<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dave Brown on Medium]]></title>
        <description><![CDATA[Stories by Dave Brown on Medium]]></description>
        <link>https://medium.com/@dave_15229?source=rss-157998f5c02d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*WRBSmv_u2JXQw6sSu-19cg.jpeg</url>
            <title>Stories by Dave Brown on Medium</title>
            <link>https://medium.com/@dave_15229?source=rss-157998f5c02d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 09:49:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@dave_15229/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Breaking Material Barriers to Lab Automation]]></title>
            <link>https://medium.com/@dave_15229/breaking-material-barriers-to-lab-automation-3c62f7f9cd9f?source=rss-157998f5c02d------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c62f7f9cd9f</guid>
            <category><![CDATA[lab-automation-systems]]></category>
            <category><![CDATA[community]]></category>
            <category><![CDATA[lab-robotics]]></category>
            <category><![CDATA[materials-science]]></category>
            <category><![CDATA[london]]></category>
            <dc:creator><![CDATA[Dave Brown]]></dc:creator>
            <pubDate>Wed, 25 Jun 2025 15:21:20 GMT</pubDate>
            <atom:updated>2025-06-25T15:29:14.634Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ctpaS6DCDAxlCfebi7R1Qg.png" /></figure><h4>2025/June/25</h4><p>I write this, not from the perspective of a material scientist, I am a biologist by training, however this was the title of our event yesterday at London Lab Automation Community (<a href="https://www.linkedin.com/groups/9844182/">#LLabAC</a>) event at Queen Mary University London. The idea, spawned together with <a href="https://www.linkedin.com/in/dr-moe-elbadawi/">Dr Mohamed Elbadawi</a> and <a href="https://www.linkedin.com/in/mojan-omidvar-a15aa8114/">Dr Mojan Omidvar</a> at <a href="https://www.qmul.ac.uk/media/news/2025/queen-mary-news/pr/queen-mary-recognised-as-top-global-university-.html">QMUL</a>, was to prompt discussion about the challenges of automating material science as opposed to the now sometimes impressively slick automation of drug screening. How can other fields reach the same standard? (Please share your own opinions in the comments)</p><p>Within our growing London Community the term Automation extends from robotic arms and liquid handling instruments, to controlled reactors and bioreactors to microfluidics and even to data pipelines. We are a broad inclusive community, whose members interests currently range from chemistry, and catalysis, to pharmaceuticals, cell biology, cell and gene therapy to material science. We set about building our event at QMUL around academic talks, and live instrument demos.</p><p>Arriving on the day, I had got myself into a bit of a festival mindset. The idea of bringing scientists and people who work to support scientific enterprises together had me thinking of crowds flocking to Stonehenge at the solstice (but without the traffic). Seeing the instrument in their flight cases gave me a slight Glastonbury vibe and had me reminiscing about playing in bands. I began to see the scheduled talks as a lineup, with <a href="https://www.linkedin.com/in/mojan-omidvar-a15aa8114/">Mojan</a> as our headline act.</p><p>As this was LLabAC’s 5th event, I have been getting to know the application specialists from our various sponsors and I am continually impressed by how smooth the instrument setups are. Catering was ready. As guests began to arrive it was exciting to welcome friends from our previous events, as well as those joining us for the first time. It seemed to be about 50:50. If you attended and I wasn’t able to welcome you, “Hello, thanks for coming. We’re glad to have you as part of our community”.</p><p>Between conversations, I overheard from the <a href="https://opentrons.com/">Opentrons</a> Flex stand that the deck of this liquid handling instrument is highly modular allowing users to customise the hardware to a wide range of use cases. Perhaps that seems like an obvious feature, it certainly sounded like a key requirement by the end of the evenings talks, but compared to the systems available during covid this is a significant step forward, and paired with the breadth and depth of Opentrons software from no code, quick start to total control via python and API feels like it addresses a core need I recognised leading an Assay Development team.</p><p>At the other end of the foyer the <a href="https://labmanautomation.com/">Labman</a> MultiDose cycled through powder handling tasks — a robot arm selecting one of 56 powder stocks from a carousel and weighing specific amounts into vials from up to 6 different SBS plate positions. It is likely true of other lab automation vendors too, but having visited Labman HQ, and having met our speaker and MultiDose expert <a href="https://www.linkedin.com/in/sahilgazula/">Sahil Gazula</a> before, I was very aware of how he had worked with this instrument through its design and optimisation phases. It is like hearing a musician play an instrument they designed and built themselves.</p><p>Some guests arrived (if this was you, please say hello in the comments and give us more detail) who were interested in water. My first thought was water testing or waste water treatment (again, I’m a biologist) but they were interested in electrolytic water splitting. Every time a new focus becomes part of this community I am delighted. I could be wrong, but of the instruments we had on site, this made me think of the <a href="https://formulatrix.com/">Formulatrix</a> Mantis for its ability to rapidly dispense low volumes using a precise volume diaphragm. This was the first time Formulatrix have sponsored one of our events, and having used the Mantis myself for cell culture media optimisation it was great to have the opportunity to showcase this instrument. Its ability to select up to 26 different diaphragm chips from a rack and non-contact dispensing make it visually and functionally distinct from the majority of liquid handling instruments.</p><p>Despite busying myself attempting to front of house our event, I was treated to hearing the <a href="https://www.sptlabtech.com/">SPT Labtech</a> XL20 explained to an IT teacher (if this was you please let me know in the comments). To me this instrument resembles a scaled down model of a shipyard with a robotic crane arm, transferring containers between different locations and weigh stations. The SPT demo is brightly colour coded to emphasise the location swaps. On seeing the tiny QR code printed on the underside of the sample tubes and hearing that the arm takes each tube to be scanned en route to its new location, it felt the IT teacher instantly had a mental model of the database in his head.</p><p>In no time, we were on to the talks. Nervous about the second talk being delivered remotely from Glasgow, I set about introducing <a href="https://www.linkedin.com/in/dr-moe-elbadawi/">Mohamed</a>, butchering his last name under pressure despite promising myself I would not (sorry). I had come to know him as a champion of ironing out the kinks of our first visit to QMUL, and through a mutual desire to help students build their skills and networks and unlock great ways to contribute. But what would his talk cover? Moe set the whole evening up brilliantly. He had a slide showing the dramatic increase of “Lab Automation” publications per year over time, in contrast to the weak increase in “Pharmacy or Pharmaceutical Lab Automation” publications. Lab Automation in general is exciting and progressing, but in some fields that momentum is not tangible. Moe’s theme was the low-effort high-reward value addition of automating menial tasks. 1. Automating the literature search for targets, and 2. Automating the preparation of samples, in his case tablets. He was able to test whether mixtures would cure to form a polymer at high throughput using liquid handling robots.</p><p>Next up we hand <a href="https://www.linkedin.com/in/bingyuan-lu-51a6b511b/">Dr Bingyuan Lu</a> a postdoc with Professor Lee Cronin at the University of Glasgow. Thankfully, the remote talk worked fine and we are able to hear his presentation which detailed the development of microfluidic sensors and their deployment in environmental testing and personalised medicine. The main takeaway for the Assay Developer in me, was that the serpentine path of the microfluidic channel can drive rapid mixing of the molecules within the droplets. This rapid and consistent mixing eliminates the need for vortexing which would be required to mix reactions performed manually. This is an incredibly space- and energy- efficient way of mixing, no robot arms are needed in this instance. This allows for wearable devices and remote function (with finite power supply). We were also treated to a prelude of Bingyuan’s work using a liquid handler to probe the origin’s of life (please add links in the comments).</p><p>Following 3D printing and microfluidics we had a 5 minute talk from Labman’s <a href="https://www.linkedin.com/in/sahilgazula/">Sahil Gazula</a> on powder dispensing followed by a 10 minute talk by <a href="https://www.linkedin.com/in/danlorduk/">Dan Lord</a> on Opentrons’ open source software approach. These talks were focused and approachable and we had informed questions from the audience after each. I hope this left the students in the audience with an impression of what is available to help them overcome technical hurdles in academia or industrial projects in the future (please let me know in the comments).</p><p>To cap off the evening, <a href="https://www.linkedin.com/in/mojan-omidvar-a15aa8114/">Dr Mojan Omidvar</a>’s talk was full of energy and clarity. We saw the detail of the material preparation and material characterisation processes she has automated in the pursuit of specialised conductive materials for 5G and 6G antenna. She framed this within the challenges of material science as a whole, highlighting the challenge of integrating data from ≥5 material property tests employing different approaches on different instruments, which certainly captured my attention as a <a href="https://www.fulgens-consulting.com/">digitisation consultant</a>. As a biologist I recognise this as a mesoscale problem. It is much harder to investigate interactions between large molecular complexes and subcellular compartments than fewer smaller scale molecules or between whole cells. Similarly, cells function and organ function are easier to study than tissue function which sits between the two. As she ended her talk, with a summary of challenges and her forecast for the automation of material science by head began to swim with the evening’s new information. I believe, like any good headline performance, it ended with a job offer. You can form a queue.</p><p>The talent and experience in this community is astounding please reach out if you’d like to get involved, or join our <a href="https://lu.ma/LLabAC">luma event calendar</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c62f7f9cd9f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lab Protocol Evaluation with free LLM ChatBots]]></title>
            <link>https://medium.com/@dave_15229/lab-protocol-evaluation-with-free-llm-chatbots-14b473f49f33?source=rss-157998f5c02d------2</link>
            <guid isPermaLink="false">https://medium.com/p/14b473f49f33</guid>
            <category><![CDATA[chatbots]]></category>
            <category><![CDATA[covid-19-testing]]></category>
            <category><![CDATA[lab-automation-solutions]]></category>
            <category><![CDATA[protocol]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Dave Brown]]></dc:creator>
            <pubDate>Tue, 25 Feb 2025 09:41:43 GMT</pubDate>
            <atom:updated>2025-03-15T15:02:17.360Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>The main focus of my consultancy is to improve scientific processes for my clients. I do this by considering how judicious use of automation and careful tracking of samples and the appropriate metadata that accompanies them can strengthen the dataset they end up with.</p><p>At a recent London Lab Automators’ Community event, a talk by <a href="https://www.linkedin.com/in/elyor-kodirov-7547b666?miniProfileUrn=urn%3Ali%3Afs_miniProfile%3AACoAAA4AG7ABzX5MslrXbKlyLQLKpzSBGjzHBfg">Elyor Kodirov</a> detailed how large language model (LLM) chatbots can allow users to program liquid handling robots with simple instructions in every day language. I had just read a <a href="https://www.synthace.com/blog/chatgpt-for-biology-experiment-design-and-automation">great article</a> on this topic written by <a href="https://www.linkedin.com/in/chris-grant-188b127/">Chris Grant</a> co-founder of Synthace. Although in both these cases, chatbots were carefully and sensibly integrated with dedicated scientific software, it did get me wondering. How ready are LLMs (particularly free versions) to handle the complexity and specificity of scientific experiments? I wanted to see for myself.</p><h3>Case Study</h3><p>During the covid-19 pandemic, I was able to volunteer carrying out diagnostic PCR tests at the CLIA-HUB lab, which was rapidly established to provide testing for the community. In my case, the sudden unsought ability to effectively help the community around me using my training in molecular biology was mind opening. The scientists around me seemed to share a deep respect for our opportunity and obligation to do our bit, something our colleagues in medicine have done unceasingly before and since. We each sought to set the very highest standard for science, but also for best practice. One result of this striving is the fact that the qPCR testing protocol was totally opensource with the aim of being as transparent as possible.</p><p>I thought this protocol would make an ideal case study for my understanding of LLMs because; the protocol was peer reviewed and published, I didn’t write it but I have direct experience performing it, and finally because covid-19 and its diagnosis was such a prominent feature of so many lives for so long.</p><h3>Method</h3><p>I wondered what 3 popular chatbots would make of this protocol. This was my prompt.</p><blockquote>Hello. This was the protocol used for qPCR covid-19 testing. Please evaluate it? Can you identify any innovations or weak spots, mistakes, scientific inconsistencies or risks?</blockquote><p>I also attached a .pdf of the protocol which you can find at <a href="https://www.notion.so/Article-Draft-199058d7554e805b9d12c804e1d1f686?pvs=21">this doi</a>. I had to use the link for Gemini rather than the pdf which was not supported by the free version at the time.</p><p>Although my approach is rough and ready, I decided to run each chatbot three times, and compare the responses.</p><p>I immediately noticed some inconsistency between versions using the same AI, and some discrepancies between AIs. As an experienced user of the protocol myself I immediately spotted that Gemini response “v9” completely imagined (or hallucinated) the involvement of the ThermoFisher KingFisher platform which was not used in our case.</p><p>I wanted to know how well did these chatbots appear to understand the protocol? How helpful were their suggestions? Once the nine separate response were returned, I sought to compare them. I did try using a chatbot to do this task, but the formatting was a headache and it did not do a great job of identifying and clustering similar points made by different responses. Pausing my undying love of making things scalable I dove into the weeds myself to see what these bots had to say about our protocol. If anyone does want to have a go at that part of the task themselves I am happy to share the 9 separate responses along with my combined dataset.</p><h3>Results</h3><p>The majority of comments made by the bots had some grounding in fact. Although subjective, I scored 71 of the 80 non-redundant comments (89%) as relevant. Comments that were not deemed relevant ranged from comments that were incorrect saying the thermocycler conditions were not provided, to misunderstanding the assay, to full blown hallucination of instruments or techniques that were not described in the protocol. The goal of the diagnostic assay used by CLIAHUB was not to quantify the absolute amounts of viral RNA present, yet some responses pointed out the lack of standard curves or RNA normalisation steps that should definitely be applied in some cases but not in this context.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*jOTYvb1TA7UI8bQ-Trm7lQ.png" /></figure><p>Figure 1. A heatmap of 80 non-redundant topics raised in chatbot responses recognised as a positive (blue) or a negative (red). The mean of all 9 responses is shown in the rightmost column (v1–9). v1 to v3 were responses from ChatGPT 4o, v4 to v6 were Claude 3.5 Sonnet and v7 to v9 were Gemini 1.5 flash.</p><p>Of the 71 relevant non-redundant comments, 63 were deemed helpful (89% again) either as critique of the protocol described, or as opportunity for helpful improvements of the protocol. For example, multiple responses correctly identified the removal of nasopharyngeal swaps and transfer of liquid from the patient sample tubes to 96 well plates as the riskiest point both in terms of potential aerosolization of virus but also of cross contamination. Perhaps obvious but certainly true.</p><p>Overall, responses were diverse with only 1.86 out of 9 responses agreeing on each topic on average. Most topics were only raised by a single response which is evident in the sparsity of the heatmap. Seven of the 80 topics were disputed among the responses for example Gemini’s responses differed in a glass half-full or half-empty argument over whether the details of the reagents were complete enough. A table in the CLIA-HUB protocol does have missing values, because for example some instrument models to not have a product number. Similarly ChatGPT disputes the sufficiency for sample storage instructions, although two Gemini responses support the argument for insufficiency on this topic.</p><p>The fiercest critique surrounded the Ct cuttoff, the threshold at which a sample is scored as positive for covid viral gene and or the human RNase P gene used as a positive control. Some responses said the cuttoff was too strict, others too lenient. With so many rules of thumb floating around in the datasets these LLMs were trained on, it is understandable that they did not reach a consensus. Many responses correctly pointed out opportunities for optimisation of the process. From the reaction volumes, the RNA elution efficiency, to checks for primer cross-reactivity. Others would extend the utility of this protocol to include other corona viruses, or samples in alternative transport buffers. Finally there were many suggestions to improve the write up, by including additional details.</p><h3>Conclusions</h3><h3>About the covid-19 diagnostic protocol.</h3><p>All three of the chatbots tested agree that this protocol is/was innovative, highlighting the use of liquid handling robots, and the considered use of multiple controls with a clear decision tree to identify covid positive patients among those experiencing covid like symptoms. I was relieved if disappointed that the chatbots didn’t scoff at our process and present an unarguably superior approach that could be described in a few short paragraphs.</p><h3>About using free AIs to critique a protocol</h3><p>It is definitely worth gathering multiple opinions, particularly when trying to generate a checklist of things a protocol should include through this sort of consensus building. No doubt models with reinforcement learning will present better suggestions with more context, but to me that loses the appeal of the generality. Science advancing faster in organisations with more money (on average) was the status quo before the AI revolution, if that remains the case, then in my eyes, it’s definitely not as cool as it could be.</p><h3>About prompt engineering</h3><p>Until this project I found the concept of prompt engineering unpleasant for some reason, maybe just the name. Perhaps because asking the right question was such a fundamental part of my training as an academic scientist that it feels insulting. Still, this small testcase has given me valuable direct experience and helped to lower my reactive reluctance to learn more about prompt engineering from the emerging experts. If you generate expert content on prompt engineering, please do share links in the comments.</p><p>Perhaps the most obvious lesson was “Don’t ask for multiple evaluations in one prompt”. Considering that I planned to try to compare chatbots with one another a more focused prompt would have offered a better test.</p><blockquote>Hello. This was the protocol used for qPCR covid-19 testing. Please identify any weak spots. Please provide each potential weak spot with a numeric score from 0–1 representing the effort to impact ratio, with 1 as maximum impact for minimal effort and 0 as minimal impact for maximum effort, and justify each score.</blockquote><p>If you are interested on how this second round worked out, let me know in the comments and I can write a follow up. Thanks for reading.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=14b473f49f33" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Avoiding Mixture Recipe Duplicates in Electronic Lab Notebooks]]></title>
            <link>https://medium.com/@dave_15229/avoiding-mixture-recipe-duplicates-in-electronic-lab-notebooks-61ddaa671845?source=rss-157998f5c02d------2</link>
            <guid isPermaLink="false">https://medium.com/p/61ddaa671845</guid>
            <category><![CDATA[laboratory]]></category>
            <category><![CDATA[benchling]]></category>
            <category><![CDATA[duplicate-data]]></category>
            <category><![CDATA[electronic-lab-notebook]]></category>
            <dc:creator><![CDATA[Dave Brown]]></dc:creator>
            <pubDate>Thu, 06 Feb 2025 23:11:08 GMT</pubDate>
            <atom:updated>2025-02-10T09:02:44.067Z</atom:updated>
            <content:encoded><![CDATA[<p>A Benchling Community member asked a question about preventing registration of duplicate recipes. I thought I would answer here because not everyone is a member of the Community, and because the concepts are generally applicable to electronic lab notebook inventories beyond Benchling. <a href="https://community.benchling.com/ask-the-community-6/how-to-prevent-duplicate-mixture-buffer-recipe-registrations-1376?postid=5051#post5051">Here</a> is a link to the Community thread.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*D3mqqZO4sBHnKru88LPnZg.png" /><figcaption>A key advantage of digitisation of lab science is the ability to group and compare results. It is important that attempts to make the same recipe are considered as a single group, and that duplicate records of the same recipe are not introduced without a good reason. We want to block the examples marked with a red cross as duplicates of the recipe to their left.</figcaption></figure><p>To find mixtures that use exactly the same components in exactly the same ratios I did the following. You can do these steps either by exporting all mixtures into a spreadsheet or by using an API.</p><p>Please comment if you would like more detail on how to do the following using the Benchling API.</p><p><strong>Get the data</strong></p><p>Get a complete set of mixtures and their ingredients (grouped by the mixture they form).</p><p><strong>Sort</strong></p><p>Sort the data so that mixtures with the same ingredients in a different order will now have the same ingredients in the same order.</p><p><strong>Calculate ratios</strong></p><p>Calculate the ratio of ingredients in the sorted order. This can be done for the target amounts and/or the measured amounts.</p><p><strong>Label duplicates</strong></p><p>Most simply, mixtures with the same ingredients in the same amounts are “<em>exact duplicates</em>”. Because we have calculated and compared ratios rather an absolute amounts, we can also label mixtures with the same ingredients and the same ratios as “<em>scale duplicates</em>”.</p><p>But it is not that simple. We need to consider the precision of our ratios. Because users will likely enter amounts to a sensible degree of precision recipes for large and small volumes may use the same recipe inexactly. This group could also be flagged as “<em>near duplicates</em>” or “<em>precision duplicates</em>” including a precision tolerance in the label may help, e.g. “<em>precision 1e-5 duplicate</em>”.</p><p>It is worth scanning for “<em>unit duplicates</em>” this might require conversion between mass and moles for example before comparing ratios. It is entirely possible to specify the exact same mixture produced by different methods. Low dilutions of salts might be achieved from a small solid mass or by serial dilution. We can identify these “<em>method duplicates</em>” if we calculate the final concentrations of all components in our mixtures. Comparing the ratios of these concentrations and units will then reveal identical mixtures produced by distinct routes.</p><p><strong>Further considerations</strong><br>But you could go on to consider inexact matching of ingredients for example is NaCl from two different suppliers the same or different? Mixtures making the same thing from different sources, “<em>source duplicates</em>”, may be deliberate, or it may indicate siloed work habits with different users or groups working differently.</p><p><strong>Next steps</strong><br>Once you have identified the various types of duplicates you are interested in you can decide which ought to be eliminated and which are worth keeping. It is worth checking the usage of each mixture in your database and whether the mixtures you plan on deleting have detailed instructions. Perhaps one duplicate has instructions and the other doesn’t. It might be better to merge than delete. Consider a plan for deprecation of duplicates rather than immediate deletion which may effect your results tables etc.</p><p>Just to be clear. At no point did I encourage anyone to delete or archive anything without care consideration and discussion with your data team.</p><p>I believe the Benchling API can build apps to do these checks routinely, or to check each new mixture upon creation and warn users if they fall into any of these categories. This is just how I think about it. I hope that helps.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61ddaa671845" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>