Henry Dieckhaus

Recent advances in machine learning (ML)-based protein design methods have enabled the rapid in silico generation of large libraries of miniprotein binders with minimal manual input. While computational design capacity has scaled rapidly, experimental validation methods have lagged, creating a bottleneck in binder discovery pipelines. Here, we apply mRNA display to screen an ML-designed miniprotein binder library and directly compare its performance with the more widely used yeast surface display platform using a single shared DNA library. We screened 2,009 designs targeting the platelet receptor TLT-1 and 3,159 designs targeting the immune receptor B7-H3 across both platforms. While both selection methods reliably identified functional binders, we found that mRNA display preferentially enriched binders with slower dissociation rates. In addition, mRNA display achieved higher library coverage than yeast display, likely rescuing functional designs that are penalized in a cell-based expression system. Biophysical characterization of selected binders from both platforms revealed strong binding affinities and high thermal stabilities. These results showcase the power of integrating ML-based computational design tools with rapid in vitro selection technologies, providing a scalable framework for therapeutic miniprotein discovery. IMPORTANCE: Miniprotein binders offer major advantages as next-generation therapeutics, including small size, high stability, and efficient production. In this work, we conduct a side-by-side comparison of mRNA and yeast display as platforms for high-throughput evaluation of de novo miniprotein binders. The binders generated here serve as starting points for therapeutics targeting TLT-1 or B7-H3, two clinically relevant molecules.

Translation of symbols in one chemical language into another defined genetics. Yet, the co-linearity of codons and amino acids is so commonplace an idea that few even ask how it arose. Readout is done by two distinct sets of proteins, called aminoacyl-tRNA synthetases (AARS). AARS must enforce the rules first used to assemble themselves. To understand the roots of translation we must experimentally test the structural codes that the earliest AARS•tRNA cognate pairs used to recognize both amino acid and RNA substrates. We present here new results on five different facets of that problem. (i) The surfaces of structures coded by opposite strands of the same gene have opposite polarities. Core residues in proteins from one strand are surface residues in proteins from the other strand. The complementarity of base pairing thus projects into the proteome. That leads in turn to contrasting amino acid and RNA substrate binding modes. (ii) E. coli reproduces in vivo a nested hierarchy of active excerpts, or “urzymes”, similar to those we had designed as models for ancestral AARS. (iii) A third novel deletion produced in vivo and a new Class II urzyme suggests how to design bidirectional urzyme genes. (iv) Codon middle-base pairing provides a basis to constrain Class I and II AARS family trees. (v) AARS urzymes acylate Class-specific subsets of an RNA library, showing RNA substrate specificity for the first time. Four new tree-building tools augment these results to compose a viable platform for experimental study of the origins of genetic coding.

Protein design plays a key role in our efforts to work out how genetic coding began. That effort entails urzymes. Urzymes are small, conserved excerpts from full-length aminoacyl-tRNA synthetases that remain active. Urzymes require design to connect disjoint pieces and repair naked nonpolar patches created by removing large domains. Rosetta allowed us to create the first urzymes, but those urzymes were only sparingly soluble. We could measure activity, but it was hard to concentrate those samples to levels required for structural biology. Here, we used the deep learning algorithms ProteinMPNN and AlphaFold2 to redesign a set of optimized LeuAC urzymes derived from leucyl-tRNA synthetase. We select a balanced, representative subset of eight variants for testing using principal component analysis. Most tested variants are much more soluble than the original LeuAC. They also span a range of catalytic proficiency and amino acid specificity. The data enable detailed statistical analyses of the sources of both solubility and specificity. In that way, we show how to begin to unwrap the elements of protein chemistry that were hidden within the neural networks. Deep learning networks have thus helped us surmount several vexing obstacles to further investigations into the nature of ancestral proteins. Finally, we discuss how the eight variants might resemble a sample drawn from a population similar to one subject to natural selection.

The reflexive translation of symbols in one chemical language to another defined genetics. Yet, the co-linearity of codons and amino acids is so commonplace an idea that few even ask how it arose. Readout is done by two distinct sets of proteins, called aminoacyl-tRNA synthetases (AARS). AARS must enforce the rules first used to assemble themselves. The roots of translation lie in experimentally testing the structural codes that the earliest AARS•tRNA cognate pairs used to recognize both amino acid and RNA substrates. We review here new results on five different facets of that problem. (i) The surfaces of structures coded by opposite strands of the same gene have opposite polarities. The corresponding proteins then fold up “inside out” relative to one another. The inversion symmetry of base pairing thus projects into the proteome. That leads in turn to contrasting amino acid and RNA substrate binding modes. (ii) E. coli reproduces in vivo the nested hierarchy of active excerpts we had designed as models—protozymes and urzymes—for ancestral AARS. (iii) A third novel deletion produced in vivo and a new Class II urzyme suggest how to design bidirectional urzyme genes. (iv) Codon middle-base pairing provides a basis to constrain Class I and II AARS family trees. (v) AARS urzymes acylate Class-specific subsets of an RNA library, showing RNA substrate specificity for the first time. Four new phylogenetic routines augment these results to compose a viable platform for experimental study of the origins of genetic coding. Significance Statement The origin of genetic coding poses questions distinct from those faced studying the evolution of enzymes since the first cells. Modern enzymes that translate the code range in size from ∼330 to ∼970 amino acids. Ancestral forms cannot have been nearly as complex. Moreover, such primitive enzymes likely could enforce only a much-reduced coding alphabet. Structural and molecular biology data point to a broad sketch of events leading to the code. That research platform will enable us to see how Nature came to store information about the physical chemistry of amino acids in the coding table. That, in turn, allowed searching of a very broad amino acid sequence space. Selection could then learn how to assemble amino acids into functional, reflexive catalysts. Those catalysts had rates and fidelities consistent with bootstrapping the modern coding alphabet. New phylogenetic algorithms need to be developed to fully test that putative sketch experimentally.

Many of the biosynthetic pathways for ribosomal synthesized and post-translationally modified peptide (RiPP) natural products make use of multi-domain enzymes with separate recruitment and catalysis domains that separately bind and modify peptide substrates. This “division of labor” allows RiPP enzymes to use relatively open and promiscuous active sites to perform chemistry at multiple residues within a peptide substrate seemingly regardless of the surrounding context. Defining, measuring, and predicting the seemingly broad substrate promiscuity of RiPPs necessitates high throughput assays, capable of assessing activity against very large libraries of peptides. Using mRNA display, a high throughput peptide display technology, we examine the substrate promiscuity of the RiPP cyclodehydratase, LynD. The vast substrate profiling that can be done with mRNA display enables the construction of deep learning models for accurate prediction of substrate processing by LynD. These models further inform on epistatic interactions involved in enzymatic processing. This work will facilitate the further elucidation of other RiPP enzymes and enable their use in the modification of mRNA display libraries for selection of modified peptide-based inhibitors and therapeutics.

There is strong interest in accurate methods for predicting changes in protein stability resulting from amino acid mutations to the protein sequence. Recombinant proteins must often be stabilized to be used as therapeutics or reagents, and destabilizing mutations are implicated in a variety of diseases. Due to increased data availability and improved modeling techniques, recent studies have shown advancements in predicting changes in protein stability when a single point mutation is made. Less focus has been directed toward predicting changes in protein stability when there are two or more mutations, despite the significance of mutation clusters for disease pathways and protein design studies. Here, we analyze the largest available dataset of double point mutation stability and benchmark several widely used protein stability models on this and other datasets. We identify a blind spot in how predictors are typically evaluated on multiple mutations, finding that, contrary to assumptions in the field, current stability models are unable to consistently capture epistatic interactions between double mutations. We observe one notable deviation from this trend, which is that epistasis-aware models provide marginally better predictions on stabilizing double point mutations. We develop an extension of the ThermoMPNN framework for double mutant modeling as well as a novel data augmentation scheme which mitigates some of the limitations in available datasets. Collectively, our findings indicate that current protein stability models fail to capture the nuanced epistatic interactions between concurrent mutations due to several factors, including training dataset limitations and insufficient model sensitivity. Significance Protein stability is governed in part by epistatic interactions between energetically coupled residues. Prediction of these couplings represents the next frontier in protein stability modeling. In this work, we benchmark protein stability models on a large dataset of double point mutations and identify previously overlooked limitations in model design and evaluation. We also introduce several new strategies to improve modeling of epistatic couplings between protein point mutations.

Amino acid mutations that lower a protein’s thermodynamic stability are implicated in numerous diseases, and engineered proteins with enhanced stability can be important in research and medicine. Computational methods for predicting how mutations perturb protein stability are, therefore, of great interest. Despite recent advancements in protein design using deep learning, in silico prediction of stability changes has remained challenging, in part due to a lack of large, high-quality training datasets for model development. Here, we describe ThermoMPNN, a deep neural network trained to predict stability changes for protein point mutations given an initial structure. In doing so, we demonstrate the utility of a recently released megascale stability dataset for training a robust stability model. We also employ transfer learning to leverage a second, larger dataset by using learned features extracted from ProteinMPNN, a deep neural network trained to predict a protein’s amino acid sequence given its three-dimensional structure. We show that our method achieves state-of-the-art performance on established benchmark datasets using a lightweight model architecture that allows for rapid, scalable predictions. Finally, we make ThermoMPNN readily available as a tool for stability prediction and design.

Allostery is a major driver of biological processes requiring coordination. Thus, it is one of the most fundamental and remarkable phenomena in nature, and there is motivation to understand and manipulate it to a multitude of ends. Today, it is often described in terms of two phenomenological models proposed more than a half-century ago involving only T(tense) or R(relaxed) conformations. Here, methyl-based NMR provides extensive detail on a dynamic T to R switch in the classical dimeric allosteric protein, yeast chorismate mutase (CM), that occurs in the absence of substrate, but only with the activator bound. Switching of individual subunits is uncoupled based on direct observation of mixed TR states in the dimer. This unique finding excludes both classic models and solves the paradox of a coexisting hyperbolic binding curve and highly skewed substrate-free T–R equilibrium. Surprisingly, structures of the activator-bound and effector-free forms of CM appear the same by NMR, providing another example of the need to account for dynamic ensembles. The apo enzyme, which has a sigmoidal activity profile, is shown to switch, not to R, but to a related high-energy state. Thus, the conformational repertoire of CM does not just change as a matter of degree depending on the allosteric input, be it effector and/or substrate. Rather, the allosteric model appears to completely change in different contexts, which is only consistent with modern ensemble–based frameworks.

Amino acid mutations that lower a protein’s thermodynamic stability are implicated in numerous diseases, and engineered proteins with enhanced stability are important in research and medicine. Computational methods for predicting how mutations perturb protein stability are therefore of great interest. Despite recent advancements in protein design using deep learning, in silico prediction of stability changes has remained challenging, in part due to a lack of large, high-quality training datasets for model development. Here we introduce ThermoMPNN, a deep neural network trained to predict stability changes for protein point mutations given an initial structure. In doing so, we demonstrate the utility of a newly released mega-scale stability dataset for training a robust stability model. We also employ transfer learning to leverage a second, larger dataset by using learned features extracted from a deep neural network trained to predict a protein’s amino acid sequence given its three-dimensional structure. We show that our method achieves competitive performance on established benchmark datasets using a lightweight model architecture that allows for rapid, scalable predictions. Finally, we make ThermoMPNN readily available as a tool for stability prediction and design.

Objectives: Automated whole brain segmentation from magnetic resonance images is of great interest for the development of clinically relevant volumetric markers for various neurological diseases. Although deep learning methods have demonstrated remarkable potential in this area, they may perform poorly in nonoptimal conditions, such as limited training data availability. Manual whole brain segmentation is an incredibly tedious process, so minimizing the data set size required for training segmentation algorithms may be of wide interest. The purpose of this study was to compare the performance of the prototypical deep learning segmentation architecture (U-Net) with a previously published atlas-free traditional machine learning method, Classification using Derivative-based Features (C-DEF) for whole brain segmentation, in the setting of limited training data. Materials and Methods: C-DEF and U-Net models were evaluated after training on manually curated data from 5, 10, and 15 participants in 2 research cohorts: (1) people living with clinically diagnosed HIV infection and (2) relapsing–remitting multiple sclerosis, each acquired at separate institutions, and between 5 and 295 participants' data using a large, publicly available, and annotated data set of glioblastoma and lower grade glioma (brain tumor segmentation). Statistics was performed on the Dice similarity coefficient using repeated-measures analysis of variance and Dunnett–Hsu pairwise comparison. Results: C-DEF produced better segmentation than U-Net in lesion (29.2%–38.9%) and cerebrospinal fluid (5.3%–11.9%) classes when trained with data from 15 or fewer participants. Unlike C-DEF, U-Net showed significant improvement when increasing the size of the training data (24%–30% higher than baseline). In the brain tumor segmentation data set, C-DEF produced equivalent or better segmentations than U-Net for enhancing tumor and peritumoral edema regions across all training data sizes explored. However, U-Net was more effective than C-DEF for segmentation of necrotic/non-enhancing tumor when trained on 10 or more participants, probably because of the inconsistent signal intensity of the tissue class. Conclusions: These results demonstrate that classical machine learning methods can produce more accurate brain segmentation than the far more complex deep learning methods when only small or moderate amounts of training data are available (n ≤ 15). The magnitude of this advantage varies by tissue and cohort, while U-Net may be preferable for deep gray matter and necrotic/non-enhancing tumor segmentation, particularly with larger training data sets (n ≥ 20). Given that segmentation models often need to be retrained for application to novel imaging protocols or pathology, the bottleneck associated with large-scale manual annotation could be avoided with classical machine learning algorithms, such as C-DEF.