Negative Extrapolation for Targeted Motif Suppression in Molecular Generators

Abstract Broadly trained molecular generators can cover diverse chemical space, but may also produce motifs that are undesirable in a given discovery context, e.g., reactive motifs. Adapting a pretrained generator to suppress such chemistry is therefore an important practical problem. However, most model steering methods enrich desirable compounds or optimize a predefined reward, whereas selectively suppressing an undesirable region of chemical space without compromising the quality and diversity of the remaining distribution remains challenging. Here, we adapt negative extrapolation (NE), a parameter-space editing strategy in which a generator is fine-tuned on molecules containing unwanted structural motifs and the resulting weight update is reversed, to suppress these motifs directly within the generator. We evaluated NE using two custom string-based autoregressive generators, a public pretrained recurrent prior, and a three-dimensional flow-matching generator, with unedited generation and conventional fine-tuning on molecules without the unwanted motifs as principal comparators. Across the string-based generators, NE reduced the occurrence of four structural motif classes (reactive, metal-binding, charged, and assay-interference motifs) more strongly than conventional fine-tuning. We further extended NE through selective parameter editing and combinations of learned update directions, enabling generative validity to be preserved across different model architectures. In the three-dimensional flow-matching generator, corrected NE reduced the fraction of valid molecules containing any of the unwanted motifs from 31.1% to 2.2% and increased the fixed-budget yield of unique, motif-free molecules that were valid molecular graphs and passed three-dimensional chemical and geometric consistency checks from 60.0% to 80.8%. Together, these results support NE as a simple approach for selectively editing molecular generators across architectures and across string-based and three-dimensional representations, while preserving molecular diversity and without requiring a reward-optimization loop. Scientific Contribution: We apply the parameter-arithmetic principle of reversing a learned update to an important problem in molecular generation: selectively suppressing a defined unwanted region of chemical space while preserving the quality and diversity of the remaining distribution. Across string-based and three-dimensional generators, we further investigate how parameter arithmetic can be adapted through selective parameter editing and combinations of independently learned update directions to retain model validity when full-model extrapolation is disruptive.