Large scale active-learning-guided exploration for in vitro protein production optimization

Abstract

Lysate-based cell-free systems have become a major platform to study gene expression but batch-to-batch variation makes protein production difficult to predict. Here we describe an active learning approach to explore a combinatorial space of~4,000,000 cell-free buffer compositions, maximizing protein production and identifying critical parameters involved in cell-free productivity. We also provide a one-step-method to achieve high quality predictions for protein production using minimal experimental effort regardless of the lysate quality.

Combining mechanistic and machine learning models for predictive engineering and optimization of tryptophan metabolism

“…Likewise, while the one-pot library construction used in this study had an estimated coverage of 48% of the full combinatorial design space, the amount of strains used for training the algorithms only covered 3%, yet enabled predictive engineering following a single design-build-test-learn cycle. This could be used to argue that more engineering iterations on even smaller datasets, potentially coupled to mixed exploitation and exploration approaches as recently demonstrated for cell-free production 54 , should be a valid avenue for ML-guided engineering of even less genetically tractable chassis, and for which no high-throughput screening method may even exist. With regards to this, we performed a follow-up test running the ART and EVOLVE approaches in explorative and exploitative modes, respectively.…”

Cell-Free Systems: A Proving Ground for Rational Biodesign

“…In the absence of a predictive model fully connecting all its components to protein output, Caschera et al ( 2018 ) used an evolutionary design of experiments approach to iteratively train an ensemble neural network model to optimize conditions for cell-free protein synthesis. More recently Borkowski et al ( 2020 ) trained a similar model on ~4,000 reactions, improving yields by 34 times. Importantly, they discovered a training dataset of only 20 compositions which was informative enough to allow the model to generalize its predictions to different lysates and conditions.…”

Robust Cell-Free Expression of Sub-Pathological and Pathological Huntingtin Exon-1 for NMR Studies. General Approaches for the Isotopic Labeling of Low-Complexity Proteins

“…Details of the lysate production and the batch CF protocol used here have been previously reported [ 25 ]. As in all CF protocols, almost all parameters can be optimized in order to improve production yields [ 60 ]. In our hands and in line with most protocols, the Mg 2+ concentration is the most critical parameter.…”

“…The ensemble of these experiments shows that the correct concentration balance of all biomolecules is key for the efficiency of translation and that the addition of loaded tRNA CUA in suppression experiments, which probably disrupts this balance, cannot be recovered by additional amounts of EF-Tu alone. However, the simultaneous optimization of multiple CF reaction parameters at once may lead to an improvement of CF suppression yields in a similar way that it improved standard CF reactions [ 60 ].…”