Koodli, R. V. | Search bioRxiv

bioRxiv2025

MedSAGE: Bridging Generative AI and Medicinal Chemistry for Structure-Based Design of Small Molecule Drugs

While generative AI is transforming the de novo design of proteins, its effectiveness for structure-based design of small molecules remains limited. Current methods, including diffusion models, often produce small molecules with difficult-to-synthesize structures, poor medicinal chemistry properties, and limited target selectivity. To address these limitations, we introduce MedSAGE, a novel generative AI framework that adapts diffusion models specifically for de novo small-molecule design. Rather than using atoms or strings, we develop a novel representation for generating molecules using fragments relevant for medicinal chemistry. Chemical and geometric information characterizing these fragments is embedded in a smooth and interpretable latent space. We also develop an algorithm to optimize the connectivity between generated fragments while preserving chemical validity and synthesizability. In a benchmark of multiple methods across 25 therapeutically relevant protein targets, MedSAGE achieved state-of-the-art performance, producing synthesizable, drug-like molecules with predicted affinity and selectivity closely matching known drugs and drug candidates. Compared to large-scale virtual screening, MedSAGE produced molecules with high predicted affinity over 100 times more efficiently. Our results demonstrate that MedSAGE is already practically useful and paves the way for next-generation tools in structure-guided drug design.

biophysics↗

bioRxiv2021

Redesigning the Eterna100 for the Vienna 2 folding engine

The rational design of RNA is becoming important for rapidly developing technologies in medicine and biochemistry, spurring development of numerous RNA secondary structure design algorithms and benchmarks to evaluate their performance. However, the problem of RNA design is dependent upon the reverse problem of RNA structure prediction through "folding engines" that predict structure from sequence. We hypothesized that differences in RNA folding engines could impact design algorithms, and recruited an online community of RNA design experts to modify the widely-used RNA secondary structure design benchmark, Eterna100, to address unsolvability of some cases when changing the folding engine used (Vienna 1.8 updated to Vienna 2.6). We tested this new Eterna100-V2 benchmark with five RNA design algorithms, and found that while overall rankings remained similar, the performance of RNA design algorithms that depended on folding engines in their training did indeed depend on which underlying parameter set was used in training. This work demonstrates that the design "difficulty" of RNA structures is intrinsically linked to thermodynamic models, and suggests that future RNA design algorithms that are agnostic to thermodynamic models will result in optimal performance and development. Eterna100-V1 and Eterna100-V2 benchmarks and example solutions are freely available at https://github.com/eternagame/eterna100-benchmarking. Author SummaryDesigning RNA molecules that fold to a desired target structure is an algorithmic problem gathering increasing attention due to the emergence of RNA-based therapies and the need for rational design of RNAs. The Eterna100 dataset, a collection of target structures with increasing design difficulty, designed and selected by players of the crowdsourced RNA game Eterna, has been widely used to benchmark RNA design algorithms. However, these puzzles were originally developed using the now-deprecated version 1 of the ViennaRNA folding engine. In this manuscript, we introduce an updated benchmark, called the Eterna100-V2. We found that nineteen puzzles using Vienna 1 were unsolvable in Vienna 2, but that Eterna participants were able to re-design the puzzles with minimal modifications to make them solvable in Vienna 2. We confirmed that the rankings of 5 RNA design algorithms remained consistent between Eterna100-V1 and -V2. However, discrepancies in performance from algorithms that relied on thermodynamic models in their training suggest that algorithms will benefit from being agnostic to thermodynamic models as these models continue to improve.

bioinformatics↗