An Integrated Laboratory Robotic System for Autonomous Discovery of Gene Function

Abstract

Progress in laboratory automation depends not only on automating the physical aspects of scientific experimentation, but also on the intellectual aspects. We present the conceptual design, implementation, and our user-experience of “Adam,” which uses machine intelligence to autonomously investigate the function of genes in the yeast Saccharomyces cerevisiae. These investigations involve cycles of hypothesis formation, design of experiments to test these hypotheses, physical execution of the experiments using laboratory automation, and the analysis of the results. The physical execution of the experiments involves growing specific yeast strains in specific media and measuring growth curves. Hundreds of such experiments can be executed daily without human intervention. We believe Adam to be the first machine to have autonomously discovered novel scientific knowledge.

Introduction

We wish to automate all aspects of laboratory science, not just the physical experimentation, but also the intellectual aspects of hypothesis formation, experiment planning, and results analysis. A “Robot Scientist” is a physically implemented robotic system that applies techniques from artificial intelligence (AI) to execute cycles of automated scientific experimentation. ¹ This contrasts with standard laboratory automation that normally focuses on just the physical aspects of experimentation. Our Robot Scientist “Adam” executes, with minimal human intervention, a complex combination of operations on yeast cell cultures at medium to high throughput and, moreover, is capable of modifying those operations according to the behavior of the organisms. ² Again, this contrasts with standard laboratory automation that is normally characterized by medium- or high-throughput execution of a linear sequence of a relatively small number of different operations. ^{3 5}

Automating Scientific Discovery

Automation has been integral to many of the changes in human society since the 19th century. The advent of computer science in the mid-20th century has made practical the idea of automating aspects of scientific discovery. Computers were initially used to automate simple linear processes, for example, to collect and process laboratory instrument data, perform astronomical calculations, and create ballistic tables. Later, AI began to be used to automate aspects of planning experiments and analyzing results. Meta-DENDRAL, developed in the 1960s, was the first automated system for scientific hypothesis generation. It used inference to suggest chemical structures responsible for mass spectrometry data. ⁶ BACON, ⁷ ABACUS, ⁸ FAHRENHEIT, ⁹ and IDS ¹⁰ were all impressive examples of automated data-driven discovery systems that could discover scientific laws as algebraic equations. A more recent example uses iterative cycles of algorithmic correlation to distil natural laws of geometric and momentum conservation, using data captured from the motion-tracking studies of a range of simple and complex oscillators and pendula. ¹¹

However, none of the systems described fully “closes the loop;” they either do not collect their own data, or do not use analyzed results to update their knowledge base, or do not automatically perform cycles of scientific discovery.

Our Robot Scientist “Adam” goes a step further than these other systems and also automatically performs experiments. Adam uses a detailed logical model of yeast metabolic pathways, ^12,13 along with bioinformatic and AI methods to form sets of hypotheses about yeast functional genomics. Adam then designs and executes physical experiments to test these hypotheses. The experimental observations are optical density (OD) readings, from which growth curve parameters are derived and automatically analyzed by statistical and machine learning methods, and the results are used to confirm or refute the hypotheses. Adam then decides whether to execute further cycles of hypotheses generation and experimentation, or whether it can infer new knowledge and update its model (see Fig. 1). Human intervention is limited to creating yeast library stocks and supplying consumables.

Figure 1.

Hypothesis generation and elimination.

Why Yeast?

The focus of Adam is on generating novel biological knowledge. We therefore deigned Adam to work with the model organism Saccharomyces cerevisiae (baker's yeast). It has a relatively small genome, and it is the best-understood eukaryotic organism. It grows fast and is ideal for experimentation in microtiter plates. It also has an unrivaled collection of mutant strains. The genetically modified haploid strains used in Adam's experiments each have a single gene deleted, and Adam's library contains over 4500 such strains. ¹⁴ More than 900 yeast genes still have unidentified functions ¹⁵ and there are reactions in the metabolic pathways that are catalyzed by enzymes for which related genes are unidentified (referred to as “locally orphan”).

System Description: Technologies and Methods

The robotic system executes iterative cycles of sample preparation and monitoring (see Fig. 2) and comprises numerous pieces of automated equipment (see Fig. 3) arranged as three subsystems (see the next section). AI software creates plate layouts that specify yeast strains, the required volumes and concentrations of growth media and metabolites, and the layout of the experiments on the microtiter plates. The plates are then created and processed, with growth data automatically recorded, and then analyzed to confirm or refute hypotheses. Finally AI software decides whether to perform further cycles or whether new knowledge has been inferred.

Figure 2.

One cycle of automated experimentation for a single microtiter plate.

Figure 3.

Layout of the robotics and instrumentation in the Adam Robot Scientist: (1) Liconic STR602 freezer, (2) Zymark Presto liquid handler, (3) Thermo 384 multidrop, (4) two Caliper Twister 2 robot arms, (5) Caliper Sciclone i1000 liquid handler, (6) Bio-Tek 405 plate washer, (7) Velocity 11 VSpin plate centrifuge, (8) three Liconic STX40 incubators, (9) two Molecular Devices Spectramax 190 plate readers, (10) Variomag plate shaker, (11) IAI Corporation Scara robot arm, (12) two pneumatically actuated plate slides, (13) two high efficiency particulate air (HEPA) filters, and (14) aluminum and rigid transparent plastic enclosure.

Process Description

The robotic system comprises three subsystems. Subsystem 1 picks selected yeast strains from frozen library microplate stocks and inoculates pregrowth microplates containing YPD-rich growth medium, ^a according to a “picking” file generated at the plate layout design stage. Subsystem 2 incubates the pregrowth plates and monitors cell growth, allowing the yeast to multiply until cell concentrations in most of the wells have reached a predefined OD measured at a wavelength of 595 nm. It then harvests and normalizes the yeast cell concentrations, dispenses the required media and metabolite solutions into the experiment plate according to the relevant “volume” files generated at the plate layout design stage, and inoculates the experiment plate with the yeast strains from the pregrowth plate. This starts the experiments that will be monitored by subsystem 3. The pregrowth plate is then removed as waste. Subsystem 3 incubates the experiment plates and monitors cell growth via regular OD readings; after 72 h the microplate is sent to waste and growth curve analysis triggered for both the pregrowth and experiment plate results.

Major Software Components

Relational Database. All data generated are held in a purpose-designed relational database implemented in MySQL. ^b This database system is easily capable of handling the data volume (tens of millions of rows for the growth curve data alone), is fast and robust, is freely available, and interfaces easily to many programming languages. The current schema stores the hypotheses, experiment designs, plate layouts, observations, analyzed results, derived statistical results and metadata (including yeast library plate contents, picking history and environment sensor data). It also includes laboratory chemical stock details and records which batch was used and for what purpose. Web interfaces in standard web languages (Java, JSP, XHTML, and XML) are accessed via an Apache Tomcat server and are used to monitor current experiments and results, and to maintain laboratory chemical stocks.

Hypothesis Generation. Adam begins an investigation by searching its logical yeast model for “locally orphan enzymes” that are on metabolic pathways. Locally orphan enzymes are enzymes that are known to be in yeast, but for which the genes encoding them are not known. The focus on metabolic pathways is a heuristic, which makes it more likely that a deletant strain will exhibit a phenotypic growth effect. Once locally orphan enzymes are identified, Adam then generates hypotheses about which genes could encode them. This is done as follows: using KEGG ¹⁶ Adam looks for genes from other organisms that encode enzymes catalyzing the same enzymic function; Adam then uses bioinformatics (primarily sequence similarity via PSI-BLAST ¹⁷ ) to identify homologous genes within the yeast genome; these homologous genes are then hypothesized to encode the locally orphan enzymes. This type of hypothesis formation process is an example of abductive inference. ¹⁸

Experiment Planning. Once a hypothesis has been generated, the next step is to automatically decide on the experiments required to test it. This is done by assuming the hypothesis to be true, and then inferring the deductive consequences of this assumption. Specifically, as Adam has access to a library of single-gene deletant strains, it is able to infer that the specific deletant strain identified by the hypothesis will be missing the locally orphan enzyme. Then all reactions and corresponding metabolites, within a number of steps downstream of the reaction(s) encoded by the locally orphan enzyme are identified. Adam then infers that, if the hypothesis is correct, adding these metabolites to the defined growth medium will differentially affect the growth of the deletant compared with the wild-type. Adam does not assume that the deletant strain will be auxotrophic as almost all the genes that result in auxotrophy have probably already been identified. What remain to be discovered are likely those genes that only have qualitative phenotypes when removed because of isoenzymes, promiscuous enzymes, and so on. ² This collection of subordinate hypotheses grouped under the primary hypothesis we term a “hypotheses set.”

The availability of the required metabolites and yeast strains were ascertained and then Adam designs the physical experiments. Not all hypotheses in each hypotheses set can necessarily be translated into experiments. The end result of this process is a set of viable “trial” groups of individual “tests,” where a trial is related to a specific hypotheses set. A typical trial is made up of four “test” conditions: deletant yeast strain with metabolite, deletant yeast strain without metabolite, wild-type yeast strain (BY4741) with metabolite, and wild-type yeast strain without metabolite. This experimental design enables Adam to compare the growth of the deletant strain in the presence of the relevant metabolite with the growth of the standard wild-type. Each test also has a number of “test instances,” which is the requested number of replicates to be run to give statistically meaningful results. An evolutionary multiobjective optimizer ¹⁹ can then be used to determine which trials can be run together, without changing the metabolite solutions (the system can dispense no more than eight different metabolites or medium components), using factors, such as trial priority, number and identity of metabolites required, availability of metabolites, their cost, and suggested size of the “trial set.”

Plate Layout Creation. Individual test instances are arranged into plate layouts according to experiment design templates. A trial normally contains 96 test instances and corresponds to a single plate. Latin square ²⁰ layouts are used to help minimize plate-edge effects and systematic noise. Such layout designs are only practical under full automation and require a high-throughput liquid handler able to dispense different volumes to each well (such as Adam's Sciclone i1000, Sciclone i1000, Caliper Life Sciences, Hopkinton, MA). This process creates two sets of plate layouts: one for the “pre-growth” plates and one for the “experiment” plates.

Pick and Volume File Creation. The final stage before running the system is creation of yeast library “picking” files for the pregrowth plates and “volume” files for the i1000 liquid handler. Picking files detail which library plate well will inoculate which pregrowth well and the picking order and are designed to prevent library plates thawing by being out of the freezer for too long, whilst also trying to minimize the overall plate creation time. The software also checks how many times specific wells have been picked, and the age of the yeast stock (cell viability decreases over time). Eight volume files are created for each experiment plate: one file for defined medium, one for water, and the remaining six for metabolite solutions. Volumes in the water file are adjusted at runtime by the yeast inoculum normalization calculation.

Growth Curve Analyses. Once a trial has completed, Adam analyses the OD versus time-series measurements for each plate well. OD is a proxy for cell count and can be modeled to a growth curve from which growth parameters can be derived. These data are often noisy, and there are often missing observations; these are handled by a set of data analysis programs we have developed. We apply an empirical approach based on cubic spline polynomials. This incorporates techniques for handling missing data points and noise and comparisons indicate that on our data it provides improved results when compared with other empirical methods (e.g., see reference ²¹ ) or classical sigmoidal modeling approaches (e.g., see reference ²² ). The growth parameters extracted include maximum growth rate (and hence doubling time), lag time, start and end time points of the period of maximum exponential growth, OD at the end point of the period of maximum exponential growth, global maximum OD reached, and also a signal-to-noise ratio.

Parametric and nonparametric statistical tests, random forests, and Monte Carlo resampling (used to establish significance) are then applied to determine whether there is a significant difference observed when adding the metabolite(s) to the deletant strains compared with the wild-type. This enables Adam to decide which hypotheses have evidence consistent with them and which do not.

Control of Robotics and Instrumentation

The equipment control software was provided by the system integrators, Caliper Life Sciences. Three instances of the Zymark Control Supervisor (ZCS) software using “iLink” protocols are executed (one for each subsystem), using separate PCs running Windows XP. Each iLink method interacts with an MS Excel spreadsheet that stores system control and plate tracking data, and also with the instrument control programs (ICPs) that control individual instruments and robots. Some ICPs are as supplied by their manufacturers but most were developed by Caliper. A fourth PC coordinates the three independent ZCS instances via Caliper software written in MS Excel Visual Basic, providing the nondeterministic action of subsystems required by unpredictable growth and a simple graphical user interface for user interaction, error handling, and status monitoring. Experiment data are output in the form of text files containing the OD readings from each Spectramax plate reader (SoftMax Pro software v. 4.8; MDS Analytical Technologies, Concord, Ontario, Cananda).

Results

We believe Adam to be the first machine demonstrated to be able to discover novel scientific knowledge. Adam has hypothesized and experimentally confirmed the function of 12 genes encoding locally orphan enzymes in yeast (see reference ² ). For three of these genes we manually confirmed that Adam was correct by purifying the protein product and demonstrating enzymic function. For six more of the genes we discovered that strong empirical evidence already existed in the scientific literature that supported Adam's conclusions, that is, that they were not in fact true local orphans, and that Adam only identified them as such because its bioinformatic database was incomplete. This inadvertently provided us with positive controls showing Adam's methodology does work. For one hypothesis Adam came to an inaccurate conclusion; this was because the hypothesized gene candidate YIL033C (BCY1) was predicted to be a glutaminase (enzyme class 3.5.1.2), and Adam experimentally confirmed this using 11 metabolites predicted to have differential effect on a glutaminase deletant. However, YIL033C has a cAMP-dependent protein kinase subunit that is involved in regulating metabolism, so this may be sufficient to explain the same phenotype. This exposes a weakness in Adam's current metabolic model that does not include control mechanisms. However, it is possible that YIL033C is both a kinase and a glutaminase, and scientific evidence also exists to support this theory (see reference ²³ ). See our website for more details on Adam's results, its hardware, and its software. ^c

Discussion

Procuring Laboratory Automation

For the benefit of those new to laboratory automation we give here a brief summary of the procurement process.

It is important for the user to carefully research and define the proposed biological application, manually testing key experimental aspects where possible. Then he or she should investigate the availability, suitability, and capital, running, and maintenance costs of any potentially appropriate equipment. An outline design will help clarify system requirements, including nonfunctional aspects such as access for cleaning and maintenance, capacities for consumables and wastes, and ease of their replenishment and removal.

Integration companies should then be invited to submit proposals against detailed requirements documentation, in accordance with relevant procurement legislation. The selected integrator should clearly understand the requirements, be happy to work closely with the end-user during the design and build, and afterward to refine the system. Proposed control software should be reliable, user-friendly, flexible (to allow later refinements and improvements to the application), expandable (to allow additional instruments or functionality to be added subsequently, if required), have good error reporting and recovery, good event tracking, and be fully maintainable.

Subsequent stages involve the build of the full system, integration of the control software, factory acceptance tests (FAT), delivery to end location (preferably by a specialist equipment removals company), site acceptance tests (SAT), commissioning, refinement of biological application, and integration with the laboratory software. The user should detail the FAT specification, for agreement with the integrator, and this should specify a comprehensive set of tests to exercise all the basic capabilities of the system and associated equipment but normally without organisms and chemicals. The SAT should repeat the FAT and then test the ability of the system to execute the required biological protocols to an acceptable standard of reproducibility in the laboratory environment, with actual organisms and chemicals where required.

Practical Experience of Adam

Here we discuss some of the more important issues that arose with the use of Adam.

Format and Storage of the Yeast Library. The initial intention was to store an entire library of haploid deletant yeast strains (over 4500) in the online automated freezer to give Adam maximum flexibility in designing experiments. The automated −20 °C Liconic STR602 freezer has ample capacity and has proven very reliable. It stores the yeast strains in high concentration in 96-well deep-well plates. However, viability of the frozen stocks drops over just a few weeks, so it is only practical to make up and store the strain stocks requested by the Adam experiment design process. Possible reasons for loss of viability are that the library plates are not sufficiently sealed so that the defrost cycle of the freezer desiccates them, and plates thaw quickly when out of the freezer (but use of an insulating or cooling block tends to cause other problems because of condensation and icing). In addition, the partial thawing of the samples allows cells to settle in the wells; to counter this, cell density has been increased and picking depth now depends on the number of previous picks.

Growth in Liquid Media. Growing yeast in media and metabolite solutions allows the use of conventional liquid handling. However, the cells tend to settle to the bottom of wells, reducing access to oxygen and fresh nutrients. Shaking Liconic STX40 incubators are used but a completely uniform distribution of the cells throughout the 3-dimensional volume of liquid (bulk flow) is not attained. Experimenting with various shaking speeds and plate types revealed that a reasonably regular distribution of the cells near the well bottom is achieved with an orbital shaking speed of 200 rpm (amplitude 0.5 mm) in plates with flat-bottomed circular wells. The time between unloading plates from an incubator and reading them should be kept constant because OD increases with time because of cell settling. In addition, yeast cells can aggregate, so that OD readings depend on where the reader beam passes through the sample; a Variomag Teleshake in subsystem 3 helps counter this (10 s at 1000 rpm in a chaotic pattern).

Bacterial Infection and Cross-Contamination. A typical experiment lasts 4 days and it is critical to minimize opportunities for bacterial infection and cross-contamination. Precautions taken include a rigid transparent system enclosure fed by two HEPA ^d filters, adherence to aseptic technique when preparing solutions and connecting them to the system, and regular system cleaning using alcohol. All media contain the antibiotic G-418, all yeast strains are G-418 resistant, and the library of strains is held in deep well plates to minimize cross-contamination. All solid and liquid waste is disposed of away from the liquid handling decks, and opportunities for aerosol contamination are minimized. Finally, periodic tests are run using wild-type yeast to identify any emerging contamination issues.

Condensation and Frost. The three incubators in Adam are set at +30 °C for optimal yeast growth and are maintained at 90% humidity to prevent drying of the cultures. When a plate is unloaded into the cooler outside environment, condensation can develop under its lid, giving spurious OD readings and noisy growth curves. Removing lids before reading would risk contamination. Consequently, the temperature within the system enclosure is kept at around +28 °C, by a combination of the laboratory air-conditioning and the heat from the instruments within the enclosure, and condensation is minimized. However, this warm humid environment exacerbates the difficulty of maintaining the frozen state of the library plates; a cooling block designed to keep the library plate frozen outside the freezer suffered from condensation and icing, and library plates and their lids developed condensation which later froze, bonding lid and plate, and causing plates to jam on loading or unloading from the freezer. Omitting the lids causes desiccation of samples and risks contamination, whilst using re-sealable pierce-able film risks a sampling tip becoming stuck in a well and often leaves behind small droplets of sample at the seam, which then re-freeze.

Reliable Plate Handling. Existing standards ^{24 27} are obviously a major contribution to interoperability of plates, tip racks, and so on within automated systems. However, there is still considerable variation between microplate types in relation to optimal grasping positions for robotic handling, and grasping tends to rely on friction rather than positive location. Perhaps a future Society for Biomolecular Screening (SBS) standard will specify shallow depressions at defined points on the plate sides, so that positive grasping could be achieved using corresponding pins on gripper fingers.

The design of plate trays or drawers, varies widely between different items of equipment. They are often deep relative to the height of a microplate and are a “tight fit,” requiring plates to be grasped close to their upper rim, potentially reducing grasp reliability. Ideally a plate handling robot should be presented with an area larger than the microplate, onto which plates may be placed with wide tolerance; when the plate is taken into the instrument a locating mechanism should press the plate into one corner of the tray to provide precise positioning, which provides a precise location from which the plate is subsequently picked. Increasingly, manufacturers are incorporating such mechanisms but in our system only the ELx405 washer from Bio-Tek has such an arrangement and this provides very high pick–place reliability.

Originally, movement of microtiter plates from our Liconic incubators to the Molecular Devices readers was inadequately reliable. There are three such moves: one involving a Twister II arm and the subsystem 2 incubator and reader, the other two involving the IX Scara arm and the two subsystem 3 incubators and reader. Even with the excellent repeatability of the IX Scara robot arm (±0.01 mm) plates would often fail to locate correctly in the reader drawer, leading to jamming. Plates are a close fit in this drawer and although it has beveled guides, the lack of compliance in the Scara renders these ineffective. Consequently, plates have to be placed with sufficient accuracy to seat directly into the drawer. Investigation revealed that the incubator pick zones were slightly longer than the recess in the reader drawer, leading to variation in the grasped position which, in turn, led to inaccurate placement into the reader. End stops on the incubator unload trays were machined so that the length of the pick zone became equal to that of the recess in the reader drawer, leading to much higher reliability.

Many plate handling robots can pick and place reliably only where plate locations are horizontal, but few items of automation equipment have leveling screws. Leveling errors of a millimeter or two can be cumulative across a pick and place move and can reduce grasp reliability or “jolt” the plate and disturb its contents.

Conclusions

Few laboratory automation systems are as flexible or functionally complex as Adam. This level of diversity and flexibility is challenging from the point of view of achieving the essential high level of reliability. Nevertheless, Adam achieved its design goal of being the first machine to discover novel scientific knowledge. ²

Acknowledgments

The authors thank the reviewers for their helpful suggestions and comments. This work was funded by the U.K. Biotechnology and Biological Sciences Research Council (grant BBD00425X1), by a Society for Biomolecular Screening 2 award from the Higher Education Funding Council for Wales, and by fellowships from the Royal Academy of Engineering/Engineering and Physical Sciences Research Council, and by the Royal Commission for the Great Exhibition of 1851 for Dr. Amanda Clare.

Competing Interests Statement: The authors certify that they have no relevant financial interests in this manuscript.