Capitalism — the propaganda and constant narrative that flood our mindset — are far crueler than we realize. Sometimes, toxicity subtly lives and breeds around us, as simple as a man taking pride in your achievements — ones he had zero involvement in — yet claiming he helped you, like many others, which makes him a “good person” who deserves to be praised and celebrated at all times. Such cohort narcissism breeds so easily in our capitalistic culture that even awareness doesn’t protect us from the suffering of this world.That is why it is important to live our little lives consciously and with presence of mind.
To make the most of it — if we can’t create grand change in society, we can at least stop contributing to its ills.
There are many ways to be impactful for future generations, and one of them is to be a researcher — a curious explorer to the core.
My determination and love became one when I began working for a bioscience company. I started to see the value of my work in science, and how deeply and rapidly it could impact lives for generations.
This same principle of conscious, intentional action guides how I approach science — every choice in design matters, just as in life.
Just as conscious living means making deliberate choices that shape our lives, conscious design in science means making deliberate choices that shape discovery. In synthetic biology, every experiment is a choice — an investment of time, resources, and opportunity. The question is no longer just what can we build? but what should we build first, and why? That’s where AI becomes more than a tool — it becomes a compass for navigating the vast design space of biological possibility.
From Philosophy to Practice: Why Synthetic Biology Needs AI
Synthetic biology has always been a game of possibilities. Every gene, promoter, and pathway arrangement opens a new door — but each door costs time, resources, and patience to test.
Recently, the use of AI in synthetic biology has evolved in two distinct phases:
- The early stage — where machine learning and biodesign tools helped with relatively narrow tasks, like predicting protein structure from amino acid sequences.
- The deep learning stage — where transformers and Large Language Models (LLMs) are being used for more ambitious goals, such as predicting physical outcomes directly from nucleic acid sequences.
The promise?
In the future, AI could consider the full web of contextual factors — the polyfactorial environment of a biological system — and perform biomolecular modeling end-to-end without exhaustive lab testing.
The Challenge
Let’s frame the core problem:
How can we efficiently identify optimal combinations of genetic design parameters in a synthetic metabolic pathway — maximizing product yield while minimizing metabolic burden — without exhaustively testing all possible designs in the lab?
The combinatorial explosion is huge. If you have just three tunable genes, each with multiple promoter strengths, ribosome binding site strengths, and gene order permutations, the design space becomes massive. Testing every single possibility is impractical.
This is where AI-driven experimental design steps in.
SAFE: A 100% In-Silico Design–Build–Test–Learn Loop
In this case study, we’ll walk through SAFE — a safe, purely computational example of how active learning can accelerate the synthetic biology design cycle.
Here’s the twist:
No wet lab.
No real sequences.
No actionable protocols.
100% synthetic, for demonstration only.
The focus is purely on the machine learning strategy — how an algorithm can navigate a large design space, predict promising candidates, and improve over time using a Design–Build–Test–Learn (DBTL) loop.
How It Works
- Design Space — Define abstract genetic designs with discrete choices (e.g., “weak” vs. “strong” promoters, different gene orders).
- Synthetic Oracle — Simulate the output (product titer) and cost (metabolic burden) using a fictional, nonlinear function.
- Surrogate Model — Train a Gaussian Process regression model to predict outcomes for unseen designs.
- Acquisition Function — Use Expected Improvement (EI) to choose the next batch of designs worth “testing.”
- Iterate — Repeat the DBTL loop, improving results over each iteration.
Full Code: Github
1. Define the Design Space
Every possible “design” in our model is abstracted down to:
- Promoter strength (weak, medium, strong)
- RBS strength (weak, medium, strong)
- Copy numbers for genes A, B, and C (1 to 3)
- Gene order (one of six permutations)
@dataclass(frozen=True)
class Design:
promoter_strength: str
rbs_strength: str
gene_copy_A: int
gene_copy_B: int
gene_copy_C: int
gene_order: str
def design_id(self) -> str:
return f"P:{self.promoter_strength}|R:{self.rbs_strength}|A{self.gene_copy_A}B{self.gene_copy_B}C{self.gene_copy_C}|O:{self.gene_order}"
2. Simulate Measurements
The synthetic oracle acts as our safe, imaginary lab — given a design, it outputs a fictional titer (product yield) and burden (metabolic cost).
def synthetic_oracle(design, rng):
p = _cat_to_num(design.promoter_strength)
r = _cat_to_num(design.rbs_strength)
base = 2.5 * math.log(1 + 4*p*r) # diminishing returns
imbalance = ...
titer_obs = rng.normal(base, 0.10)
burden_obs = ...
return titer_obs, burden_obs
3. Train the Surrogate Model
We use a Gaussian Process Regressor with a Matern kernel to predict outcomes for unseen designs.
def build_model():
kernel = ConstantKernel(1.0) * Matern(length_scale=1.0, nu=2.5) + WhiteKernel(noise_level=1e-3)
gpr = GaussianProcessRegressor(kernel=kernel, normalize_y=True)
pre = ColumnTransformer([
("cats", OneHotEncoder(handle_unknown="ignore"), CAT_COLS),
("nums", "passthrough", NUM_COLS),
])
return Pipeline([("pre", pre), ("gpr", gpr)])
4. Pick the Next Designs to Test
We use Expected Improvement (EI) to decide which designs are most promising.
def expected_improvement(mu, sigma, best):
z = (mu - best) / sigma
return (mu - best) * norm.cdf(z) + sigma * norm.pdf(z)
5. Iterate the DBTL Loop
- Start with a few random designs
- Train the model
- Pick top candidates using EI
- “Test” them via the oracle
- Repeat
for it in range(iters):
proposals = propose_next(observed, candidates, lambda_burden, model)
new_obs = simulate_measurements(proposals)
observed = pd.concat([observed, new_obs])
6. Watch the Optimization Curve
SAFE learning curve: best-so-far (titer − λ·burden) improves steadily with each active-learning iteration (λ = 0.2, batch = 5).
Closing Thoughts
SAFE may be a simulation, but its lesson is real: in both life and science, intentionality is the multiplier. AI in synthetic biology doesn’t just accelerate experiments — it helps us decide which experiments are worth running at all. And just as living consciously can spare us from wasting years on the wrong paths, designing consciously can spare us from wasting resources on the wrong hypotheses. In the lab, as in life, progress comes not from doing everything, but from doing the right things — with purpose.