Combinatorial pathway optimization is a powerful approach in metabolic engineering to improve strain performance. While machine learning (ML) has shown promise in guiding the Design-Build-Test-Learn (DBTL) cycle, most applications have been limited to small design spaces, thereby restricting the potential of predictive and exploration-exploitation strategies. In this work, we applied two DBTL cycles to optimize p-coumaric acid production in Saccharomyces cerevisiae. The first cycle involved constructing a large combinatorial library of 18 genes and 20 promoters (170 million possible designs). In the second cycle, we employed a gradient bandit-based machine learning recommendation strategy, tuned to balance exploration and exploitation. Our results show that this balanced strategy outperforms greedy, feature importance-based approaches, leading to greater diversity in strain performance and improved top-producer identification. Notably, applying the same strategy to an alternative parent strain yielded the highest p-coumaric acid titer (1.23 g/L), a 2.37-fold improvement over the original. These findings highlight the value of ML-guided exploration in large design spaces and demonstrate that balancing exploration and exploitation is critical for successful strain optimization.

Motivation: Metabolic kinetic models are widely used to model biological systems. Despite their widespread use, it remains challenging to parameterize these Ordinary Differential Equations (ODE) for large scale kinetic models. Recent work on neural ODEs has shown the potential for modeling time-series data using neural networks, and many methodological developments in this field can similarly be applied to kinetic models. Results: We have implemented a simulation and training framework for Systems Biology Markup Language (SBML) models using JAX/Diffrax, which we named jaxkineticmodel. JAX allows for automatic differentiation and just-in-time compilation capabilities to speed up the parameterization of kinetic models, while also allowing for hybridizing kinetic models with neural networks. We show the robust capabilities of training kinetic models using this framework on a large collection of SBML models with different degrees of prior information on parameter initialization. We furthermore showcase the training framework implementation on a complex model of glycolysis. Finally, we show an example of hybridizing kinetic model with a neural network if a reaction mechanism is unknown. These results show that our framework can be used to fit large metabolic kinetic models efficiently and provides a strong platform for modeling biological systems. Implementation: Implementation of jaxkineticmodel is available as a Python package at https://github.com/AbeelLab/jaxkineticmodel.

Background: Assembly algorithm choice should be a deliberate, well-justified decision when researchers create genome assemblies for eukaryotic organisms from third-generation sequencing technologies. While third-generation sequencing by Oxford Nanopore Technologies (ONT) and Pacific Biosciences (PacBio) has overcome the disadvantages of short read lengths specific to next-generation sequencing (NGS), third-generation sequencers are known to produce more error-prone reads, thereby generating a new set of challenges for assembly algorithms and pipelines. However, the introduction of HiFi reads, which offer substantially reduced error rates, has provided a promising solution for more accurate assembly outcomes. Since the introduction of third-generation sequencing technologies, many tools have been developed that aim to take advantage of the longer reads, and researchers need to choose the correct assembler for their projects. Results: We benchmarked state-of-the-art long-read de novo assemblers to help readers make a balanced choice for the assembly of eukaryotes. To this end, we used 12 real and 64 simulated datasets from different eukaryotic genomes, with different read length distributions, imitating PacBio continuous long-read (CLR), PacBio high-fidelity (HiFi), and ONT sequencing to evaluate the assemblers. We include 5 commonly used long-read assemblers in our benchmark: Canu, Flye, Miniasm, Raven, and wtdbg2 for ONT and PacBio CLR reads. For PacBio HiFi reads, we include 5 state-of-the-art HiFi assemblers: HiCanu, Flye, Hifiasm, LJA, and MBG. Evaluation categories address the following metrics: reference-based metrics, assembly statistics, misassembly count, BUSCO completeness, runtime, and RAM usage. Additionally, we investigated the effect of increased read length on the quality of the assemblies and report that read length can, but does not always, positively impact assembly quality. Conclusions: Our benchmark concludes that there is no assembler that performs the best in all the evaluation categories. However, our results show that overall Flye is the best-performing assembler for PacBio CLR and ONT reads, both on real and simulated data. Meanwhile, best-performing PacBio HiFi assemblers are Hifiasm and LJA. Next, the benchmarking using longer reads shows that the increased read length improves assembly quality, but the extent to which that can be achieved depends on the size and complexity of the reference genome.

Combinatorial pathway optimization is an important tool in metabolic flux optimization. Simultaneous optimization of a large number of pathway genes often leads to combinatorial explosions. Strain optimization is therefore often performed using iterative design-build-test-learn (DBTL) cycles. The aim of these cycles is to develop a product strain iteratively, every time incorporating learning from the previous cycle. Machine learning methods provide a potentially powerful tool to learn from data and propose new designs for the next DBTL cycle. However, due to the lack of a framework for consistently testing the performance of machine learning methods over multiple DBTL cycles, evaluating the effectiveness of these methods remains a challenge. In this work, we propose a mechanistic kinetic model-based framework to test and optimize machine learning for iterative combinatorial pathway optimization. Using this framework, we show that gradient boosting and random forest models outperform the other tested methods in the low-data regime. We demonstrate that these methods are robust for training set biases and experimental noise. Finally, we introduce an algorithm for recommending new designs using machine learning model predictions. We show that when the number of strains to be built is limited, starting with a large initial DBTL cycle is favorable over building the same number of strains for every cycle.