Protein representations from language models have yielded state-of-the-art performance across many tasks in computational protein engineering. In recent years, progress has focused on parameter count, with recent models' capacities surpassing the size of the very datasets they were trained on. Here, we propose an alternative direction. We show that large language models trained on codons, instead of amino acid sequences, provide high-quality representations that outperform comparable state-of-the-art models across. These results suggest that, in addition to commonly studied scale and complexity, the information content of biological data provides an orthogonal direction to improve the power of machine learning in biology.

How noncoding DNA determines gene expression in different cell types is a major unsolved problem, and critical downstream applications depend on improved solutions. To advance this, we developed Enformer, a hybrid convolutional and transformer model trained to predict thousands of epigenomic experiments including gene expression from human and mouse reference genomes. I will discuss Enformer’s performance on variant effect prediction which may be relevant for regulatory DNA sequence design.

We have developed and implemented a machine learning model to predict protein expression. The model was coupled to an in silico screening procedure that systematically designs and assesses thousands of constructs in a high-throughput manner. We will share our plans to improve the model by (1) streamlining internal data registration, (2) considering yield values instead of classes, (3) incorporating protein sequence embeddings based on AI language models, and (4) leveraging external datasets. Limited availability of training data is a key blocker, so we are exploring sharing data via a pre-competitive consortium in collaboration with EMBL-EBI and other academic and industry partners.

Bovine enterokinase light chain is used for affinity-tag removal. Expression in E. coli leads to insoluble inclusion bodies. Directed evolution yielded 321 unique variants, with up to >11,000-fold increased soluble expression, mainly due to stability. Codon optimisation improved expression at 37°C. However, non-optimised codons and expression at 30°C gave the highest activities. Partial least squares analysis revealed that soluble variants tended to combine stabilising mutations outside the active site.

The pET plasmids constitute the most popular protein production system. Using synthetic experimental evolution, we have improved the performance of several of the pET genetic modules and bacterial strains. In addition, we have developed an extremely simple method for making recombinant DNA. The approach and the genetic modules will be combined into a next-generation pET system.

Recombinant protein production allows us to create smart materials and catalysts. Our mission is to find solutions to produce proteins, enzymes and metabolites using the yeast Pichia pastoris, which is an efficient alternative for recombinant production combining the simplicity of bacterial expression systems with some essential features of higher eukaryotic hosts. We can build on over 15 years of experience in toolbox development and gene expression using Pichia pastoris to overcome hurdles in recombinant production. We adopt available technology to our needs and evaluate innovative new strategies for the expression of our proteins and enzymes.

Mammalian cells are the preferred host cells for therapeutic protein production and have been engineered to contain desired attributes for increased protein production. To identify novel engineering targets, laborious and time-consuming empirical approaches have been attempted. Here, I present a genome-wide CRISPR-Cas9 screening platform for CHO and HEK293 cells using a virus-free, recombinase-mediated, cassette exchange-based gRNA integration method to identify novel targets for high productivity and culture-stress resistance.

In recombinant protein production with CHO cells, bottlenecks in productivity or product purity issues require a particular cellular or clonal mechanism to be analyzed. Emerging analytical techniques allow ever more detailed insights into cellular processes involved in protein expression or cultivation performance. Thus, we performed targeted studies on CHO sub-OMICs, including the miRNome, cell surfaceome, as well as secreted HCPs and extracellular vesicles, to address specific issues of biopharmaceutical production.

In the last years, we have been dealing with the production of different recombinant proteins in human cells, by means of PEI-based transfection and further transient gene expression. We have been constantly pursuing the improvement of the process, in terms of higher expression levels. We have explored different parameters and conditions (some of them already published for other recombinant proteins), and we have implemented in our process those changes that improved protein yield. In our talk, we will detail such continuous process, and show where we started, and where we are now.