#32 · ML / AI methods

Controlling training form and guidance locality for generative design under limited data

The dominant thread is explicit control over what a generative model is trained on or guided by when labeled data are limited. A controlled study of low-data molecular generation reports that randomizing SAFE atom and fragment order raises novel molecule yield 450-fold over canonical SAFE, isolating training representation as the lever. LOCOGEN reformulates diffusion guidance for regulatory DNA design as a locality problem, protecting learned motif positions while damping out-of-distribution guidance. DAPE-PD adapts the autoregressive Evo model into a bidirectional masked diffusion model for promoter generation. Separately, Evo 2 in-context learning is shown to enable few-shot design of proteins, antibodies, and phage genomes, and BEND 2.0 benchmarks DNA language models on variant effect prediction and design. What is new is treating data form, guidance locality, or context as the explicit determinant of performance in these data-constrained settings. The items span molecules, regulatory DNA, and genomes, so the shared thread is methodological rather than a single task.