EP 8: Implementation of AI in scientific research

Summary

This Data Science With Sam episode has Sam interview Lucas Simon about applying AI, machine learning, and data science to biomedical research at Baylor College of Medicine’s Therapeutic Innovation Center. The source turns the wiki’s existing AI For Science and Bioinformatics Domain Gap themes into a concrete molecular-data workflow: sequencing pipelines create gene expression matrices, Computational Biology analyzes them, and Single-Cell RNA Sequencing creates data shapes where Biomedical Deep Learning and autoencoder representations become more useful. Its core synthesis is that biomedical AI depends on the representation layer as much as on models: raw reads, count matrices, feature engineering, high-performance computing, team IT support, and biological interpretation all shape whether deep learning reveals real cell structure.

Key Claims

• Lucas Simon leads a small data-science group at the Therapeutic Innovation Center at Baylor College of Medicine, where the group applies data science to molecular data for early cancer therapeutics.
• The source says biomedical research has changed because sequencing and related technologies make large molecular datasets routine rather than exceptional.
• Gene expression matrices are presented as a central data object: roughly 20,000 expressed genes can be measured across samples or cells, producing matrices that are hard to interpret by inspection.
• Lucas makes a source-scoped distinction between Bioinformatics and Computational Biology: bioinformatics turns raw sequencing reads into analysis-ready matrices, while computational biology asks downstream biological questions from those matrices.
• Sequencing pipelines create storage and compute pressure because raw sequencing samples can contain tens of millions of reads and very large intermediate files.
• Molecular Feature Engineering is framed as a discovery opportunity, because researchers can summarize raw molecular signals in nontraditional ways rather than accepting only standard count-matrix representations.
• Single-Cell RNA Sequencing changes the deep-learning setting by moving from bulk RNA-seq experiments with perhaps hundreds of samples to cell-level datasets that can contain tens of thousands to around a million cells.
• Biomedical Deep Learning becomes more plausible in this data shape because there can be many more cell-level observations than genes.
• Autoencoders are used as an example of biological meaning emerging from a neural network: compressed hidden-space clusters can correspond to different cell types.
• The source emphasizes that model optimization alone is not the endpoint; visualization, output analysis, and biological interpretation are needed to turn learned structure into scientific insight.
• Tool choice is presented pragmatically: Lucas’s team uses a mix of R and Python, and commonly uses Keras, which sits on TensorFlow and has both Python and R interfaces.

Key Quotes

“bioinformatics” - Lucas’s term for the raw-read-to-matrix preparation layer.

“computational biology” - Lucas’s term for downstream analysis of analysis-ready molecular matrices.

“profiling means measuring” - Lucas’s clarification that gene-expression profiling is measurement, not prediction.

Connections

• Data Science With Sam, Sam (Data Science With Sam), Lucas Simon, Baylor College of Medicine, and Therapeutic Innovation Center - show, host, guest, institution, and research-center context.
• AI For Science, Human-Driven Scientific AI, Bioinformatics Domain Gap, Domain Expert Alignment, and AI Verification - broader scientific-AI and expert-interpretation frame.
• Bioinformatics, Computational Biology, Sequencing Data Pipeline, Gene Expression Matrix, and Molecular Feature Engineering - source’s molecular-data workflow branch.
• Single-Cell RNA Sequencing, Biomedical Deep Learning, and Single-Cell Autoencoder Representation - deep-learning branch enabled by cell-level sequencing data.
• Keras, TensorFlow, and Machine Learning Engineering - practical modeling and software-tool context.
• AI Drug Discovery Platform, AI Clinical Validation In Drug Discovery, Generative Biology, and AI Protein Design - adjacent wiki branch for AI in molecular biology and drug discovery.

Contradictions

• No direct contradiction found.
• The source extends Data, AI, and Scientific Research: A Coffee Chat by making the biology/bioinformatics boundary more operational: raw-read processing, matrix construction, feature representation, hardware, and cell-level data volume decide which AI methods are credible before biological interpretation begins.