30 January 2024 in Journal of Chemical Information and Modeling
8 April 2025 in Journal of Chemical Information and Modeling
28 March 2024 in PNAS Nexus
We introduce the kernel-elastic autoencoder (KAE), a self-supervised generative model based on the transformer architecture with enhanced performance for molecular design. KAE employs two innovative loss functions: modified maximum mean discrepancy (m-MMD) and weighted reconstruction ( ). The m-MMD loss has significantly improved the generative performance of KAE when compared to using the traditional Kullback–Leibler loss of VAE, or standard maximum mean discrepancy. Including the weighted reconstruction loss , KAE achieves valid generation and accurate reconstruction at the same time, allowing for generative behavior that is intermediate between VAE and autoencoder not available in existing generative approaches. Further advancements in KAE include its integration with conditional generation, setting a new state-of-the-art benchmark in constrained optimizations. Moreover, KAE has demonstrated its capability to generate molecules with favorable binding affinities in docking applications, as evidenced by AutoDock Vina and Glide scores, outperforming all existing candidates from the training dataset. Beyond molecular design, KAE holds promise to solve problems by generation across a broad spectrum of applications.
9 August 2024 in The Journal of Physical Chemistry A
5 January 2026 in Journal of Chemical Theory and Computation
21 April 2026 in ChemRxiv
Recent advances in deep learning have improved benchmark performance for chemical property prediction, yet reliable transfer to new chemical domains remains limited. A contributing factor is that many models treat molecules primarily as static graphs, ignoring the causal logic of how they are constructed. This review surveys multistep synthesis planning (2020-2026) and argues that the field is undergoing a fundamental transition: from an Era of Navigability (2018-2023), focused on the computational feasibility of finding any route through combinatorial search space, to an Era of Validity (2024-Present), focused on the chemical correctness of those routes. We organize the literature around two dominant paradigms, search-based planning and direct sequence generation, and analyze how their design choices relate to different notions of validity. To resolve the ambiguity of current "solvability" metric, which frequently exceeds 99% by measuring only topological connectivity, we introduce a formalized Hierarchy of Chemical Validity (Solv-N). This framework distinguishes between syntactic (Solv-0) and topological (Solv-1) success, which are largely solved, and the higher-order constraints of selectivity (Solv-2) and executability (Solv-3), which remain open challenges. We critically examine how legacy benchmarks and inflated virtual inventories obscure this distinction, and we conclude with a roadmap for synthesis-aware foundation models evaluated under explicit Tier 2-3 constraints.