Kei Terayama

Finding diverse and high-quality candidates in vast design spaces remains an important challenge in scientific research, including molecular design and material discovery.

Recent advances in generative artificial intelligence have made in silico molecular design a powerful approach for exploring chemical space toward specific goals. However, despite the need for trial-and-error adjustment of generative strategies and reward formulations, most methods implicitly fix the searchable chemical space, significantly limiting flexibility in practical design. This paper introduces ChemTSv3, an exploration framework with a flexible architecture that accommodates diverse design scenarios for adaptive molecular design. Specifically, molecular representations are unified as nodes, including string-based encodings, molecular graphs, and protein sequences. Molecular generations and editing operations are abstracted as transitions between nodes, allowing graph-based modifications, sequential mutations, and large-language-model-driven transformations to be handled within the same formulation. Representations and transition types can be dynamically switched to adapt the search space to the stage and nature of the design task. Here we show that this flexibility enables efficient exploration across diverse design spaces, from drug-like small molecules to proteins.

Accurate assessment of ligand coordinate-density consistency across different resolutions remains challenging in macromolecular crystallography. We introduce the atomic Box Correlation Coefficient (aBCC), an atom-level metric for evaluating the consistency between ligand atomic coordinates and electron density in a resolution-standardized framework. To predict aBCC values from electron-density maps, we developed QAEmap, a machine learning model based on three-dimensional convolutional neural networks (3D-CNNs). The model was trained using Fourier-truncated electron-density maps and corresponding ligand coordinates generated from high-resolution structures in the Protein Data Bank. It was evaluated using both Fourier-truncated electron-density maps and experimentally determined PDB structures. was evaluated using both Fourier-truncated electron-density maps and experimentally determined PDB structures.The prediction accuracy gradually decreased with decreasing resolution, but remained reliable up to ~3.5 angstrom. These results demonstrate that aBCC enables resolution-standardized atom-wise evaluation of coordinate-density consistency across different resolutions and provide a foundation for further development and refinement of machine learning-based coordinate validation.

A series of FLAP molecules are one of the functional fluorescent probes that can sense subtle changes in local environments such as viscosity, polymer free volume, and external mechanical force. In this work, potential energy surfaces (PESs) of a representative Flexible Aromatic Photofunctional (FLAP) molecule, FLAP0, which bears anthraceneimide wings fused with a cyclooctatetraene (COT) core, have been investigated using density functional theory (DFT) and time-dependent DFT (TD-DFT). The calculation results indicate that, upon excitation with 350–400 nm light, the accessible states are the second (S2) and third (S3) singlet excited states. FLAP0 in these states undergoes nonadiabatic transitions to the first excited (S1) state via S3/S2 and S2/S1 conical intersections (CIX). Notably, a transition state (TS) on the S1 PES near the S2/S1 CIX exhibits a symmetry-breaking motion rather than flapping of the anthraceneimide wings, depriving the driving force for planarization. After reaching the S1 state with a shallow V-shaped geometry, therefore, FLAP0 planarization is thermally induced by COT hydrogen wagging. Due to the low barrier for this motion, the V-shaped and planar forms exist in thermal equilibrium on the S1 surface rather than undergoing unidirectional transformation. This equilibrium facilitates access to diverse energy minima around the planar S1 geometry, leading to the structured green emission bands. Consequently, the putative “vibronic structure” in solution is revealed to be the sum of emissions from these distinct planar forms.

Recent advances in generative AI have sparked anticipation of a future where functional molecules that potentially outperform existing ones can be freely designed. However, as a comprehensive understanding of the property space defined by known functional molecules is lacking, assessing whether new AI/human-designed molecules truly surpass existing ones is challenging. To address this, we computed ~50 experimentally observable properties for >5 million molecules curated from three datasets covering commercially available, reported, and artificially constructed compounds using density functional theory. Based on them, we developed MolAtlas, a visualization system that reveals statistical boundaries and property relationships within an observable property space, offering a reference framework for evaluating molecular novelty. As an example, by analyzing frontier orbital energies, we proposed an empirical requirement for molecular air-stability. We further demonstrated the utility of the system by identifying a small fluorescent compound and characterizing charge-retaining molecules for liquid electrets. Thus, MolAtlas provides a data-driven foundation for navigating the molecular property space and accelerating functional molecule design.

Autonomous oscillations are fundamental nonequilibrium processes vital to life that contribute to homeostasis and robustness against noise from the external environment. Inspired by these biological systems, autonomously oscillating materials have been extensively developed to create lifelike objects. However, excitable dynamics, which are characteristic of threshold-triggered signaling in neurons, remain unexplored in the design of synthetic oscillatory materials despite their unique potential in signal processing to acquire adaptive functionalities. Herein, we report oil droplets on a water surface that exhibit autonomous oscillations with excitable dynamics driven by the assembly of two oppositely charged surfactants. The oscillation consists of a repetitive cycle of gradual expansion and rapid contraction of the interfacial areas of the oil droplets. Using AI-based image analysis, we quantified the dynamics and developed models to elucidate the underlying mechanism: the interfacial area gradually increased as surfactants accumulated at the oil–water interface and rapidly decreased upon threshold-triggered release of nanoemulsions, which effectively removed surfactants from the interface. Furthermore, we demonstrated that the oscillation is influenced by environmental fluctuations, introducing stochasticity into an otherwise deterministic excitable oscillator. This autonomous oscillatory system, analogous to the “integrate-and-fire” model of neural activity, provides a minimal physicochemical framework for designing adaptive liquid-like soft materials with complex excitable dynamics.

Inverse analysis of semiconductor devices often suffers from non‐unique solutions, known as multivaluedness, within a large‐scale parameter space. Here, we show that a tandem neural network (tandem NN), which couples a pretrained forward model to an inverse model and jointly minimizes prediction and reconstruction losses, overcomes this challenge for amorphous In–Ga–Zn–O thin‐film transistors. Trained on 1000 simulated transfer curves covering six intrinsic material parameters varied across ranges three to five orders of magnitude wider than in previous studies, the network infers multiple physical parameters from a single current–voltage curve in less than 1 ms with R ² = 0.99. The inferred parameters reproduce experimental current–voltage characteristics of lab‐fabricated devices without additional fitting, confirming physical validity. Compared with conventional TCAD iterative fitting, the tandem NN provides significant acceleration and paves the way for autonomous experimentation for materials discovery, digital‐twin frameworks in next‐generation transistor manufacturing, and other multivalued inverse problem domains.

The rational design of coordination complexes with precisely controlled excited-state dynamics remains a major challenge in the discovery of functional materials, particularly in multidimensional chemical spaces. In this context, aluminum complexes are particularly attractive because of their earth abundance, low toxicity, and unique ability to form structurally robust, luminescent helicates with versatile photophysical properties. Here, we report an AI-assisted computational–experimental strategy for metal complexes to design aluminum(III) dinuclear helicates exhibiting efficient long-wavelength emission. By integrating density functional theory calculations with asynchronous Bayesian optimization, we systematically modulated ligand electron density, conjugation length, and symmetry to control inter- 2 and intraligand charge–transfer processes. On the basis of the optimized methods, we synthesized five helicates, which displayed deep-red emission with quantum yields of up to 0.96 (λmax = 603 nm) and broadband emission extending into the near-infrared region (λmax = 671 nm), with considerable spectral weight extending to approximately 800 nm, remarkably properties in aluminum complexes. Transient absorption spectroscopy revealed that suppression of interligand charge transfer minimizes nonradiative decay and enables high quantum yields. The large apparent Stokes shifts arise from the combined effects of intraligand charge transfer and pronounced excited-state structural relaxation, which work together to stabilize the emissive state. In addition, the helicates further exhibit circularly polarized luminescence after optical resolution. These findings establish a scalable and generalizable AI-assisted strategy for designing sustainable coordination luminophores beyond purely organic systems.

π-Conjugated hydrocarbons are foundational molecular scaffolds in materials science. However, their design and discovery are limited by the lack of systematic frameworks for exploring their chemical space while preserving the structural validity. We present a unit-based molecule generation framework tailored to π-conjugated hydrocarbons, named CARBOT. This framework comprises a minimal set of chemically rational unit transformations, including vinyl- and ethynyl-group addition and cyclization via C–C bond formation. CARBOT is provably sound and complete within the scope of neutral Kekulé hydrocarbons, with carbons restricted to sp and sp² hybridization, excluding allenes and carbenes. To accelerate access to aromatic frameworks, additional composite transformations are introduced. A depth-first search combined with a substructure-matching score enables rediscovery of a wide range of π-conjugated hydrocarbons within the defined space, including nonplanar nanographenes, highly strained carbon nanobelts, and Möbius hydrocarbons, and it can be scaled up to molecules as large as C₂₀₀H₁₀₀. The molecule generator is expected to provide a systematic foundation for the design and discovery of functional π-conjugated molecules.

Advances in experimental techniques have expanded the volume of biological data. This has increased the demand for structured information extraction from papers, with large language models (LLMs) considered promising. However, challenges remain, including limited validation in biology and unclear applicability to multimodal tasks that integrate text with domain-specific figures, such as microscopic images and scatter plots. Here, we developed a multimodal LLM (MLLM)-based workflow to extract the experimental conditions and phase status from the text and figures of experimental papers on liquid-liquid phase separation and validated the effect of various inputs, prompts, and MLLM types. As a result, the MLLM-based extraction methods achieved F1-scores over 0.80 by processing each figure as a processing unit and inputting domain-specific prompts reflecting manual extraction guidance. This study demonstrates the potential and limitations of MLLMs for extracting experimental information using a focused set of LLPS papers as a model case and provides insights into the possibility of advancing multimodal approaches in biology.