Negative selection in the thymus limits autoimmunity by eliminating T cells that react strongly to self. Individual T cells, however, are only exposed to a small fraction of all self-peptides during their “training” in the thymus, and how tolerance is generalized to the remaining “test” self-peptides across peripheral tissues in the body remains an open question. We show that this can be achieved because the immune system satisfies two conditions necessary for generalization in machine learning settings. Consequently, sparse, random sampling of only 10% of self-peptides in the thymus is sufficient to avoid reactivity to 90% of peripheral self. We support this result and validate predictions from our model with diverse experimental data. Overall, we provide a plausible answer to a long-standing question underlying adaptive immunity, and we highlight how generalization, a fundamental challenge faced by nearly every learning algorithm, is tackled by the immune system.

We introduce AutoTuneX, a data-driven, AI system designed to automatically predict optimal parameters for transcript assemblers - tools for reconstructing transcripts from the reads in a given RNA-seq sample. AutoTuneX is built by learning parameter knowledge from existing RNA-seq samples and transferring this knowledge to unseen samples. On 1588 human RNA-seq samples tested with two transcript assemblers, AutoTuneX predicts parameters that resulted in 98% of samples achieving more accurate transcript assembly compared to using default parameters, with some samples experiencing up to a 600% improvement in AUC. AutoTuneX offers a new strategy for automatically optimizing use of sequence analysis tools.

We study Turnpike with uncertain measurements: reconstructing a one-dimensional point set from an unlabeled multiset of pairwise distances under bounded noise and rounding. We give a combinatorial characterization of realizability via a multi-matching that labels interval indices by distinct distance values while satisfying all triangle equalities. This yields an ILP based on the triangle equality whose constraint structure depends only on the two-partition set 𝒫_y={(r,s,t): y_r+y_s=y_t} and a natural LP relaxation with {0,1}-coefficient constraints. Integral solutions certify realizability and output an explicit assignment matrix, enabling an assignment-first, regression-second pipeline for downstream coordinate estimation. Under bounded noise followed by rounding, we prove a deterministic separation condition under which 𝒫_y is recovered exactly, so the ILP/LP receives the same combinatorial input as in the noiseless case. Experiments illustrate integrality behavior and degradation outside the provable regime.

Optimizing synonymous codon sequences to improve translation efficiency, RNA stability, and compositional properties is challenging because the search space grows exponentially with protein length and objectives interact through long range RNA structure. Dynamic programming-based methods can provide strong solutions for fixed objective combinations but are difficult to extend to additional constraints. Deep generative models require large-scale, high-quality mRNA sequence datasets for training, limiting applicability when such data are scarce. Reinforcement learning naturally handles sequential decision-making but faces challenges in codon optimization due to delayed rewards, large action spaces, and expensive structural evaluation. We present CodonRL, a reinforcement learning framework that learns a structural prior for mRNA design from efficient folding feedback and demonstration-guided replay, and then enables user-controlled multi-objective trade-offs during inference. CodonRL uses LinearFold for fast intermediate reward computation during training and ViennaRNA for final evaluation, warms up learning with expert sequences to accelerate convergence for global structure objectives, and introduces milestone-based intermediate rewards to address delayed feedback in long range optimization. On a benchmark of 55 human proteins, CodonRL outperforms GEMORNA, a state-of-the-art codon optimization method, across multiple metrics, achieving 9.5% higher codon adaptation index (CAI), 25.4 kcal/mol more favorable minimum free energy (MFE), and 3.4% lower uridine content on average, while improving codon stabilization coefficient (CSC) in over 90% of benchmark proteins under matched constraints. These gains translate into designs that are predicted to be more efficiently translated, more structurally stable, and less immunogenic, while supporting continuous objective reweighting at inference time.

Motivation Genomic language models (gLMs) face a fundamental efficiency challenge: one must either maintain separate specialized models for each biological modality (DNA and RNA) or develop large multimodal architectures. Both approaches impose significant computational burdens-modality-specific models require redundant infrastructure despite inherent biological connections, while multi-modal architectures demand increased parameter counts and extensive cross-modality pretraining.Results To address this limitation, we introduce CodonMoE (Adaptive Mixture of Codon Reformative Experts), a lightweight adapter that transforms DNA language models into effective RNA analyzers without RNA-specific pretraining. Our theoretical analysis establishes CodonMoE as a universal approximator at the codon level, capable of mapping arbitrary functions from codon sequences to codon-dependent RNA properties given sufficient expert capacity. Across four RNA prediction tasks spanning stability, expression, and regulation, DNA models augmented with CodonMoE significantly outperform their unmodified counterparts, with the HyenaDNA+CodonMoE series achieving state-of-the-art results using 80% fewer parameters than specialized RNA models. By maintaining sub-quadratic complexity while achieving superior performance, our approach provides a principled path toward unifying genomic language modeling, leveraging more abundant DNA data and reducing computational overhead while preserving modality-specific performance advantages.Availability and implementation Source code for the method and to reproduce the results is available at https://github.com/Kingsford-Group/CodonMoE.

Motivation Ribosome dynamics are vital in the process of protein expression. Current methods rely on ribosome profiling (Ribo-seq), RNA-seq profiles, and full genomic context. This restricts their use in de novo sequence design, like messenger RNA (mRNA) vaccines. Simulation-only approaches like the Totally Asymmetric Simple Exclusion Process (TASEP) oversimplify translation by focusing solely on codon elongation times.Results We present seq2ribo, a hybrid simulation and machine learning framework that predicts ribosome A-site locations using only an mRNA sequence as input. Our method first employs a novel structure-aware TASEP (sTASEP), which models translation using a comprehensive set of fitted parameters that include codon wait times and structural features, such as local angles, base-pairing, and discrete positional buckets. The ribosome locations generated by sTASEP are then processed by a polisher model, which learns to refine the simulated ribosome distributions. seq2ribo provides high-fidelity predictions of ribosome locations across diverse cell types (iPSC, HEK293, LCL, and RPE-1), significantly outperforming baselines. seq2ribo is the first method to achieve meaningful positional correlation with observed ribosome profiles from sequence alone, reaching transcript-level Pearson correlations up to 0.920 and within-transcript shape correlations up to 0.186, where all baselines yield near-zero values on these metrics. seq2ribo also reduces elementwise error by up to 37.7% relative to the sequence-only Translatomer baseline. By adding a task-specific head, seq2ribo achieves Pearson correlations up to 0.732 with experimental translation efficiency (TE) across several cell lines, and up to 0.903 with measured protein expression. By operating from sequence alone, seq2ribo provides a new tool for synthetic biology, enabling the rational design and optimization of mRNA sequences without the need for expression-level data or genomic context.Availability seq2ribo is available at https://github.com/Kingsford-Group/seq2ribo.

Genomic language models (gLMs) face a fundamental efficiency challenge: either maintain separate specialized models for each biological modality (DNA and RNA) or develop large multi-modal architectures. Both approaches impose significant computational burdens - modality-specific models require redundant infrastructure despite inherent biological connections, while multi-modal architectures demand massive parameter counts and extensive cross-modality pretraining. To address this limitation, we introduce CodonMoE (Adaptive Mixture of Codon Reformative Experts), a lightweight adapter that transforms DNA language models into effective RNA analyzers without RNA-specific pretraining. Our theoretical analysis establishes CodonMoE as a universal approximator at the codon level, capable of mapping arbitrary functions from codon sequences to RNA properties given sufficient expert capacity. Across four RNA prediction tasks spanning stability, expression, and regulation, DNA models augmented with CodonMoE significantly outperform their unmodified counterparts, with HyenaDNA+CodonMoE series achieving state-of-the-art results using 80% fewer parameters than specialized RNA models. By maintaining sub-quadratic complexity while achieving superior performance, our approach provides a principled path toward unifying genomic language modeling, leveraging more abundant DNA data and reducing computational overhead while preserving modality-specific performance advantages.

A multigraph is uniformly k-edge-connected if there are exactly k edge-disjoint paths between any pair of vertices. For example, a uniformly k-edge-connected graph is obtained from a k-edge-connected graph by collapsing the nodes connected by more than k edge-disjoint paths into supernodes. We characterize the class of uniformly 3edge-connected graphs, giving a synthesis involving two operations by which every uniformly 3-edge-connected multigraph can be generated. Slightly modified syntheses give the planar uniformly 3-edge-connected graphs and the uniformly 3-edge-connected graphs with the fewest possible edges, generalizing the well-known Harary graphs. In proving the correctness of the synthesis, we also show the existence of a particular type of induced, non-separating cycle in near 3-regular graphs, which is of interest in its own right. (c) 2025 Published by Elsevier B.V.

In computational chemistry, molecular docking—predicting the binding structure of a small molecule ligand to a protein—is vital for understanding interactions between small molecules and their protein targets, with broad applications in drug discovery ([Morris and Lim-Wilby, 2008][1]). Traditional docking methods rely on energy-based scoring functions and optimization algorithms to identify ligand-binding structures. However, these methods are often slow and inaccurate due to the extensive search space for binding structures and the complexity of scoring function landscapes ([Corso et al., 2023][2]). Deep learning techniques have emerged as promising alternatives for molecular docking. These methods harness the power of neural networks to understand the complex interactions between protein pockets and ligands from large datasets. They can be classified into two categories: regression-based methods (Stark et al., 2022; [Lu et al., 2022][3]), which offer greater computational efficiency compared to traditional docking methods but have yet to offer substantial improvements in accuracy ([Corso et al., 2023][2]); and generative models, particularly diffusion models ([Corso et al., 2023][2]; [Qiao et al., 2024][4]; [Nakata et al., 2023][5]; [Lu et al., 2024][6]; [Schneuing et al., 2024][7]), which lead to significant accuracy improvements compared to regression-based methods. Most diffusion-based docking methods aim to address the blind docking problem, where the ligand is docked without prior knowledge of the specific protein pocket—a site of the protein with potential ligand-binding capabilities—and only the overall protein structure is provided ([Yim et al., 2024][8]). However, in many drug discovery cases, specific pockets have already been identified ([Zheng et al., 2013][9]), meaning that the docking process only needs to determine how the ligand binds to the pocket. This defines the pocket-based docking problem. Pocket-based docking leverages the physical and chemical information of protein side chains around the pocket, significantly reducing the search space for ligand binding structures and improving the understanding of atomic-level interactions. Additionally, focusing on the local protein substructure rather than the global structure enhances computational efficiency, which allows for the use of more complex neural architectures. We introduce DTMol, a novel diffusion model designed to tackle the pocket-based docking problem. Our model integrates a pretrained molecular representation framework with a diffusion transformer architecture. The advantages of this design are twofold: first, the molecular representation models we employ are pretrained on separate datasets of small molecules and protein pockets, which are larger than the datasets of ligand-pocket interactions. Consequently, these pretrained models encode typical structural features of both elements, enhancing our model’s ability to represent them and predict binding structures. Second, the transformer architecture used in our diffusion model is more capable of effectively capturing all-atom interactions between the pocket and the ligand. We test our method on the docking benchmark PDBBind v2020 and PoseBuster, and compare it with other methods. DTMol achieves 77.65% top-1 success rate (RMSD ¡ 2Å) on PoseBuster, outperforms the second-best docking software Gnina (65.65%) We also test the performance of our method on a real-world virtual screening task using Janus kinase 2 (PDB ID: 6BBV) as the target protein. A TR-FRET screening experiment is performed on the union of top ligands identified via physics-based docking and the diffusion model, in order to measure their inhibitory activity against the JAK2 protein. The results shows that our model achieved the only positive rank correlation score compared to other traditional docking methods and machine learning-based methods. To summarize, the main contributions of this work are: ### Competing Interest Statement The authors have declared no competing interest. [1]: #ref-21 [2]: #ref-5 [3]: #ref-17 [4]: #ref-26 [5]: #ref-22 [6]: #ref-18 [7]: #ref-28 [8]: #ref-39 [9]: #ref-41

1562 Background: Traditional clinicogenomic analysis workflows in oncology require substantial bioinformatics expertise and custom coding, creating a bottleneck where clinical researchers must rely on analysts for data interpretation. This dependency hinders hypothesis development, exploration and discovery, as researchers cannot directly interact with their data in real-time. DeepSeal addresses these challenges by providing a user-friendly, chat-based interface that enables immediate analysis of clinicogenomic data and seamless generation of results, empowering clinical researchers to independently explore and validate hypotheses. Methods: DeepSeal, an integration of a large language model with clinical and molecular databases and bioinformatics tools, was evaluated by replicating the findings of Riaz et al. (Cell, 2017), in advanced melanoma patients treated with nivolumab. Through natural language prompts, DeepSeal performed multiple analyses including differential gene expression analysis and Gene Set Enrichment Analysis (GSEA) to generate comprehensive molecular profiles of responders versus non-responders and to identify molecular signatures associated with treatment response. Results: DeepSeal successfully replicated the key findings of Riaz et al., identifying significant differential expression of immune-related genes in treatment responders. GSEA executed through DeepSeal’s chat interface further revealed enrichment of immune-related pathways critical for response, including B cell activation (GO:0042113), T cell activation (GO:0042110), and regulation of adaptive immune response (GO:0002819). The chat interface enabled rapid hypothesis testing and visualization generation, with analyses completed in minutes without the need for programming expertise. Conclusions: DeepSeal demonstrates the feasibility of enabling clinical researchers to independently analyze complex clinicogenomic data through natural language interaction. By successfully replicating and validating findings from Riaz et al., it generates reliable insights without programming expertise, offering a transformative approach to accelerate translational research. This removal of technical barriers between researchers and their data has the potential to substantially speed hypothesis testing and discovery in oncology, ultimately enhancing the pathway from molecular insights to improved patient care.

Summary:Modern biological research critically depends on public databases. The introduction and propagation of errors within and across databases can lead to wasted resources as scientists are led astray by bad data or have to conduct expensive validation experiments. The emergence of generative artificial intelligence systems threatens to compound this problem owing to the ease with which massive volumes of synthetic data can be generated. We provide an overview of several key issues that occur within the biological data ecosystem and make several recommendations aimed at reducing data errors and their propagation. We specifically highlight the critical importance of improved educational programs aimed at biologists and life scientists that emphasize best practices in data engineering. We also argue for increased theoretical and empirical research on data provenance, error propagation, and on understanding the impact of errors on analytic pipelines. Furthermore, we recommend enhanced funding for the stewardship and maintenance of public biological databases. Availability and implementation:Not applicable.

Codon sequence design is crucial for generating mRNA sequences with desired functional properties for tasks such as developing mRNA vaccines or gene editing therapies. Yet existing methods lack flexibility and controllability to adapt to various design objectives. We propose a novel machine learning-based framework, ARCADE, that enables flexible and controllable multi-objective codon design. Leveraging inherent knowledge from pretrained genomic language models, ARCADE extends activation engineering, a technique originally developed for controllable text generation, beyond discrete feature manipulation such as concepts and styles, to steering continuous-valued biological metrics. Specifically, we derive biologically meaningful semantic steering vectors in the model's activation space, which directly control properties such as the Codon Adaptation Index, Minimum Free Energy, and GC content. Experimental results demonstrate the flexibility of ARCADE in designing codon sequences with multiple objectives, underscoring its potential for advancing programmable biological sequence design. Our implementation is available at https://github.com/Kingsford-Group/arcade.

The Turnpike problem aims to reconstruct a set of one-dimensional points from their unordered pairwise distances. Turnpike arises in biological applications such as molecular structure determination, genomic sequencing, tandem mass spectrometry, and molecular error-correcting codes. Under noisy observation of the distances, the Turnpike problem is NP-hard and can take exponential time and space to solve when using traditional algorithms. To address this, we reframe the noisy Turnpike problem through the lens of optimization, seeking to simultaneously find the unknown point set and a permutation that maximizes similarity to the input distances. Our core contribution is a suite of algorithms that robustly solve this new objective. This includes a bilevel optimization framework that can efficiently solve Turnpike instances with up to 100,000 points. We show that this framework can be extended to scenarios with domain-specific constraints that include duplicated, missing, and partially labeled distances. Using these, we also extend our algorithms to work for points distributed on a circle (the Beltway problem). For small-scale applications that require global optimality, we formulate an integer linear program (ILP) that (i) accepts an objective from a generic family of convex functions and (ii) uses an extended formulation to reduce the number of binary variables. On synthetic and real partial digest data, our bilevel algorithms achieved state-of-the-art scalability across challenging scenarios with performance that matches or exceeds competing baselines. On small-scale instances, our ILP efficiently recovered ground-truth assignments and produced reconstructions that match or exceed our alternating algorithms. Our implementations are available at https://github.com/Kingsford-Group/turnpikesolvermm.

MOTIVATION:Sequences equivalent to their reverse complements (i.e. double-stranded DNA) have no analogue in text analysis and non-biological string algorithms. Despite this striking difference, algorithms designed for computational biology (e.g. sketching algorithms) are designed and tested in the same way as classical string algorithms. Then, as a post-processing step, these algorithms are adapted to work with genomic sequences by folding a k-mer and its reverse complement into a single sequence: The canonical representation (k-nonical space). RESULTS:The effect of using the canonical representation with sketching methods is understudied and not understood. As a first step, we use context-free sketching methods to illustrate the potentially detrimental effects of using canonical k-mers with string algorithms not designed to accommodate for them. In particular, we show that large stretches of the genome ("sketching deserts") are undersampled or entirely skipped by context-free sketching methods, effectively making these genomic regions invisible to subsequent algorithms using these sketches. We provide empirical data showing these effects and develop a theoretical framework explaining the appearance of sketching deserts. Finally, we propose two schemes to accommodate for these effects: (i) a new procedure that adapts existing sketching methods to k-nonical space and (ii) an optimization procedure to directly design new sketching methods for k-nonical space. AVAILABILITY AND IMPLEMENTATION:The code used in this analysis is available under a permissive license at https://github.com/Kingsford-Group/mdsscope.

Three-dimensional chromosome structure plays an important role in fundamental genomic functions. Hi-C, a high-throughput, sequencing-based technique, has drastically expanded our comprehension of 3D chromosome structures. The first step of Hi-C analysis pipelines involves mapping sequencing reads from Hi-C to linear reference genomes. However, the linear reference genome does not incorporate genetic variation information, which can lead to incorrect read alignments, especially when analyzing samples with substantial genomic differences from the reference such as cancer samples. Using genome graphs as the reference facilitates more accurate mapping of reads, however, new algorithms are required for inferring linear genomes from Hi-C reads mapped on genome graphs and constructing corresponding Hi-C contact matrices, which is a prerequisite for the subsequent steps of the Hi-C analysis such as identifying topologically associated domains and calling chromatin loops. We introduce the problem of genome sequence inference from Hi-C data mediated by genome graphs. We formalize this problem, show the hardness of solving this problem, and introduce a novel heuristic algorithm specifically tailored to this problem. We provide a theoretical analysis to evaluate the efficacy of our algorithm. Finally, our empirical experiments indicate that the linear genomes inferred from our method lead to the creation of improved Hi-C contact matrices, which are more effective in accurately capturing the structures of topologically associated domains.

The Beltway and Turnpike problems entail the reconstruction of circular and linear one-dimensional point sets from unordered pairwise distances. These problems arise in computational biology when the measurements provide distances but do not associate those distances with the entities that gave rise to them. Such applications include molecular structure determination, genomic sequencing, tandem mass spectrometry, and molecular error-correcting codes (since sequencing and mass spec technologies can give lengths or weights, usually without connecting them to endpoints). Practical algorithms for Turnpike are known when the distance measurements are accurate, but both problems become strongly NP-hard under any level of measurement uncertainty. This is problematic since all known applications experience some degree of uncertainty from uncontrollable factors. Traditional algorithms cope with this complexity by exploring a much larger solution space, leading to exponential blowup in terms of both time and space. To alleviate both issues, we propose a novel alternating optimization algorithm that can scale to large, uncertain distance sets with as many as 100,000 points. This algorithm is space and time-efficient, with each step running in 𝒪(m log (m)) time and requiring only 𝒪(√(m)) working space for a distance set of size m. Evaluations of this approach on synthetic and partial digest data showcase improved accuracy and scalability in the presence of uncertain, duplicated, and missing distances. Our implementation of the algorithm is available at https://github.com/Kingsford-Group/turnpikesolvermm .

Minimizers and syncmers are sketching methods that sample representative k-mer seeds from a long string. The minimizer scheme guarantees a well-spread k-mer sketch (high coverage) while seeking to minimize the sketch size (low density). The syncmer scheme yields sketches that are more robust to base substitutions (high conservation) on random sequences, but do not have the coverage guarantee of minimizers. These sketching metrics are generally adversarial to one another, especially in the context of sketch optimization for a specific sequence, and thus are difficult to be simultaneously achieved. The parameterized syncmer scheme was recently introduced as a generalization of syncmers with more flexible sampling rules and empirically better coverage than the original syncmer variants. However, no approach exists to optimize parameterized syncmers. To address this shortcoming, we introduce a new scheme called masked minimizers that generalizes minimizers in manner analogous to how parameterized syncmers generalize syncmers and allows us to extend existing optimization techniques developed for minimizers. This results in a practical algorithm to optimize the masked minimizer scheme with respect to both density and conservation. We evaluate the optimization algorithm on various benchmark genomes and show that our algorithm finds sketches that are overall more compact, well-spread, and robust to substitutions than those found by previous methods. Our implementation is released at https://github.com/Kingsford-Group/maskedminimizer. This new technique will enable more efficient and robust genomic analyses in the many settings where minimizers and syncmers are used.

Most sequence sketching methods work by selecting specific k-mers from sequences so that the similarity between two sequences can be estimated using only the sketches. Because estimating sequence similarity is much faster using sketches than using sequence alignment, sketching methods are used to reduce the computational requirements of computational biology software packages. Applications using sketches often rely on properties of the k-mer selection procedure to ensure that using a sketch does not degrade the quality of the results compared with using sequence alignment. Two important examples of such properties are locality and window guarantees, the latter of which ensures that no long region of the sequence goes unrepresented in the sketch. A sketching method with a window guarantee, implicitly or explicitly, corresponds to a Decycling Set, an unavoidable sets of k-mers. Any long enough sequence, by definition, must contain a k-mer from any decycling set (hence, it is unavoidable). Conversely, a decycling set also defines a sketching method by choosing the k-mers from the set as representatives. Although current methods use one of a small number of sketching method families, the space of decycling sets is much larger, and largely unexplored. Finding decycling sets with desirable characteristics (e.g., small remaining path length) is a promising approach to discovering new sketching methods with improved performance (e.g., with small window guarantee). The Minimum Decycling Sets (MDSs) are of particular interest because of their minimum size. Only two algorithms, by Mykkeltveit and Champarnaud, are previously known to generate two particular MDSs, although there are typically a vast number of alternative MDSs. We provide a simple method to enumerate MDSs. This method allows one to explore the space of MDSs and to find MDSs optimized for desirable properties. We give evidence that the Mykkeltveit sets are close to optimal regarding one particular property, the remaining path length. A number of conjectures and computational and theoretical evidence to support them are presented. Code available at https://github.com/Kingsford-Group/mdsscope.