Lei Jiang

Molecular function is governed by three-dimensional conformational space, yet foundation models and large language models naturally operate over discrete tokens rather than continuous molecular geometry. Here we introduce Chem3DLLM, a foundation-model interface that makes molecular conformational space accessible through representation, conditioning and physics-informed feedback. Chem3DLLM combines reversible compression of molecular tokenization (RCMT), which encodes atom types, coordinates and bond connectivity into compact language-compatible sequences; protein alignment, which conditions ligand conformational generation on binding-pocket context for structure-based drug design; and reinforcement learning with scientific feedback (RLSF), which uses molecular energetics to refine generated conformations. RCMT gives about threefold character compression while preserving deterministic reconstruction with root-mean-square deviation (RMSD) = 0. In protein-pocket-conditioned generation, Chem3DLLM achieves favourable predicted docking profiles, with average and median Vina scores of -7.21 and -7.10. The same interface supports SMILES-to-3D generation on QM9 with 100\% validity and 100\% uniqueness.Ablation indicates that RLSF adds a modest physics-informed refinement to a representation-driven framework. These results suggest a route for connecting foundation models to 3D molecular design through reversible representation, biological conditioning and physics-informed feedback.

Borehole images are measured by logging tools in a well, providing a microresistivity map of the rock properties surrounding the borehole. These images contain valuable information related to changes in mineralogy, porosity, and fluid content, making them essential for petrophysical analysis. However, due to the special design of borehole imaging tools, vertical strips of gaps occur in borehole images. We develop an effective approach to fill these gaps using a convolutional neural network with partial convolution layers. To overcome the challenge of missing training labels, we introduce a self-supervised learning strategy. Specifically, we replicate the gaps found in borehole images by randomly creating vertical blank strips that mask out certain known areas in the original images. We then use the original images as label data to train the network to recover the known areas masked out by the defined gaps. To ensure that the missing data do not impact the training process, we incorporate partial convolutions that exclude the null-data areas from convolutional computations during forward and backward propagation of updating the network parameters. Our network, trained in this way, can then be used to reasonably fill the gaps originally appearing in the borehole images and obtain full images without any noticeable artifacts. Through the analysis of multiple real examples, we determine the effectiveness of our method by comparing it with three alternative approaches. Our method outperforms the others significantly, as demonstrated by various quantitative evaluation metrics. The filled full-bore images obtained through our approach enable enhanced texture analysis and automated feature recognition.