ML biocatalysis - e.g. transamination reaction

- highly selective enzymes replace conventional chemistry

Structure- and Data-Driven Protein Engineering of Transaminases for Improving Activity and Stereoselectivity
Yu-Fei Ao et. al, Angewandte Chemie 2023. https://doi.org/10.1002/anie.202301660

3FCR transaminase substrate screening

Structure- and Data-Driven Protein Engineering of Transaminases for Improving Activity and Stereoselectivity
Yu-Fei Ao et. al, Angewandte Chemie 2023. https://doi.org/10.1002/anie.202301660

"classical" ML approaches

Structure- and Data-Driven Protein Engineering of Transaminases for Improving Activity and Stereoselectivity
Yu-Fei Ao et. al, Angewandte Chemie 2023. https://doi.org/10.1002/anie.202301660

the greifswald protein screening platform LARA

what is semantics enabled machine learning (ML) ?

• better "understanding" of the data by the ML algorithms
• semantics guided ML model building
  by automated feature extraction
• improved consitancy checking, because the ML algorithm can extract physical limits
• improved error handling, because the ML algorithm "knows" more about the predicted system

what are we trying to build ?

• generic software platform closed loop designs of the simple kind
• autonomous experiment design
• autonomous experiment exection
• autonomous experiment evaluation
• AI and machine learning
• redesign of the experiment
• continue this process until an desired outcome is reached or resources are used up

the challange

• full SiLA communication of all devices in all platforms
• platform independant process description language
• flexible programming language for complex processes
• dynamic and error tolerant scheduling
• full experimental metadata for machine learning applications
• storage and exchange of data between labs (KIWI biolab)
• fully semantic annotation of the metadata (ontologies)

benfits for machine learning applications

• rich metadata from ground up
• ontology based (s. ontology development workflow)
• allowing autonomous feature extraction

challanges to overcome

• heterogenous data
• non-standardised data structure
• non-standardised metadata - no semantics
• non-standardised device communication
• black-box software - closed source
• no advanced, comprehensive data storage

building homgeneous infrastructure from ground up

... what software components are required ?

• device control
• data readout
• data transfer
• data storage
• semantic understanding of the data by the machines
• machine learning algorithms to "work on the data"
• feedback software to the instruments
• common description language
• process execution / scheduler

SiLA servers/devices of LARA

pythonLab

https://gitlab.com/opensourcelab/pythonLab
universal, python based, automation language

pythonLabOrchestrator & Scheduler

https://gitlab.com/opensourcelab/pythonlabscheduler

pythonLabScheduler in the wild (stefan maak)

https://gitlab.com/opensourcelab/pythonlabscheduler

LabDataReader

https://gitlab.com/opensourcelab/LabDataReader

• generic data reader framework for propriatory (text based) data formats
• new data formats can be added as plugins
• rich meta data support
• automatic semantic annotation of the data
• fully written in python
• output formats: pandas data frame, JSON-LD, csv, (AnIML/JSON-LD - under development)

overview of final architecture

fully open sourced and python based

summary

• building a large, community driven open source infrastructure
• no black box, everything is adjustable and extendible
• advanced queries and automated feature extraction
• feedback loop driven protein engineering is possible with the tools in place
• machine learing applications will dramatically improve, when the right metadata is in place
• we need a modern mind set (biochemistry and machine learning)
• we are not alone, please join and contribute !