ProteinGuide: On-the-fly property guidance for protein sequence generative models

TL;DR

ProteinGuide introduces an on-the-fly conditioning framework that blends a pre-trained protein sequence prior with assay-informed predictive models to steer generation without retraining. It unifies four discrete-space model classes—DDMs, DFMs, MLMs, and AO-ARMs—under a Bayes-rule-based guidance formalism, and provides exact (DEG) and fast approximate (TAG) sampling algorithms. The authors demonstrate interpolative, extrapolative, and multi-property Pareto-extrapolative designs across in silico tasks and validate in vivo by enhancing adenine base editor activity with a single guided round, surpassing seven rounds of directed evolution. The work shows that principled, on-the-fly guidance can integrate experimental feedback efficiently, enabling iterative, design-build-test-learn cycles across diverse protein design problems without retraining complex models. Overall, ProteinGuide broadens practical access to principled multi-property design by leveraging predictive models within existing pre-trained generators, accelerating protein engineering workflows.

Abstract

Sequence generative models are transforming protein engineering. However, no principled framework exists for conditioning these models on auxiliary information, such as experimental data, without additional training of a generative model. Herein, we present ProteinGuide, a method for such "on-the-fly" conditioning, amenable to a broad class of protein generative models including Masked Language Models (e.g. ESM3), any-order auto-regressive models (e.g. ProteinMPNN) as well as diffusion and flow matching models (e.g. MultiFlow). ProteinGuide stems from our unifying view of these model classes under a single statistical framework. As proof of principle, we perform several in silico experiments. We first guide pre-trained generative models to design proteins with user-specified properties, such as higher stability or activity. Next, we design for optimizing two desired properties that are in tension with each other. Finally, we apply our method in the wet lab, using ProteinGuide to increase the editing activity of an adenine base editor in vivo with data from only a single pooled library of 2,000 variants. We find that a single round of ProteinGuide achieves a higher editing efficiency than was previously achieved using seven rounds of directed evolution.

Figures (21)

• Figure 1: Overview of ProteinGuide.a) Contrasting unguided generation with guided generation by way of ProteinGuide. A sequence is denoted, $x$, the pre-trained generative model is denoted $p_\theta(x)$; $y$ denotes a property of interest that we seek to guide toward using predictive model $p_\phi(y \mid x)$. When guiding, we blend information from the predictive model with that of the pre-trained model in a statistically correct manner, resulting in sequences sampled from the property-informed, reweighted distribution, $p_{\theta, \phi}(x \mid y)$. Masking of one sequence position is denoted with the symbol "?", and pink letters indicate key mutations that are introduced only when guiding with the predictive model, rather than when using unguided generation. b) Unifying view of training and sampling for masked diffusion/flow models, MLMs, and AO-ARMs enables guidance on a broad set of model classes. Colored dots visually tag individual sequences, each of which has a different amount of masked tokens. The sequence being masked during training is M K T R S V. The shared training objective involves randomly masking out parts of the sequence and then predicting the masked sequence given the remaining unmasked context over different masking rates. Once any type of model is trained, one can then use a sampling algorithm originally developed for any of the other models. Consequently, we can obtain computationally tractable exact sampling for continuous time models and also accelerated approximate sampling. c) Schematic contrasting ProteinGuide with post hoc filtering and fine-tuning. ProteinGuide combines a pre-trained generative prior with assay-labeled data through correct statistical conditioning, enabling more reliable enrichment for high-property sequences than the other approaches shown.
• Figure 2: Guiding ProteinMPNN with experimental stability data to generate sequences with enhanced stability.a) Schematics for guiding ProteinMPNN with large-scale experimental stability measurements. ProteinMPNN, an inverse folding model, is guided by a stability predictive model trained on the large-scale stability dataset from tsuboyama2023mega to generate sequences with stability better than the wild-type sequence. b) Five methods for sampling (in legends) are assessed across eight proteins (horizontal axis) by their success rate on 100 generated samples. The height of each bar represents the percentage of sequences predicted to be at least as stable as wild type ($\Delta \Delta G \leq 0$), while the hatched sub-portion indicates the fraction of those stable sequences that are also predicted to correctly fold into the desired structure (RMSD $\leq 2$Å). In most cases the hatched sub-portion covers the full height of the bar. Error bars represent standard errors of proportions, calculated as $\sqrt{p(1-p)/n}$, where $p$ is the empirical proportion of success and $n$ is the sample size. c) Distributions of refolded RMSD for the generated sequences (lower is better). The dashed horizontal line indicates the threshold below which a sequence is considered to fold into the desired structure (RMSD $\leq 2$Å). d) Distributions of oracle predicted stability for the generated sequences (lower is better). The dashed horizontal line indicates the threshold below which a sequence is predicted to be at least as stable as the wild-type ($\Delta \Delta G \leq 0$).
• Figure 3: ProteinGuide guides ESM3 with assay-informed predictive models for extrapolative design.a) Experimental set-up for the extrapolative setting, where the goal is to generate sequences with higher fitness compared to those in the labeled dataset. The set of sequences (around 100) with highest property values in the labeled dataset were excluded for both fine-tuning and predictive model training. Fine-tuning is performed on sequences on $M$ top variants from the assay-labeled data set. $M$ is chosen for each dataset scanning through different % threshold on top-scoring variants as a hyperparameter, and choosing the one that had highest likelihood for held-out high-scoring variants. The predictive model used in guidance is trained on 2,000 sequences randomly sampled from the labeled dataset. b) For each dataset (subpanel title), each method is scored by its success rate on 100 samples. The height of each bar represents the percentage of sequences predicted by the oracle model to be above a particular threshold based on the quantile of the training set (horizontal axis) and are not identical to any sequence in the labeled set (i.e., novelty $> 0$), while the hatched sub-portion indicates the fraction of those sequences that are also predicted by ESMFold lin2023evolutionary to fold into a structure with pLDDT $\geq$ 0.8 (Supplementary Figure \ref{['si_fig:fitness_structure']}). Error bars represent standard errors of proportions, calculated as $\sqrt{p(1-p)/n}$, where $p$ is the empirical proportion of success and $n$ is the sample size. c) Diversity (left) and novelty (right) of the generated sequences. Higher diversity and novelty is desirable provided that the sequences also have high property value. Diversity is computed as the averaged pairwise hamming distance among the generated sequences. "Labeled" diversity is computed on 500 randomly sampled sequences from the labeled dataset for each protein. For each sequence, novelty is computed as its minimum hamming distance with sequences in the labeled dataset. The novelty of FT generated sequences is very low novelty and is not readily discernible. d) Distributions of oracle predicted property value for the generated sequences. "Labeled" consists of all sequences in the labeled dataset.
• Figure 4: Multi-property guidance with ProteinGuide on PbrR to improve lead (Pb) binding and reduce off-target zinc (Zn) binding.a) Schematic diagrams of the data splits used for training and evaluating models for multi-property guidance on PbrR. b) Pareto-extrapolation task: scatterplot of oracle-predicted log-fold change for model generations for ESM C (Unguided), FT, and ProteinGuide. Experimental values of the initial library of variants from wang2025active are also shown in gray. The target region requires a Pb log-FC of 2 compared to the wild-type and a Zn log-FC of -0.6 compared to wild-type. c) Left panel shows the proportion of sequences that were in the target region for each of ESM C (Unguided), FT, and ProteinGuide. Right panel shows the average Hamming distance between sequences generated by each model, as a measure of diversity. d) Controllable generation task: the top left panel shows outliers and the region corresponding to the remaining four plots (in yellow shading). Top middle panel shows each of three target regions for the Pareto sampling task as quadrants delineated by colored dashed lines. Each target region has it's bottom left vertex anchored on the approximate Pareto frontier. The remaining three plots (top right and bottom row) show sequences generated by guidance for each target region using the classifier thresholds in the subpanel titles. Subpanel titles also show success rate in parentheses.
• Figure 5: Engineering adenine base editor with ProteinGuide.a) Schematic representation of the ProteinGuide ABE experiment. b) Scatterplots of the number of mutations from TadA ($x$-axis) vs. the predicted probability of being active ($y$-axis) for every variant in library 1 (teal) and library 2 (pink) of FMIF (left) and ESM3 (right). c) A histogram of the experimentally measured log enrichments for both library 1 (teal) and library 2 (pink). ProteinGuide was used to generate the round 2 library such that variants would be enriched for having high activity (predicted activity in log enrichment units higher than the gray dashed line). d) Similar to b), but with experimentally measured log enrichments shown on the vertical axis. e) Sequence logo plots of the differences in the position-wise amino acids frequencies between library 1 and library 2 for FMIF (left) and ESM3 (right). Letters above the horizontal line at $y=0$ represent position/amino acids that are enriched in library 2 relative to library 1, and those below 0 represent positions/amino acids that are depleted in library 2 relative to library 1. Black letters above the plots show mutations obtained in the experimental evolution campaigns leading to ABE8. f) Comparison of the activity of two of the top variants after ProteinGuide with experimentally obtained variants with known activities (ABE1.1, ABE7.10, ABE8). Left: activity estimated based on the titers of individual transformations in the selection strain, shown are colony forming units (cfu) on selective normalized to non-selective condition. Right: activity calculated based on log-enrichment in a pooled selection assay measured by sequencing. Error bars represent standard deviations over three independent transformations for each experiment.

Theorems & Definitions (9)

• Proposition 1
• Corollary 1
• proof : Proof Sketch
• proof : Proof of Proposition \ref{['prop:main']}
• proof : Proof of Corollary \ref{['corollary:sample-permutation']}
• Lemma 1
• proof
• Lemma 2
• proof