Baidu Miaoda 3.5's No-Code Ambition: Can It Truly Move the Bioinformatics Iceberg?

The release of Baidu Miaoda 3.5 has shown me, a bioinformatics postdoc who deals with pipelines and wet-lab data every day, an interesting tension. No-code platforms have surged to the top in market share, but will scientific research scenarios that truly require "code thinking" always remain their blind spot?

Let's start with the conclusion: The "simpler, more versatile" upgrade of Miaoda 3.5 is undoubtedly progress at the commercial application level. However, its technical architecture determines that it excels at handling business logic with clear rules, while struggling to directly adapt to scientific workflows like bioinformatics, which rely heavily on domain knowledge, messy data formats, and heterogeneous computing resources. But this precisely gives us an opportunity for reflection—the gap between dry-lab (computational) and wet-lab (experimental) work might be bridged by some capabilities of such platforms, provided the platform evolves from an "application generator" into a "domain workflow engine."

From a technical perspective, the essence of no-code platforms is abstracting generic components, allowing users to build logic through drag-and-drop and configuration. Baidu Miaoda 3.5 claims to have connected the "full chain from creative generation to product delivery," which sounds like encapsulating AI model invocation, front-end/back-end deployment, and data storage into black boxes. For standardized scenarios like e-commerce, customer service, and marketing, this indeed significantly lowers the development barrier. But in bioinformatics, a typical protein structure prediction task involves multiple steps from sequence retrieval, multiple sequence alignment, homology modeling to molecular dynamics simulation. The data formats at each step (e.g., FASTA, PDB, DCD) and computing resource requirements (GPU, CPU clusters) are highly heterogeneous. The existing component libraries of no-code platforms likely cannot cover these specialized tools, let alone handle noisy data generated during wet-lab experiments.

More critically, the definition of "correctness" in research environments is often not fixed. For example, AlphaFold2 prediction results need to be combined with experimental validation to judge confidence, involving extensive manual judgment and iteration. If no-code platforms force users to fix processes into templates, it would instead stifle the flexibility of scientific exploration. I've seen too many graduate students try to run RNA-seq analysis using Excel or GUI tools, only to give up because they couldn't handle batch effects, switching back to Python scripts. This lesson shows that for scientific computing requiring fine-grained control, a purely no-code approach may be counterproductive.

That said, the "more versatile" direction of Miaoda 3.5 is worth noting. If it can really provide scalable API interfaces, allowing user-defined Python modules or R packages to be inserted as components, then it has the potential to become a bridge connecting dry-lab and wet-lab work. For instance, wet-lab teams could use it to quickly build an interface for data entry and preliminary quality control, while dry-lab teams call custom statistical models via the backend. This "semi-no-code" mode might be more pragmatic than pure no-code.

Frost & Sullivan ranked Miaoda first among China's AI-native no-code platforms. This data mainly reflects acceptance in the enterprise application market, especially among SMEs lacking technical teams. For research institutions, budgets and talent structures differ; they tend to prefer purchasing professional software directly or developing in-house rather than relying on general-purpose platforms. But trends are changing: NIH and several EU projects have started trying low-code platforms to manage biological sample data and experimental workflows, such as using Node-RED or Knime. Although these platforms aren't as "AI-native" as Baidu Miaoda, they excel in native support for scientific data formats. If Baidu Miaoda can add parsing capabilities for common bioinformatics data formats (like HDF5, BAM) and provide preset statistical analysis templates in the future, its application space in the scientific community will expand significantly.

As someone who has long dealt with AI models, my biggest concern is the weakening of model interpretability behind "no-code." In protein structure prediction, we often need to adjust model hyperparameters or modify attention mechanisms to adapt to specific families. These operations are almost impossible to implement in a no-code interface. Miaoda emphasizes "

Original link: https://www.leiphone.com/category/industrynews/aRNiQFaWXJssTnLJ.html