Title: Enhancing Genotype-Driven Molecular Generation Through Evidence-Based Multi-Objective Latent Perturbation in Diffusion Models

Abstract:

The creation of effective anticancer treatments is hindered by significant tumor heterogeneity and the lack of clear molecular targets across various cancer subtypes. While generative models conditioned on cancer genotypes present a viable path for personalized drug discovery, current methods fall short in explicitly optimizing for a combination of sensitivity, synthesizability, and mechanistic binding plausibility. To address this, we propose a latent-space optimization strategy for a pretrained diffusion model that maps genotypes to drugs. This approach employs a learnable perturbation within the molecular latent space, refined through gradient ascent to maximize a composite reward function. This function integrates predicted drug sensitivity (measured by AUC), drug-likeness (QED), and synthetic accessibility (SAS).

To ensure biological realism, both the reward structure and the evaluation framework are grounded in experimentally derived data from cancer cell lines and validated pharmacologic signals, thereby anchoring the generation of candidates in tangible clinical evidence. Furthermore, we assess mechanistic consistency plausibility using a multi-agent LLM pipeline that leverages the attention mechanisms inherent in the diffusion model. Our experiments, conducted across 15 cancer cell lines from three distinct held-out evaluation sets, reveal consistent and significant improvements over existing baselines in terms of sensitivity, drug-likeness, synthesizability, and chemical validity.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC