This paper presents a deep generative framework for molecular design that tackles a key limitation of Generative Adversarial Networks (GANs) when applied to discrete molecular data. Although GANs have shown strong potential in continuous domains, their use in molecular sequence generation remains limited due to training instability and rapid mode collapse, even in advanced variants such as Wasserstein GAN with Gradient Penalty (WGAN-GP). In many cases, these models fail to produce meaningful or valid outputs. To address this issue, we combine Gumbel-Softmax relaxation with the SELFIES encoding scheme, allowing gradients to propagate effectively over discrete sequences. This design improves training stability and enables the model to avoid early collapse, making adversarial training practical in this setting. The model is trained on SMILES-based data converted to SELFIES encoding and evaluated using the GuacaMol distribution-learning benchmark. The results show high validity (97.25%), strong uniqueness (92.96%), and very high novelty (99.21%). Unlike standard GAN configurations that often fail to converge, the proposed approach maintains stable training and produces chemically meaningful and diverse molecules with competitive performance. Overall, the findings suggest that with appropriate design choices, GAN-based models can become a viable, energy-efficient alternative for discrete molecular generation, providing a practical step toward sustainable, fast-tracked generative models for drug discovery.

This paper presents a deep generative framework for molecular design that tackles a key limitation of Generative Adversarial Networks (GANs) when applied to discrete molecular data. Although GANs have shown strong potential in continuous domains, their use in molecular sequence generation remains limited due to training instability and rapid mode collapse, even in advanced variants such as Wasserstein GAN with Gradient Penalty (WGAN-GP). In many cases, these models fail to produce meaningful or valid outputs. To address this issue, we combine Gumbel-Softmax relaxation with the SELFIES encoding scheme, allowing gradients to propagate effectively over discrete sequences. This design improves training stability and enables the model to avoid early collapse, making adversarial training practical in this setting. The model is trained on SMILES-based data converted to SELFIES encoding and evaluated using the GuacaMol distribution-learning benchmark. The results show high validity (97.25%), strong uniqueness (92.96%), and very high novelty (99.21%). Unlike standard GAN configurations that often fail to converge, the proposed approach maintains stable training and produces chemically meaningful and diverse molecules with competitive performance. Overall, the findings suggest that with appropriate design choices, GAN-based models can become a viable, energy-efficient alternative for discrete molecular generation, providing a practical step toward sustainable, fast-tracked generative models for drug discovery.