Learning from data

A wide variety of artificial intelligence systems have been in use for many years and the number is constantly increasing. This not only applies to everyday private life. Areas of application have long been identified in many industries and companies in which AI can make unpleasant everyday tasks easier, improve product quality, open up new service offerings and support or even automate decision-making. In the laboratory environment, however, reality still looks different. This is surprising at first glance, as the potential uses of AI in the laboratory are just as varied and rewarding as in all other industries that have been relying on AI support for several years.

There are two main reasons why the use of AI in the laboratory is currently still so low despite its great potential: Firstly, AI requires digital data – which is not yet available in many laboratories. They therefore first need a suitable infrastructure in order to be able to benefit from AI. On the other hand, companies often do not yet recognise that the use of AI is a possible solution for an individual problem that has already been identified. In addition, many laboratory managers fear high development costs for their own AI solution with an unclear end result. This can be countered in a structured AI project with an initial assessment.

Preparation: Making data available

Artificial intelligence systems sometimes require large amounts of digital data. Before they are used for AI development, initial data analyses can be carried out in advance. These analyses require a smaller amount of data and often provide initial answers to specific questions. Whether companies have already collected this data or are just starting to collect it has an impact in terms of time at most. Data scientists use a large proportion of the available data to train the AI model. This teaches it to recognise patterns and correlations in the data, to remember them and to apply this knowledge independently. The better the training data, the better the knowledge. The other part of the data serves as a test environment: the data scientists use this data to test the quality of the AI.

1. Device manufacturers currently lack large amounts of data on how their devices are used (e.g. operating times, temperature data, energy history logs, vibration analyses, etc.) and are therefore unable to offer services such as failure predictions or intelligent maintenance intervals.
Software manufacturers for LIMS and middleware solutions can offer AI modules for generic simplification in the laboratory, but there is a lack of test and process data from the laboratories to feed the modules with corresponding data.
2. Laboratories are the biggest losers. In many places, test data is still transferred to a LIMS via USB stick or even manually, or lab notebooks are filled out manually. This not only means a high time factor, but also a risk of error. In addition, laboratories currently have little chance of meeting their specific challenges with their own customised AI solution, as large digital data collections are usually lacking.

The solution would be: digitalisation and networking. This includes, in particular, a standardised communication interface so that all laboratory devices from different manufacturers and disciplines can be easily linked with different laboratory software. One such open and manufacturer-independent standard is LADS (Laboratory Analytics Device Standard), which is currently being developed under the leadership of the industry association Spectaris and in collaboration with around 40 contributors from the global laboratory sector. It is based on the OPC UA standard, which was initiated by manufacturing companies in the early 1990s and is now the established standard in the industrial environment worldwide.

As soon as previously self-sufficient laboratory devices can be connected to laboratory software, a huge potential for optimisation opens up for laboratories and manufacturers of laboratory devices and software. This would enable a large number of laboratories to automate their test procedures, processes and documentation. This reduces the risk of errors, increases flexibility and frees up already scarce personnel resources for more valuable tasks than transferring data from one system to another. And this also creates - as a by-product, so to speak - the opportunity to collect data on a large scale - the basis for the use of AI.

From data to the finished AI tool

The development of an artificial intelligence system is fundamentally independent of the industry in which it is used. Data scientists abstract from the industry-specific use case to a data-centred view and select the appropriate model and suitable methods to train the model. Whether an AI should predict the failure of a gas turbine or a pipetting robot is therefore basically irrelevant to the development of the AI.

Companies often start by collecting data indiscriminately and hope that they will be able to extract some knowledge from it later on. In fact, it is possible to check whether statistical methods and neural networks find inconsistencies, patterns and correlations in this data that were previously unknown to those responsible. However, the chances of success are unclear in advance and the economic viability can only be assessed in retrospect.
Ideal, on the other hand, is when a specific task or problem has already been identified in the company, the solution to which can advance the company's development and competitiveness. In the case of a device manufacturer, for example, this could be device failures, the cause and early detection of which is important in order to develop more robust devices and minimise unplanned device downtimes for customers. In laboratories, for example, the optimisation of quality control would be a possible reason for developing an AI solution.

This is what it can look like: Based on a specific use case, data scientists can start a structured AI project in which the benefits are also transparent at all times: With data exploration, they clarify which data is required before the project starts, whether data that is already available is suitable for the respective use case and in what form the data must be available. Only then, in the second step, is an initial AI model created using rapid prototyping, which provides information on whether fully comprehensive AI modelling is economically viable and makes sense in terms of content. If so, data scientists and software developers develop and train the AI model, build data pipelines for a continuous data flow in later practical use, integrate the AI into user interfaces and existing IT infrastructure and optionally develop maintenance mechanisms and routines.

Assessing mass spectra with AI

The AI project at Microsynth, where artificial intelligence is soon to be used in laboratories, is currently going through exactly the same process. Microsynth AG, headquartered in Balgach, Switzerland, specialises in specific DNA and RNA oligonucleotide syntheses. Before the oligonucleotides produced are sent to customers, they undergo strict quality control: experts use mass spectrometry to determine whether the composition of the oligonucleotides produced meets the customer's specifications. For this purpose, one person analyses each mass spectrum produced and assesses whether the peaks correspond to the expected picture or whether there are unusual peaks. At Microsynth, up to 2,000 measurements are taken every day, which have to be assessed within a few hours and before the oligonucleotides are dispatched - an important task, but also a major time and cost factor. Those responsible therefore asked themselves whether an AI could support or even automate quality control. The idea: an AI checks the measurement results and only calls in human support if the control does not meet expectations. In this way, the AI would speed up the evaluation process and the human experts would have more time for other valuable work.

Microsynth received positive feedback on the feasibility of an AI tool and a homework assignment from the commissioned data scientists: Data is needed. In order for the desired AI to be able to assess the mass spectra independently, it needs spectra of various oligonucleotides as a database. In this way, it learns what the correct results are and which results deviate from the norm. Although the creation of the database involves work, it pays off with all the more precise AI results. The obvious thing to do is to systematically save the data from the many daily measurements and label them with the labels „correct spectrum“ or „unclear spectrum“. However, an AI model developed by Google may also be able to support the prediction of mass spectra.

Article from "LABO" dated 03 May 2023