Guiding Generative Models for Protein Design: Prompting, Steering and Aligning

Abstract: Generative AI models learn probability distributions from data and produce novel samples that capture the salient properties of their training sets. Proteins are particularly attractive for such approaches given their abundant data and the versatility of their representations, ranging from sequences to structures and functions. This versatility has motivated the rapid development of generative models for protein design, enabling the generation of functional proteins and enzymes with unprecedented success. However, because these models mirror their training distribution, they tend to sample from its most probable modes, while low-probability regions, often encoding valuable properties, remain underexplored. To address this challenge, recent work has focused on guiding generative models to produce proteins with user-specified properties, even when such properties are rare or absent from the original training distribution. In this review, we survey and categorize recent advances in conditioning generative models for protein design. We distinguish approaches that modify model parameters, such as reinforcement learning or supervised fine-tuning, from those that keep the model fixed, including conditional generation, retrieval-augmented strategies, Bayesian guidance, and tailored sampling methods. Together, these developments are beginning to enable the steering of generative models toward proteins with desired, and often previously inaccessible, properties.

• The paper presents a comprehensive review of methods to align generative protein models with specific functional targets using both training and inference-time interventions.
• It details approaches ranging from supervised fine-tuning and various reinforcement learning algorithms to advanced inference-time techniques such as prompt conditioning and Bayesian re-weighting.
• The study identifies key challenges like data bias and out-of-distribution failures, while outlining future directions that integrate experimental feedback and physics-informed strategies.

Guiding Generative Models for Protein Design: Prompting, Steering, and Aligning

Overview

The paper "Guiding Generative Models for Protein Design: Prompting, Steering and Aligning" (2511.21476) offers a comprehensive review of emerging methodologies for controlling generative models in protein design. The analysis is centered around systematically categorizing interventions by whether they operate during model training or inference, and provides an in-depth comparison of methods ranging from supervised fine-tuning and reinforcement learning (RL) to modern inference-time steering approaches like prompt engineering, retrieval-augmented generation (RAG), output guidance, activation steering, Bayesian re-weighting, and sampling strategies.

Problem Formulation and Motivation

Protein generative models (GMPDs) such as diffusion models, autoregressive and masked protein LLMs (pLMs), and hybrid sequence-structure generators have enabled de novo generation of binders, enzymes, and molecular scaffolds. Yet, these models largely adhere to the statistical properties present in their training corpora, intrinsically biasing the samples towards high-probability, naturalistic regions of sequence/structure space. This is insufficient when engineering for rare or out-of-distribution functions—e.g., extremophile stability in sequences from mesophilic datasets, or novel catalytic mechanisms absent from annotated protein databases.

The core technical challenge is to generate sequences $x$ exhibiting user-defined properties $y$—formally, to sample from $p(x|y)$ rather than $p(x)$—even when the support of $p(x)$ and $p(x|y)$ are disjoint or $y$-annotated data are extremely sparse.

Train-Time Methods

Supervised Fine-Tuning

Supervised fine-tuning (SFT) adapts general pLMs or generative diffusion models to specific protein subclasses or functional targets using curated data, shifting $p(x)$ toward the conditional distribution over high-fidelity examples. SFT exhibits strong results in domain adaptation for enzyme families [madani2023large, munsamy_conditional_2024], genome editors [ruffolo_design_2024], and biological nanomachines [king2025generative]. However, SFT is fundamentally data-limited, cannot express nuanced preferences over a continuum of fitness, and is prone to catastrophic forgetting [shumailov2024ai].

Reinforcement Learning

RL surrogates the missing property annotations or explicit sequence–property mapping with a reward or preference signal. A variety of RL algorithms have been translated from general sequence modeling:

• REINFORCE and policy gradient (AB-Gen, [xu2023ab]) provide basic optimization over sequence property functions.
• Proximal Policy Optimization (PPO), a trust-region method, stabilizes updates and is the workhorse for RL from experimental or human feedback, as in RLXF [blalock_functional_2025], ProteinZero [wang2025proteinzero], and model alignment to experimentally measured objectives.
• AlphaZero-style MCTS has outperformed classical tree searches for backbone-level design and multi-chain complexes [renard_model-based_2024, lin2025highplay].
• Direct Preference Optimization (DPO) [rafailov_direct_2024] directly leverages ranked sequence pairs encoding experimental preference, dispensing with reward modeling and providing robust preference alignment [widatalla_aligning_2024, stocco_guiding_2025].
• Group Relative Policy Optimization (GRPO) [guo2025deepseek-r1_2025] enables groupwise normalization for stability and sample efficiency.

Empirically, PPO- and DPO-type RL fine-tuning enables targeting extreme and combinatorially rare design optima, e.g., nanomolar binders or epitope-masked immune-invisibile scaffolds [gasser_tuning_2025, stocco_guiding_2025]. Notably, RL-augmented GMPDs are now routinely used to align model-generated samples with both in silico oracles and empirical fitness measurements [blalock_functional_2025, wang2025proteinzero, park2024improving].

Inference-Time Control

Inference-time control enables rapid adaptation without expensive retraining or risk of model drift. The assumption is that the pretrained GMPD provides a sufficiently rich base distribution for target property emergence.

Key modalities:

• Prompt/context conditioning: Conditional generation using function tags (enzyme commission, taxon, functional keywords) has proved effective for ZymCTRL [munsamy_conditional_2024], ProGen [madani2023large], Evo [evo1, brixi2025genome], and adapter-based approaches [yang_function-guided_2025]. Masked sequence reconstruction is used to explore local sequence neighborhoods or to constrain key residues.
• Retrieval-Augmented Generation (RAG): External retrieval modules dynamically augment generative context at inference—e.g., Protriever retrieves homologs to inform functional sampling in vectorized libraries [weitzman_protriever_2025].
• Gradient/Output Guidance: Techniques such as PPLM and ColabDesign/BindCraft inject gradients from output oracles or property predictors into the sampling process, biasing toward high-fitness regions even with black-box models [pacesa_one-shot_2025, cho_boltzdesign1_2025].
• Activation Steering: Latent state vector manipulation (e.g., via sparse autoencoders) enables precise control over activation trajectories, enabling "surgical" steering of, e.g., fold type, hydrophobicity, or function [adams_mechanistic_2025, garcia_interpreting_2025, corominas2025sparse].
• Bayesian Guidance: Output distributions are re-weighted using external, typically Bayesian, scoring functions—enabling rigorous property-guided sampling across model classes [xiong2025guide].
• Sampling Controls: Altering top-k/top-p sampling or using deterministic search strategies (MCTS/beam search) to navigate multi-modal sequence landscapes [ferruz2022protgpt2, brixi2025genome].

Recent theoretical advances show the equivalence of flow matching, masked language modeling, and diffusion schemes under a common discrete-state control framework [xiong2025guide]—further empowering portable, inference-time intervention across architectures.

Practical and Theoretical Implications

The reviewed techniques unlock the capacity to design proteins outside the evolutionary feasible regions, a long-standing challenge in protein engineering. RL-based alignment (especially with DPO and GRPO), RAG, and context steering now allow navigation and sampling of sequence/function regions not traversed by evolution, opening the door for synthetic scaffolds, non-natural enzymes, programmable binders, and immune-orthogonal biomolecules.

However, two broad limitations persist:

1. Generalization/OOD Failure: Both steering and fine-tuning presuppose that the pretrained model covers the functional region of interest. If not, more aggressive intervention (physics-based scoring [baek2021accurate, noe_boltzmann_2019, schymkowitz2005foldx], combinatorial mutagenesis, or vastly expanded training data [vince_breaking_2025]) are required.
2. Dataset/Annotation Biases: Supervised and RL methods are bottlenecked not only by the quantity but the diversity and representativeness of functional protein data [gordon2024protein, ding2024protein]. Recent advances in high-throughput experimentation and synthetic library screening promise to raise the ceiling on efficient model alignment, especially when paired with self-driving labs [rapp_self-driving_2024, qian_accelerating_2025].

Speculation and Future Directions

Methodological unification, architecture-agnostic intervention, and integration of physical and evolutionary priors will converge, enabling robust generalization beyond currently curated protein space. The field is poised to shift toward feedback-rich, closed-loop design cycles where computation, wet-lab data, RL alignment, and rapid validation are deeply intertwined—ultimately transferring the lessons of generative modeling from NLP to programmable biology.

Scaling models further without addressing the underlying data and inductive bias issues (e.g., with more diverse phylogenetic sampling or physics-informed rewards) will yield diminishing returns [hou2025understanding, pugh_likelihood_2025]. Flexible, hybrid approaches that combine inference-time steering, RL alignment, and experimental loop closure will likely dominate the next generation of protein generative design.

Conclusion

This review rigorously delineates and compares model- and inference-time interventions for conditional protein generation. It demonstrates that while current methods provide powerful levers for moving beyond the natural protein sequence–function distribution, significant theoretical and practical challenges remain in robustly bridging the alignment gap to engineer novel proteins with precise, user-specified properties. Future success will depend on systematic advances in data diversity, multi-modal alignment, and integration with empirical feedback, setting the stage for generalizable and controllable protein design.