Eric J Ma's Website

In this blog post, I explore the ESM3 model, focusing on its handling of missing modalities in multi-modality training. I dissect the model's architecture, input and output configurations, and the strategic use of default values for absent data. By examining the source code and conducting a toy example, I illustrate how embeddings are calculated and how they shift in vector space when modalities are missing. This deep dive reveals the model's elegant design and its potential for multi-modality integration. Has this piqued your curiosity yet?

Post SciPy 2024, I had a chance to try out pixi, a new environment manager from the prefix.dev team. I went cold turkey on my laptop, removing ~/anaconda, and haven't looked back. In this (very long) blog post, I detail my experience switching from mamba to pixi, the ways that pixi makes it easier to manage environments, how pixi helps with onboarding onto a project, supports containerization, GPU access, and seamless integration with Docker, and how it facilitates publishing to PyPI and running tests. The switch has streamlined my workflow significantly. Was this enough to get you curious about how pixi can optimize your development process too?

In part 3 of the series on protein language models, I explore the critical phase of evaluating protein sequences generated by language models, emphasizing the importance of practical, bioinformatics-based evals to narrow down candidates for lab testing. I explore both sequence-based and structure-based evals, highlighting their roles in filtering and ranking sequences to prioritize for experimental validation. Additionally, I offer insights on fostering collaboration between computational and laboratory teams to enhance protein design efforts. How can these evals and collaborations accelerate protein engineering?

In part 2 of my three-part series on PLMs in protein engineering, I do a deep dive into the training methods of protein language models, specifically focusing on masked language modeling and autoregressive training. I explain how these models are trained, highlighting the complexities and considerations involved in training, such as model choice, masking fraction, and the need for curated training sets. With these insights, I aim to shed light on the intricate process of preparing protein language models for protein design. Curious about how these models could revolutionize protein engineering?

In part 1 of a three-part series on protein language models, I do a deep dive into the fascinating world of protein language models (PLMs) for protein engineering, drawing parallels between PLMs and GenAI models like GPT-4. I explore three distinct applications: patent-breaking, prioritized deep mutational scans, and candidate expansion via natural mimicry, highlighting the goal of generating plausible protein sequences for diversification. I also touch upon the potential of PLMs for optimization, using mutational effect scores for predictive purposes. How can PLMs revolutionize protein engineering and what lies ahead in this promising field?

In this blog post, I share my enriching experience at SciPy 2024, from attending insightful tutorials on Quarto, LLMs, Anywidget, and handling large datasets, to delivering talks on fostering an open-source culture and LlamaBot. I also highlight the vibrant lightning talks, the collaborative sprints, and the engaging social activities that made this conference memorable. Not to forget, the delicious Tacoma cuisine that added flavor to the whole experience. Curious to know which tutorial inspired me to recreate my talks just for fun?

In this blog post, I share a cautionary tale from my work experience about the pitfalls of using pickle files for data storage, particularly highlighting their dependency on the computing environment and specific package versions. I encountered an issue when a notebook failed to run due to a pickle file not being compatible with the updated version of pandas. This experience led me to advocate for using native formats over pickles for better stability and reproducibility, and underscored the importance of software skills like continuous testing. How can we ensure our data storage methods don't become obsolete with evolving dependencies?

In this blog post, I reflect on two years of running quarterly docathons at work, dedicated two-day events focused on writing high-quality documentation. I discuss what docathons are, their purpose, and the simple yet effective way to organize them, emphasizing the importance of food, documentation, and optional workshops. Based on both cost and the invaluable benefits of improved documentation practices, I also discuss the significant return on investment these docathons have yielded. How can such a straightforward event substantially enhance the quality of documentation and team engagement? Read on to find out.

In this blog post, I share insights from a co-op performance calibration, highlighting the crucial difference between conversational and communication skills in the hiring process. I recount an experience where a candidate's excellent conversational abilities masked their technical skills, leading to a dilemma on whether to hire them. Drawing from Andy Grove's 'High Output Management,' I emphasize the importance of using interviews to gauge technical abilities effectively, advocating for code reviews as a high-bandwidth method to assess candidates' skills. This approach minimizes the risk of being misled by mere conversational charm. How can we better distinguish between conversational prowess and genuine communication skills in interviews?

In this blog post, I explore the importance of reproducibility and portability in data science, focusing on data access patterns. I introduce pins, an open-source tool that enables data scientists to reference data from a central source of truth and manage data versions explicitly. By using pins, we can avoid common pitfalls like non-reproducible analyses and streamline the process of accessing and versioning data. This approach not only enhances productivity but also ensures that data is accessed in a consistent and error-free manner. Curious about how pins and analogous tools can robustify your data science workflow?