essential for understanding molecular behaviour, predicting binding energetics, and ensuring biological and chemical feasibility. Generative AI introduces powerful exploratory capabilities, but without integration into scientifically grounded modelling and simulation frameworks, AI-generated molecules can struggle to reliably translate into viable therapeutic candidates. The future of small-molecule design is therefore unlikely to be defined by AI replacing classical methods, but rather by integrating both into unified discovery workflows. That is what we have observed with BIOVIA Generative Therapeutics Design (GTD), when combined with the generative NVIDIA MoIMIM algorithm.

Better Together: BIOVIA GTD Seamlessly Integrates NVIDIA MolMIM

BIOVIA Generative Therapeutics Design is an integrated scientific software solution on the 3DEXPERIENCE platform that accelerates the discovery and optimization of small-molecule therapeutics by combining generative AI with established computational chemistry and data management capabilities. The solution provides scientists with a unified environment in which they can define and model biological targets, anti-targets (off-targets and liabilities), and ADMET properties, and combine design objectives such as potency, selectivity, and drug-like properties into a Target Product Profile (TPP). The algorithms automatically generate candidate molecules that are then iteratively optimized to identify those that best satisfy these multiple criteria.

Since its inception, GTD has employed various methods to generate new, related molecules from the best-so-far input molecules. The methods are based on mechanistically correct chemical transformations, group replacements, atom- or bond-replacements, matched molecular pair substitutions and group or reaction enumeration. These methods yield valid structures in every case because they are fundamentally based on the rules of organic chemistry.

NVIDIA MolMIM is a generative model that proposes new small-molecule structures by learning patterns from large collections of known compounds. At a high level, it works by translating molecules into an internal “latent space” numerical representation that implicitly captures their underlying chemical features. Molecules that are chemically similar tend to cluster near one another in this space, enabling smooth exploration of chemical space without performing discrete operations on atoms and bonds.

Once this chemical space has been learned, MolMIM can generate new molecules by starting from an existing compound and making controlled changes, or by making random perturbations to explore nearby regions of chemical space to generate related but novel structures. These generated molecules can then be ranked using external criteria — for example, predicted properties, docking scores or other structure-based evaluations — and the model can iteratively propose new candidates based on the top-ranked structures. In other words, MolMIM is a perfect fit for the GTD workflow of iterative optimization via molecular structure refinement.

The Sum of the Parts: Combining AI with Classical Simulation

Together, MolMIM and the “classic” GTD chemical transformations work well to propose chemically reasonable variations and new scaffolds, while relying on downstream physics-based modelling and domain expertise to determine which designs are most likely to succeed experimentally. When optimizing against a TPP, we have found that MolMIM and chemical transformations work better together than either approach does alone. Evidently, the size and directions of the paths they make through chemical space are in some way complementary.

In this emerging paradigm, innovation is no longer driven by AI or classical simulation alone, but by their convergence into a unified scientific experience. In the future of drug discovery, the greatest breakthroughs occur when the whole is greater than the sum of its parts.

Watch the video to find out how you have seamless access to NVIDIA MolMIM through BIOVIA GTD on the 3DEXPERIENCE platform: