I started this blog a few years ago when I began my journey studying immunology, to have a place to write about what I was learning and about applications of AI to the field. My studies were a success: I completed a Masters in Microbiology-Immunology, explored interesting applications, and met some awesome people along the way. At the end of last year, I decided to return to my origins and re-joined fast.ai, which had grown into Answer.AI while I was away. Why now? AI is an increasingly hated field, and for good reason. I wrote about my decision: There are tons of false and over-hyped claims about what AI can do, but underneath those, it is a genuinely useful technology. I want to use AI in ways that don’t degrade my own critical thinking skills and that don’t pretend that resources are unlimited. The biggest companies have made their vision for the technology seem like the only option. It isn’t. OpenAI, Anthropic, Google, and xAI don’t own AI. Their values and decisions aren’t the only version of what AI technology can be. I understand why many people are anti-AI. My new job is rooted in the same objections. I returned to the field because I want to figure out how to use AI in ways that protect human creativity, autonomy, and problem-solving. You can read my full essay here. I am excited to continue exploring how to design AI tools that better encourage human understanding and human autonomy, particularly in the area of education. I have written several essays in the last year, posted at fast.ai and answer.ai, about how AI intersects with education, coding, and science: Breaking the Spell of Vibe Coding Stop Saying Boredom is Good for Kids I Don’t Want a Learning Dashboard for My Child Risks and Limitations of AI in the Life Sciences My friends all hate AI; I just joined an AI startup Going forward, you can find me at fast.ai, answer.ai, on twitter, and on LinkedIn. I look forward to reading your responses. Create a free GitHub account to comment below.

As a former mathematician, I was used to nobody reading what I wrote. So when I first began blogging in 2015, I never expected that several of my blog posts would go viral or to have multiple journalists contact me (including from NPR, Wired, and Fortune), make the front page of Hacker News (over 10 times), receive conference keynote invitations, and be interviewed on podcasts. I do not consider myself a “natural” writer. In college, I tried to avoid classes that required essays, because writing was a struggle for me. It wasn’t until I was 30 that I set out to practice writing more. I share tips I use for blogging here, which include being willing to put a lot of time into a single post, incorporating high quality information, and having a clear idea of my intended audience. I have selected some of my most popular and impactful posts below. Several of these were originally posted on Medium or fast.ai (the two sites where my writing used to live). They are grouped into clusters based on theme. I hope you might enjoy reading these if you haven’t seen them before! Challenging Conventional Wisdom Questioning widely-held assumptions about tech culture, education, and health has been the basis for several popular posts. If you think women in tech is just a pipeline problem, you haven’t been paying attention (2015) In 2015, I felt burnt out and disillusioned by my experiences working in tech. I was frustrated with how much the popular conversation was still focused on “the pipeline problem”: training young girls to code while ignoring all the adult women being driven out of the tech industry by mistreatment. I spent 9 months researching and writing this post. It went viral and remains my most popular essay. This, together with my other posts, led to me being interviewed and quoted in Wired several times regarding diversity in tech (as well as other AI topics). My first post Trends to Avoid When Founding a Startup (2018) The dominant narrative for Bay Area tech startups is to try to raise venture capital, achieve exponential hypergrowth, and hire lots of computer science PhDs. I argued that these approaches not only harm employees, but lead to weaker companies and worse products. My family’s unlikely homeschooling journey (2022) Many people hold a stereotyped and outdated view of homeschooling, not realizing the explosion of innovative, non-traditional education options available in recent years. My husband and I never planned to homeschool, but we unexpectedly found that our child thrives with this approach. Your Immune System is Not a Muscle (2024) The misleadingly named “Hygiene Hypothesis” is often used to justify the misconception that all microbes are good for us. However, this theory is more accurately reframed as the “Old friends hypothesis”: humans co-evolved with friendly bacteria and some parasites. We did not co-evolve with the crowd infections of mega-cities and 100,000 global flights per day. AI Beyond Elite Institutions Machine learning isn’t just for those at billion dollar companies. These posts highlight unconventional practitioners and offer practical guidance for people in varied domains. Deep Learning: Not Just for Silicon Valley (2017) Our goal at fast.ai is making AI accessible to people outside of elite institutions, who are tackling meaningful problems in low-resource areas. This post introduced some of our earliest international fellows and the diverse range of problems they were working on. I always enjoyed writing about fascinating use cases from our deep learning community. How (and why) to create a good validation set (2017) An all-too-common scenario: a seemingly impressive machine learning model is a complete failure when implemented in production. Advice on one common culprit of this, and how to avoid it. In the early years of fast.ai, I wrote numerous posts with practical advice for machine learning. This article asked the question, “Can A.I. conquer its Excel problem? An Introduction to Deep Learning for Tabular Data (2018) Deep learning is not just for images and text. Companies such as Pinterest and Instacart are also applying it to tabular data, the type of data you might normally put in a spreadsheet. This post caught the attention of a reporter with Fortune, who ended up interviewing me and writing about the topic here. Debunking AI Hype & Holding Tech Accountable The narratives about AI put forth by major tech companies are often misleading about what is necessary, what values matter, and what types of harms can result. Google’s AutoML: Cutting Through the Hype (2018) In a 3-part series, I countered claims that all data scientists need customized, bespoke neural network architectures. While I was nervous about disagreeing with both Google’s CEO and head of AI, my posts led to an invitation to keynote the prestigous ICML AutoML workshop. Seven years later, my critiques have been proved valid, with transfer learning a cornerstone of ML and automated neural network search not commonly used. By 2023, we were supposed to all be using AutoML neural architecture search Five Things That Scare Me About AI (2019) AI ethics is not just a theoretical topic. I was (and still am) alarmed about the harms already being caused to human beings by AI systems irresponsibly applied to healthcare, employment decisions, policing, and more. The Problem with Metrics is a Big Problem for AI (2019) Overemphasizing metrics leads to a variety of real-world harms, including manipulation, gaming, and a myopic focus on the short-term. AI is metric optimization on steroids. I later turned this blog post into an academic paper, together with David Uminsky. Two disturbing case studies I keep returning to are how computerized algorithms have been used to cut healthcare and to fire teachers Deep learning gets the glory, deep fact checking gets ignored (2025) A microbiologist discovered hundreds of errors in a paper that used AI to classify enzymes. This is a case study of how challenging it can be to evaluate AI claims outside our area of expertise, as well as of the misaligned incentives that reward flashy results, but not diligent fact-checking. This has been by far my most popular post on LinkedIn. Immunology & Science Decoding T cells with AI (2024) T cells are a crucial component of the adaptive immune system. Accurately pedicting what they can bind to would impact a range of treatments. Numerous algorithms have been developed for this question, but the problem is far from solved. The surface of a T cell. I’ve enjoyed exploring how AI is being applied to immunology Scientists Just Connected the Dots Between Viruses and… Everything (2025) For a long time, catching frequent viruses was considered both inevitable and harmless. But it turns out that common, seemingly-mild viruses have disturbing long-term health impacts. A thread about viruses Thanks for joining me on this walk through the past! Also, you can subscribe to be notified of new blog posts by submitting your email below: I look forward to reading your responses. Create a free GitHub account to comment below.

Deep learning is glamorous and highly rewarded. If you train and evaluate a Transformer (a state-of-the-art language model) on a dataset of 22 million enzymes and then use it to predict the function of 450 unknown enzymes, you can publish your results in Nature Communications (a very well-regarded publication). Your paper will be viewed 22,000 times and will be in the top 5% of all research outputs scored by Altmetric (a rating of how much attention online articles receive). However, if you do the painstaking work of combing through someone else’s published work, and discovering that they are riddled with serious errors, including hundreds of incorrect predictions, you can post a pre-print to bioRxiv that will not receive even a fraction of the citations or views of the original. In fact, this is exactly what happened in the case of these two papers: Functional annotation of enzyme-encoding genes using deep learning with transformer layers | Nature Communications Limitations of Current Machine-Learning Models in Predicting Enzymatic Functions for Uncharacterized Proteins | bioRxiv A Tale of two Altmetric Scores This pair of papers on enzyme function prediction make for a fascinating case study on the limits of AI in biology and the harms of current publishing incentives. I will walk through some of the details below, although I encourage you to read the papers for yourself. This contrast is a stark reminder of how hard it can be to evaluate the legitimacy of AI results without deep domain expertise. The Problem of Determining Enzyme Function Enzymes are what catalyze reactions, so they are crucial for making things happen in living organisms. Enzyme Commission (EC) numbers provide a hierarchical classification system for thousands of different functions. Given a sequence of amino acids (the building blocks of all proteins, including enzymes), can you predict what the EC number (and thus, the function) is? This seems like a problem that is custom-made for machine learning, with clearly defined inputs and outputs. Moreover, there is a rich dataset available, with over 22 million enzymes and their EC numbers listed in the online database UniProt. An Approach with Transformers (AI model) A research paper used a transformer deep learning model to predict the functions of enzymes with previously unknown functions. It seemed like a good paper! The authors used a reasonable, well-regarded neural network architecture (two transformer encoders, two convolutional layers, and a linear layer) that had been adopted from BERT. They looked at regions with high attention to confirm that these were biologically significant, which suggests that the model had learned underlying meaning and provided interpretability. They used a standard training, validation, and test split on a dataset with millions of entries. The researchers then applied the model to a dataset where no “ground truth” was known to make ~450 novel predictions. For these novel predictions, they randomly selected three to test in vitro and confirmed that the predictions were accurate. A transformer model, shown on the left, was used to predict Enzyme Commission numbers for uncharacterized enzymes in E. coli. Three of these were tested in vitro (Fig 1a and Fig 4 from Kim, et al.) The Errors The Transformer model in the Nature Communications paper made hundreds of “novel” predictions that are almost certainly erroneous. The paper had followed a standard methodology of evaluating performance on a held-out test set, and did quite well on that (although later investigation suggests there may have been data leakage). The results claimed for enzymes where no ground truth is known were full of errors. For instance, the gene E. coli YjhQ was predicted to be a mycothiol synthase, but mycothiol is not synthesized by E. coli at all! The gene yciO, which evolved from the gene TsaC, had already been shown a decade earlier in vivo to not have the same function as TsaC, yet the Nature Communications paper concluded it did have the same function. Of the 450 “novel” results given in the paper, 135 of these results were not novel at all; they were already listed in the online database UniProt. Another 148 showed unreasonably high levels of repetition, with the same very specific enzyme functions reappearing up to 12 times for genes of E. coli, which biologically implausible. Most of the “novel” results from the transformer paper were either not novel, unusually repetitious, or incorrect paralogs (Fig 5 from de Crecy, et al.) The Microbiology Detective How did these errors come to light? After the model had been trained, validated, and evaluated on a dataset involving millions of entries, it was used to make ~450 novel predictions, and three of these were tested in vitro. It just so happens that one of the enzymes selected for in vitro testing, yciO, had already been studied extensively over a decade earlier by Dr. de Crécy-Lagard. When Dr. de Crécy-Lagard read that deep learning had predicted that yciO had the same function of another gene, TsaC, she knew from her long years in the lab that this was incorrect. Her previous research had shown that the TsaC gene is essential in E. coli even if yciO is present in the same genome and even when yciO gene is overexpressed. Moreover, the yciO activity reported by Kim et al. is more than four orders of magnitude (i.e. 10,000 times) weaker than that of TsaC. All this suggests that yciO does NOT serve the same key function as TsaC. Two enzymes with a common evolutionary ancestor, but different functions (Fig 7 from de Crecy, et al.) YciO and TsaC do have structural similarities, and YciO evolved from an ancestor of TsaC. Decades of research on protein and enzyme evolution have shown that new functions often evolve via duplication of an existing gene, followed by diversification of its function. This poses a common pitfall in determining enzyme function, because the genes will have many similarities with the ones they duplicated and then diversified from. Thus, looking at structural similarities is only one type of evidence for considering enzyme function. It is also crucial to look at other types of evidence, such as neighborhood context of the genes, substrate docking, gene co-occurrence in metabolic pathways, and other features of the enzymes. It is important to look at multiple types of evidence when classifying enzyme function (Fig 2 from de Crecy, et al.) Hundreds of Likely Erroneous Results Spotting this one error inspired de Crécy-Lagard and her co-authors to take a closer look at all of the enzymes found to have novel results in the Kim, et al, paper. They found that 135 of these results were already listed in the online database used to build the training set and thus not actually novel. An additional 148 of the results contained a very high level of repetition, with the same highly specific functions reappearing up to 12 times. Biases, data imbalance, lack of relevant features, architectural limitations, or poor uncertainty calibration can all lead models to “force” the most common labels from the training data. Other examples were proven wrong via biological context or a literature search. For instance, the gene YjhQ was predicted to be a mycothiol synthase but mycothiol is not synthesized by E. coli. YrhB was predicted to synthesize a particular compound, which was already predicted to be synthesized by the enzyme QueD. A form of E. coli with a QueD mutant was unable to synthesize the compound, showing that this is not in fact the function of YrhB. Rethinking Enzyme Classification and “True Unknowns” Identifying enzyme function actually consists of two quite different problems which are commonly conflated: propagating known function labels to enzymes in the same functional family discovering truly unknown functions The authors of the second paper observe, “By design, supervised ML-models cannot be used to predict the function of true unknowns.” While machine learning can be useful for propagating known functions to additional enzymes, there are many types of errors that can occur: including failing to propagate labels when they should, propagating labels when they should not, curation mistakes, and experimental mistakes. Unfortunately, erroneous functions are being entered into key online databases such as UniProt, and this incorrect data may be further propagated if it is used to train prediction models. This is a problem that increases over time. Need for Domain Expertise It is not news that AI work will be more highly rewarded and supported than work that closely inspects the underlying data and integrates deep domain knowledge. The aptly titled “Everyone Wants to do the Model Work, not the Data Work” paper involving dozens of machine learning practitioners working on high-stakes AI projects and found that inadequate-application domain expertise was one of a few key causes of catastrophic failures. Sources of cascading failures in machine learning systems (Fig 1 from Sambasivan, et al.) These papers also serve as a reminder of how challenging (or even impossible) it can be to evaluate AI claims in work outside our own area of expertise. I am not a domain expert in the enzyme functions of E. coli. And for most deep learning papers I read, domain experts have not gone through the results with a fine-tooth comb inspecting the quality of the output. How many other seemingly-impressive papers would not stand up to scrutiny? The work of checking hundreds of enzyme predictions is less glamorous than the work of building the AI model that generated them, yet it is even more important. How can we better incentivize this type of error-checking research? At a time when funding is being slashed, I believe we should be doing the opposite and investing even more into a range of scientific and biomedical research, from a variety of angles. And we need to push back on an incentive system that is disproportionately focused on flashy AI solutions at the expense of quality results. Related Reading: The problem with metrics is a big problem for AI Gaps and Risks of AI in the Life Sciences “AI will cure cancer” misunderstands both AI and medicine You can subscribe to be notified of new blog posts by submitting your email below: I look forward to reading your responses. Create a free GitHub account to comment below.

DNA sequencing hasn’t lived up to the hype Twenty to thirty years ago, politicians, scientific leaders, journalists, and even Nobel laureates predicted that sequencing the human genome would revolutionize how we treat disease. And while the advances in DNA sequencing that have occurred since then have improved recognition and treatment for some cancers and rare diseases, on the whole the field has not lived up to earlier hype. Time Magazine covers from 1994 and 1999 about genetics In an article titled “Why sequencing the human genome failed to produce big breakthroughs in disease”, a biology professor highlights that most common diseases are not caused by a single gene. In fact, common diseases are often linked to hundreds of gene variants, and even collectively, these variants still account for only a small fraction of disease variance. Here, I want to focus on two other key limitations of DNA sequencing, and how they are now being addressed with new approaches. What DNA can’t tell us First, DNA can’t answer many questions about how cells and organisms work in practice. A neuron in the brain has the same DNA as a liver cell, yet the two have completely different functions. This is because different segments of DNA are turned off or on in different cells. To understand how cells are actually working, you need to know about proteins and RNA (RNA is the intermediary which translates DNA into protein). Proteins are what build the structure of cells, catalyze chemical reactions within the cell, and allow communication between cells. Healthy and unhealthy cells in the same organism will usually have the same DNA. For instance, if some regions of the intestines are experiencing an IBD (irritable bowel disease) flare and others aren’t, they would all have the same DNA, yet likely different RNA and protein levels. A second big problem is that many key sequencing techniques destroy spatial information. You essentially may have to put tissue or cells into a blender in order to get rich information about DNA or RNA sequences. While this data is informative, it turns out that locations of cells within a piece of tissue, and locations of regions within a cell, are also very important! Again, considering the case in which some regions of the intestine are inflamed due to IBD, yet others aren’t, mixing them all together in a blender will lose or distort useful information. Look at the difference between crypts in a healthy segment of the colon (on the left) compared to inflamed crypts (on the right). Differences between a healthy bowel (on the left) and an inflamed bowel (on the right). Source: mypathologyreport.ca Recently, we have seen a rise in breakthroughs that allow us to obtain data about location. Spatial techniques are a necessary and exciting step beyond DNA sequencing. The power of spatial information showed up as a major theme at a conference I attended last year, and spatial techniques have been recognized by Nature Methods as “Method of the year” twice in the last 5 years. A Few Major Areas of Innovation What is genetic sequencing anyway? There are a number of different types of sequencing that have been invented in the last 30 years. Walking through a brief history will illustrate what these technologies are, and what they can and cannot do. To make it concrete, let’s look at the example of how they have been applied to cancer treatment. Sanger Sequencing: This is an older technology dating back to the 1970s, and which was the main way of sequencing DNA up until 2005. Sanger sequencing was one of the methods used in the mid-90s to identify the genes BRCA1 and BRCA2 as key genetic risk factors for breast cancer. The process of discovering BRCA1/BRCA2 involved scientists slowly zeroing in on their chromosomal locations over a period of years. Several other cancer genes were discovered during this time period as well. While Sanger sequencing is effective on smaller amounts of DNA, it can be quite slow to deal with larger volumes. It took over 10 years to sequence the first copy of the human genome using Sanger Sequencing. It is still used today as a simple and reliable way to test for known mutations (such as BRCA1/BRCA2) or for smaller tasks. High-Throughput Sequencing: New technology released in the mid-2000s allowed DNA to be cut into lots of short pieces and for millions of pieces to be sequenced in parallel at once. This approach, called high-throughput sequencing, was significantly faster than Sanger sequencing. High-throughput sequencing has many applications, including to cancer treatment, by making it cheaper and faster to sequence DNA to identify particular mutations which can influence treatment decisions. Method of the Year | Nature Methods Long-read sequencing: High-throughput sequencing has the advantage of high accuracy, but the downside of short sequence lengths. In the 2010s, technologies were released with the opposite set of strengths and weaknesses. Long-read sequencing provides the advantage of long sequence lengths, although the downside of lower accuracy. To compare, high-throughput sequencing uses DNA strands that are a few hundred base pairs long, whereas long-read sequencing uses DNA strands that are tens of thousands base pairs long. Both technologies have different strengths and are widely used today. Single-Cell Sequencing: High-throughput sequencing involves sequencing the DNA of many cells at once, but sometimes it is useful to sequence individual cells. In a tumor, different cells can have different mutations. It is possible that a small subset of the mutations may drive metastasis (the spread of cancer to other areas) or resistance to treatment. Identifying these driver mutations can guide treatment decisions, since particular driver mutations can predict the effectiveness of various drugs. Key mutations may be drowned out in the average if you sequence the entire tumor. This is one reason why it is useful to be able to sequence single cells, and not just obtain the average of many cells. Single-cell sequencing was selected as Nature Methods method of the year in 2013. A Lego interpretation of bulk RNA-seq; single-cell RNA-seq; spatial transcriptomics; and the original organ. Source: Bo Xia, @BoXia7 Multi-Omics: The study of DNA is genomics. DNA alone gives us an incomplete picture of an organism. Epigenomics can provide information about which regions of DNA are active or silenced. To understand how different cells function, as well as cells in different states of disease or health, you also need to know about their RNA and proteins. This data is contained in the field of transcriptomics (transcripts are strands of RNA transcribed from DNA) and proteomics (the proteins in a cell). Metabolomics looks at small molecules (such as sugars, amino acids, and vitamins) within the body and exposomics includes all sorts of environmental exposures. Collectively, these fields are known as -omics or multi-omics. It is valuable to combine multiple types of -omics together for richer sources of information, since each has different strengths, limitations, and insights to offer. Multi-Omics is an exciting area that draws on lots of data, with applications to cancer, infectious disease and immunology. Spatial: Spatial information lets us see all the variation within a section of the body– such as a segment of the intestines, the liver, or a cancerous tumor. This variation can often be significant for understanding disease and treatment prognosis. To better understand why spatial techniques are useful, let us dive into some background about cancer. Tumors aren’t just lumps of bad cells Cancer is defined as excess cell division. I used to think that tumors were just clumps of “bad cells”, where “good” and “bad” were binary states. This is incorrect. Tumors are not uniform, and within the category of “bad” there is a great deal of variation and heterogeneity. Different cells within a tumor may have different mutations from one another. And our immune systems sometimes build complex defense structures within tumors in attempts to more effectively fight them. How close a cancerous cell is to one of these immune structures impacts how likely the body is to destroy the cancerous cell. Notice all the variation within this tumor! Immunologist Dr. Angela Ferguson, who studies head and neck cancers, describes tumors as having a “physical landscape”. She has shown how the organization and structure within a tumor can predict and guide treatment outcomes. Her work found that cancer progression is “landscape-dependent”, where landscape refers to the locations of immune cells and structures within a tumor. It is not enough to study cancer cells in isolation. We need to understand their layout. Methods that effectively put tumor cells into a blender in order to sequence them, disrupting their spatial information, are insufficient on their own. Mapping that spatial information can hold the keys for more effective treatment. Single cell sequencing approaches allow a greater number of genes to be measured, whereas spatial approaches measure fewer genes but also provide location information. Combining these two approaches can prove powerful. Programming Libraries Applied to Spatial -Omics We are living at a time when multi-omics, spatial information, and user-friendly programming libraries are converging for easier exploration and discovery. For instance, below is an image I created using the common Python programming libraries pandas and matplotlib of data the NIH has shared about a rare liver disease. The image shows a slice of liver tissue, with gene expression overlaid in a color scale ranging from purple (low expression) to yellow (highest expression). Using Matplotlib to display a cross-section of liver with gene expression The NIH dataset contain images of slices of liver and expression of many different genes, from both healthy patients and those with a rare liver disease. https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE240429. The plot shows expression of albumin, a protein which helps transport other molecules around the body, overlaid on top of the liver, illustrating which regions produce more or less of this protein. Data science tools are invaluable for transforming and plotting such data. In addition to being able to use standard Python libraries (such as Pandas and Matplotlib) for visualing this data, there are many specialist libraries as well. The Scverse (Sc = single cell) includes a number of Python libraries focused on single cell analysis. The SC Verse includes libaries for single cell sequencing analysis There are also popular R libraries for single cell analysis and multi-omics, such as Harmony, Seurat, and mixOmics. With all of these tools and technologies, it is an exciting time to be working at the intersection of data science and microbiology. The causes of most diseases are complex and multi-factorial, although new approaches of spatial multi-omics provide unprecedented types of useful information. Related Reading: What AI can tell us about microscope slides Gaps and Risks of AI in the Life Sciences AI and Immunology You can subscribe to be notified of new blog posts by submitting your email below: I look forward to reading your responses. Create a free GitHub account to comment below.