Protein function-conditioned language models for variant effect prediction and controllable design

Protein function-conditioned generative modeling can unify two core goals in protein engineering: predicting the effects of sequence variants and designing new sequences that satisfy functional constraints. We present a framework that conditions protein language modeling on molecular function semantics encoded from the Gene Ontology (GO) as a text-attributed directed acyclic graph. By embedding GO text and ontology topology, our function embedding captures relationships among functions and supports cross-function transfer. Function-conditioned meta-pretraining across diverse protein families further enables cross-family transfer before family-specific fine-tuning. Using these GO embeddings as conditioning context, we develop Func2Seq, a transformer-based autoregressive model of P(seq|func) that supports likelihood-based variant effect scoring and controllable sequence sampling. We further introduce Func2Prot, a joint sequence-structure extension that models P(seq, str|func) and enables direct backbone generation via a backbone-angle decoder. On a standard family-specific variant effect prediction benchmark, Func2Seq improves predictive performance over strong alignment-free and alignment-based baselines, and ablations show consistent gains from meta-pretraining when paired with family adaptation. A factor analysis indicates the largest improvements for small or low-diversity families. We also include enzyme fitness case studies to probe when joint sequence-structure modeling helps. For controllable design, we focus on the AT-rich interaction domain and show functional site-level constraint preservation while exploring globally novel sequence space, emphasizing function-critical binding regions and structural plausibility under appropriate evaluation pipelines.