Module 1 of 8

Biology is a data network, and you are not an imposter

1.1 The imposter has it backwards

Most CS students assume biomedicine belongs to people who spent years memorizing pathways and pipetting in wet labs. Look at what modern biomedical research actually produces: a single sequencing run emits tens of gigabytes; one mass-spectrometry experiment measures thousands of proteins across dozens of samples; a hospital generates millions of time-stamped clinical events. The bottleneck in biomedicine is no longer generating data. It is making sense of data — at scales no human reads unaided.

That is your home turf. Dimensionality, graphs, noise, caching, distributed state, schema design: the mental tools you already own are the ones the field is short of. What you are missing is smaller than you think: the domain's vocabulary, its data realities, and a sense of what a finding must survive to matter clinically. That is what these eight modules install.

1.2 The central dogma, read as an architecture diagram

The one piece of biology everyone must know is the central dogma: information flows from DNA to RNA to protein. DNA is the stable, heritable store, written in a four-letter alphabet — A, C, G, T — where A pairs with T and C pairs with G across the double helix. Transcription copies a stretch of it into RNA, a short-lived working copy that uses U in place of T. Translation reads RNA to assemble a protein, the molecule that actually does things: catalyzes reactions, carries signals, builds structure.

Here is the same sentence in your vocabulary, with the places where the metaphor breaks — the breaks matter more than the metaphor.

This is why the course calls biology a constraint network rather than source code. The genome does not execute line by line. Every measurement you will touch — gene expression, protein abundance, a clinical lab value — is one noisy projection of a system whose state depends on context you usually cannot observe. Good biomedical AI respects that; bad biomedical AI pretends the projection is the system.

1.3 The bioinformatics blindspot

A true story shape you will meet repeatedly. A team trains a classifier to detect cancer from tissue images collected at two hospitals. AUROC: 0.99. Publication-ready — until someone notices that one hospital contributed most of the cancer cases and the other most of the controls, and the scanners differ. The model learned to recognize the scanner. On data from a third hospital it performs at chance.

In omics the same failure wears different masks: batch effects (samples processed on different days cluster by day, not by disease), leakage through patient overlap between train and test splits, and models that exploit missingness patterns rather than biology. The metric was never wrong. The question it answered was.

1.4 AI as your bilingual teammate

You will use AI assistants throughout this course in a specific role: an on-demand interdisciplinary teammate that translates between medical language and computational structure. The skill is asking for structure, with sources, rather than asking for opinions. Compare:

Weak:   "Explain HER2-positive breast cancer."

Strong: "I'm a CS student building a dataset schema. From this paper's
        abstract (pasted below), extract: (1) the disease and subtype
        studied, (2) sample types, (3) measured molecules, (4) sample
        counts per group — as JSON. Quote the exact phrase supporting
        each field. Mark any field the text does not state as null.
        Do not guess."

Three properties make the strong prompt professional: it demands a machine-usable structure, it demands evidence (the exact supporting phrase), and it makes "not stated" a legal answer. Hold on to that last one. In consumer apps a wrong guess costs a shrug; in biomedicine a confident unsupported claim is the cardinal defect. Every system you build here treats "unknown" as a first-class output.

1.5 Two working examples, one standard

Memory Lane Games builds simple, localized reminiscence games for people living with dementia — photos and prompts tuned to a player's own era and place. No diagnosis, no medical claims, no patient data in a cloud. Yet care homes and families report real engagement and calmer days, and the model scales globally because localization is content, not code. It shows a path you will study in Module 6: a modest, well-guarded interactive tool can deliver verified human impact and a business.

Evidence navigators like omics research agents (Medea, and the pxEnrich system built within this program) answer plain-language questions about disease research from public datasets. Their defining rule: the language model is never the database. Retrieval is deterministic, every claim carries its source, and when evidence is insufficient the system says so. The same rule governs your capstone, whichever form it takes.

1.6 The three guardrails

Because this program operates under NIH policy, every project in this course enforces three rules from day one. They are graded, and they are also what makes your work investable and publishable:

1. Zero-PII boundary. No personally identifiable or patient data ever leaves the local environment. Public, de-identified datasets only; identifiers stripped at ingestion.
2. Deterministic output over hallucination. User-facing answers come from retrieval over vetted content or static lookups. A generative model may rephrase retrieved content; it may never originate medical claims.
3. Edge-case validation. Automated stress tests simulate messy reality — erratic users, corrupted inputs, out-of-distribution data — before anyone else touches the tool.

Results

Visible to course editors only. Students see their own score and a verification code.