What exactly has AI done in AI biology?

[doggie]

Recommended Words

What are the major directions of AI biology?
Talents needed for the AI+biology intersection must be strong on both ends, not weak on both ends.

If you are just hopping on the trend, it will eventually become: unable to understand formulas, only able to copy code, rely on scripts for training, only read figures in papers, get stuck when asked, and unable to perform experiments.

Original Text

This article is transcribed by SimpRead 简悦, original at mp.weixin.qq.com

I always find AI biology odd. Now it’s 2025, will there really be jobs after three years of graduate study?

As long as you study relatively mainstream and technically intensive directions in “AI × biology” (such as computational biology, bioinformatics, protein structure prediction, single-cell/multi-omics analysis, AI drug discovery, etc.), there will be career opportunities three years later. But that doesn’t mean “anyone can get by easily.” This field is both popular and competitive now; the threshold and differentiation are increasing.

I’ll clarify a few key questions:

1. What types does your “AI biology” roughly include?

The so-called “AI + biology” today generally falls into several paths (not necessarily mutually exclusive):

① Computational biology / Bioinformatics (research-focused)

This involves omics data (gene expression, single-cell, multi-omics), evolution, regulatory networks, etc.

Common tools: Python/R, statistics, machine learning, deep learning, graph neural networks, plus a bunch of pipelines and Linux.

Career path: postdoctoral positions at universities/research institutes, research roles in medical schools/hospitals, “algorithm/data scientists” or “bioinformatics engineers” in biopharmaceutical or device companies.

② AI drug discovery / Structural biology direction

For example, protein structure prediction (AlphaFold lineage), molecule generation, virtual screening, molecular dynamics + deep learning.

Career path: pharmaceutical companies (especially Biotech, AI drug discovery startups), large pharma R&D departments, some tech + bio companies.

③ Medical AI / Digital health

Such as pathology images, medical imaging (CT, MRI), electronic medical records NLP, clinical decision support systems.

Career path: internet healthcare companies, imaging/device companies, hospital information centers and related cooperative project teams.

④ Purely “playing with machine learning” pseudo AI-bio

The studied field bears “intelligence” or “AI” title, but in reality, it’s just helping labs run tools, tune parameters, do some visualization.

The algorithm, engineering, math skills are weak, and biological understanding is average.

This category is the most dangerous: by graduation, you are neither hardcore CS/ML nor have substantial experimental skills in traditional biology, easily becoming “not strong enough on either side.”

You roughly need to judge which category you will be more like in the future. What really determines your job prospects is “skills and project depth,” not the professional title.

2. From 2025 onward, over the next 3–5 years, will this direction rise or fall?

Demand is rising, but thresholds are also rising, with huge differences among regions/schools/individuals.

Let me elaborate on the reality:

Industry level

Population aging, high drug development costs, trends in precision medicine all really exist, and they require ultra-large-scale data and modeling.

After large models emerge, bioinformatics and drug discovery actually need people who can “align the model with biological problems,” not just those who can use APIs.

Large pharma, Biotech, and internet giants are continuously recruiting “people who understand AI and biology/medicine,” but they prefer:

• actual project outputs (papers, open-source code, deployed products, pipelines);
• individuals who can independently close the loop from “biological hypothesis → data → model → result interpretation.”

Academic level

Bachelor’s: Generally difficult to directly do high-quality “AI + biology” roles unless you have very strong CS + research background.

Master’s/graduate: For research-type masters with solid projects, doing engineering/algorithm/data science is feasible; for “diploma-mixing” masters, competitiveness is weak.

PhD: Still the mainstream threshold for those wanting to pursue genuine research-driven jobs (research institutes, pharma research scientists).

Your question is “is there work after three years of graduate study,” but the key is what level you can accumulate in those three years.

3. More realistically: what background makes it relatively easier to find a job?

To be frank: if you meet most of the following conditions, you don’t need to be too pessimistic:

Skill structure

Programming: at least proficient in Python (data processing, modeling, writing pipelines), can read/modify open-source projects; if you also know some C++/Rust/Java, that’s a plus.

Math/statistics: basic probability, mathematical statistics, linear algebra, can understand classic machine learning/deep learning formulas, not just “library tuning.”

Machine learning/deep learning: have truly implemented or modified models from scratch (e.g., variants of autoencoders, GNN, Transformers in biological tasks), understand data preprocessing, feature engineering, model evaluation.

Biological knowledge: no need to be a wet-lab expert, but must understand the biological meaning of your data and communicate normally with experimental/clinical people.

Projects/results

Completed full-chain projects: from problem definition → data collection/cleaning → modeling → result interpretation → report or paper writing, preferably with open-source code.

One or two respectable papers (not necessarily top journals, but reflecting methodology and thinking), or genuine deployment results in a company/lab.

Able to explain a project in resume and interview “with clear algorithm details and clear biological significance.”

Such people, in the three years after 2025, their job search is mainly a matter of “which place or company type to choose,” not “whether there is a job.”

4. What pitfalls could make you very awkward after graduation?

I might be a bit blunt here:

Only chasing trends without solid foundations

If after three years your status is:

• Cannot understand machine learning, just copy code;
• Model training relies entirely on senior colleagues’ scripts, and you don’t know how to debug;
• Papers are only read for conclusions and figures, ignoring methods;
• Stuck on biological questions at depth, can only say “because the paper did so.”

That will be difficult. Because for companies, you are neither a capable project engineer nor an original research scientist, but just a “technician who can run others’ scripts,” highly replaceable.

Only doing traditional wet-lab but wants to hang an “AI” label

If your main skill is still experiments (PCR, Western blot, cell, animal experiments), but:

• You write “participated in AI projects” on your resume;
• Actually just hand over data to others to train models;
• Not familiar with code or algorithms;

Then you basically won’t pass technical interviews for “AI-related positions.” And when applying for traditional biology jobs, you’ll be asked: “Did you waste three years of experimental skills?”

This “in-between state” is what many fear most.

Overly applying “overall environment anxiety” to yourself

Indeed, macroeconomic and industry ups and downs exist; layoff news, fierce competition for PhD positions are real.

But what ultimately decides your employment is your own projects and skills, not abstract words like “AI bubble” or “bad macro environment.”

5. If you have to make a decision now, how to evaluate if this path suits you?

Here are some simple self-check questions; you can grade yourself mentally:

1. For “coding, reading math formulas, debugging” — work偏理工, do you resist it or like it?

2. Compared to “doing experiments,” which do you enjoy more:

A: Repetitive operation on the bench;

B: Sitting in front of a computer analyzing data, tuning models, looking at results?

3. Do you have the patience to “check documentation/search/source code all day to figure out a technical detail”?

4. Can you accept spending much time in three years off the bench, revolving around Linux + Python + piles of papers?

If you are not averse and even a bit excited by “coding, modeling, reading papers,” AI biology is likely an opportunity, not a pitfall.

If deep down you prefer wet-lab and are very averse to coding, seriously consider if you want to pursue “hard AI” or be a “biologist who knows some computational skills.”

6. Realistic plans you can make now

Clarify a main axis: decide if you are a “CS-leaning AI-bio person” or a “bio-leaning AI-bio person,” then shore up your weaker side, avoid shallow attempts on both ends.

Aim to complete at least 1–2 projects you can speak about for 30 minutes: including problem background, data, methods, results, limitations, and improvement directions.

Make your capabilities “visible”: GitHub, public data competition results, papers/preprints, internship experiences — all real “chips” for job hunting.

Proactively learn job descriptions of your target roles, see their tech stack requirements, then backtrack to what you should improve in these three years, not building in isolation.

7. Back to your original question: “Always feel AI biology is odd, will there be jobs after three years?”

The “odd” part lies in the fact that this field is indeed a mixed bag; under the AI label, there are both real and superficial topics and jobs.

The real version of “are there jobs after three years” is: there are many positions needing “understanding AI, data, and able to communicate with biology/clinical,” but they won’t pay for “jack of all trades, master of none.”

If you treat these three years as real skill accumulation time, striving to be the kind of person who “can explain and solidly do a bio problem with AI,” AI biology is a springboard, not a pitfall.

If you just want to ride the “hot trend” to avoid employment pressure, hoping that “just the professional label automatically brings good jobs,” that’s not only AI biology — most professions will disappoint after 2025.

What I find odd is: labs have no AI foundation but just start AI biology direction. Let zero-foundation students explore, plus companies do so too. Also, everyone has GPT write ML and deep learning code and calls it AI. I’m not sure if this has real job depth. I believe AI biology prerequisite is doing some fundamental work, not just superficially writing ML code to analyze data.

What you said is actually the most real and embarrassing issue in many “AI+X”:

Labs and companies with no AI base hang the label and let zero-foundation students plus GPT “feel around.” Your instinct that it feels hollow is not wrong, but it requires closer inspection: what is “fake AI biology,” and what counts as a truly deep direction.

To put it simply: most AI biology you see is indeed superficial.

The phenomena you mentioned can be unpacked one by one:

First, “labs have no AI foundation, just wanting to catch the trend, start AI biology direction for zero-foundation students to try.”

This happens worldwide, not only in China: many PIs are in traditional biology, pharmacology, molecular biology; recently they feel they can’t publish or get funding without AI, so they:

• write a book/read reviews, invite collaborators with some coding skills;
• let students learn ML, run public datasets;
• add “AI-driven” or “deep learning” in project titles.

Such groups easily become what you described:

“Students zero foundation + mentor half-knowledge + GPT-generated code = seemingly AI paper topics.”

Secondly, “outside companies also do the same.”

Many small/mid pharma, bio companies, even hospital info centers say they want to do “AI imaging,” “AI drug screening,” but in reality:

• buy ready-made tools/platforms;
• hire 1–2 data analysis/ML engineers;
• let them input internal data to try models, hoping for some “AI reports.”

In such environments, work easily degrades to what you said:

“letting GPT help write some model code, tune parameters, produce some figures, and call it AI.”

Thirdly, “writing ML/DL code with GPT counts as AI.”

This is a new problem: before, if you couldn’t write code, you had to learn patiently; now GPT writes about 80% of seemingly plausible code, making “fake AI” very easy in the short term.

But here is a crucial reality:

Real employers, publishers, product teams look at whether “you can make things stable, explainable, and deployable.”

GPT-generated code alone can’t uphold that. GPT lowers the entry barrier but makes “who is deep, who is shallow” clearer:

• Shallow people: let GPT write code and run it, without really understanding loss functions;

• Deep people: use GPT to boost efficiency, but can judge if the model is overfitting, if data is valid, and if conclusions hold.

So your intuition that “everyone is just skimming the surface” exists in “fake AI environments,” true indeed;

But I don’t agree if you conclude “whole AI biology is superficial and has no deep work.”

So what counts as “deep” AI biology? What exactly is “fundamental work”?

You mentioned a key but abstract term: “fundamental work.”

Intuitively, AI biology shouldn’t just be fitting models to data; there must be more solid foundations. This feeling is correct, but to explain clearly, AI biology needs to be split:

Truly deep AI biology includes several layers of “fundamentals”:

One layer is data and experimental fundamentals.

Simply put: you must know if the data is reliable and can support the model to learn meaningful things. For example:

• Omics data: RNA-seq, single-cell, ATAC-seq, proteomics — their experimental design, batch effects, sources of noise. If you know nothing and just throw expression matrices into models, results will likely look flashy but be fake.

• Structural and molecular data: protein structures, ligand conformations, molecular dynamics simulations — their physical constraints and chemical rationality. Without concept understanding, you may produce “models predict well but molecules cannot be synthesized.”

This layer is what you meant:

“Do foundational work before talking about AI.”

Foundational work doesn’t mean “waste years doing low-skill chores,” but that you must understand biology, chemistry, experiments, and data sufficiently; otherwise, AI is just a castle in the air.

The second layer is algorithms and modeling themselves.

Here “depth” doesn’t mean you must create new SOTA models but you must at least:

• know why use a certain model instead of another (e.g., why use graph networks for molecules, why transformers for sequence data);

• know model assumptions and limitations (e.g., distribution, bias, sample size of training data);

• know how to evaluate if the model is truly learning, not just memorizing (cross-validation, external validation set, prospective validation, negative control design, etc.)

When a group/company only stays at “GPT wrote me a ResNet/Transformer and I paste result curves in PPT,” that’s not deep AI biology, just playing show.

The third layer is truly closing loops with biological questions.

Deeper AI biology must ask: “What practical impact do model results have on the real world?”

Examples:

• Drug discovery: molecules prioritized by the model — can experiments validate them? If they fail, why, and how to improve the model?

• Disease subtypes: new subtypes discovered via clustering and representation learning — can clinics distinguish them? Could this guide medication?

• Gene regulation/pathways: build regulatory networks — can key nodes show expected phenotypes via knockout/overexpression?

These extend the “foundational work” you mentioned:

Data as root, model as brain, experiment as hand — these three must interlock, not fly separately.

What about “hanging AI label, no foundation, just starting out,” any way out?

Frankly, if a lab/company is like this:

• mentors/bosses know nothing about AI, just think “let’s do some machine learning to look cutting-edge”;

• projects basically “download public datasets → fit classification/regression/clustering models → publish an article”;

• no long-term, biologically clear problem lines, no serious experimental validation;

In such an environment, even if you write code with GPT every day, your competitiveness after three years is risky. Your intuition questioning “is this deep work” is not oversensitivity but clarity.

More realistically, people trained here easily fall into:

• code only “runs” but engineering and math foundations are weak;

• biological understanding stays at “know the process a bit,” lacking deep knowledge of systems/diseases/technologies;

• projects mostly on public datasets, lacking a full closed-loop experience from problem to data generation to model design to validation.

Such resumes, to recruiters, easily become representatives of a “superficial generation of AI biologists.”

On this point, I agree and support your caution.

But I want to counter one point: totally denying “starting from zero exploration” is unreasonable

You say “labs with no foundation start AI biology and let zero-base students explore,” it sounds risky but doesn’t mean no value. The key difference is:

One is, mentors may lack AI experience but have clear biological problems and long-term accumulation, e.g.:

• worked on a disease for 10+ years, very aware of experimental data pain points;

• just never used machine learning before, now want to use AI to integrate multi-omics, improve screening efficiency.

In this case, even if the AI part is learned on the fly, as long as the question is solid, data real, and experiments validate, it’s genuinely building “a bridge from zero to one,” which is valuable.

The other is, the problems are unclear, just making an “AI topic” to publish papers, choose data arbitrarily, do not care about biological significance — that’s pure “empty.”

From a student perspective, you can judge your environment with questions:

• Does this group have a clear biological/medical problem direction or change topics yearly to follow trends?

• Is the boss good at data and experiments but unfamiliar with AI, or “jacks of all trades but masters of none”?

• Does the project seriously consider validating model conclusions, or just stay at “looks like metrics are okay”?

If the first two are yes and the third is being tried, “starting AI biology from zero” can still be an upward environment;

If all three are weak, your concern is basically a realistic risk.

What you really need to figure out: what kind of person do you want to become?

Your current confusion can be translated into a blunt question:

“I don’t want to be the type who only uses GPT to write models but has no real value. How can I go deeper in AI biology?”

Objectively, you have several options:

If you still want to stay in this field, you need to consciously resist the “superficial” side. For example:

• In biology, pick something you really want to understand deeply, not everything superficially;

• Technically, at least master a full chain: from data acquisition → preprocessing → feature/representation → modeling → evaluation → interpretability → experiment validation design;

• When using GPT, treat it as a tool, not a crutch: let it help write boilerplate code or find APIs, but critical design, metrics, and questioning results must be your responsibility.

If you find you have no interest in programming or modeling, honestly face it:

You may be more suited to go deep into biology/medicine/experiments, becoming an expert in a disease mechanism or technical platform, cooperating with AI teams instead of writing models yourself.

Such roles are highly needed in truly mature teams, not “lesser options.”

If you neither want to dig deep biology nor deep AI but just want a “seemingly cutting-edge” major, I’ll frankly say:

Whether AI biology or other popular interdisciplinary fields, you will most likely be weeded out in 3–5 years.

Because GPT writes superficial code for you, AI tools help do basic analysis — these shallow tasks are easiest to be automated.

Finally, to be frank about this

Your current feeling of oddness is because you faintly sense a fact:

“The AI biology label itself no longer guarantees value; what really matters is the set of abilities and perspectives you accumulate inside.”

On this, I am fully on your side.

But I won’t say “then don’t touch AI biology,” because if you still have interest and willingness to work hard, this path still has depth and jobs, just:

• don’t assume “the lab says it’s AI, so it naturally is deep”;

• don’t assume “just can write code with GPT means mastering AI.”

What you truly need in “fundamental work” is actually:

solid biological understanding + respect for data and experiments + clear-headed judgment of models + persistence in closed-loop validation.

These are not written on the professional title or automatically granted by research direction; you can only get them by choosing the right environment, the right mentor, the right projects, and investing your own effort bit by bit.