Week 127–131 · AI Meets Algae — Computational Biology

Part 1 of 5 · Why algae needed a computational layer

The Variable Problem

A microalgae cultivation run is a six-to-nine variable optimization problem playing out in real time, and most of those variables interact with each other in ways a spreadsheet can't capture.

Light intensity, temperature, inoculum ratio, pH, reactor geometry, CO2 concentration, and nutrient levels each shift the optimum depending on which species is growing and what compound it's being grown for — a condition set tuned for Spirulina protein yield is not the condition set tuned for Haematococcus astaxanthin induction, even though both are "just" cultivation parameters. Test even a modest grid of five of these variables at four levels each, and exhaustive wet-lab screening requires over a thousand individual cultivation runs. A single photobioreactor run can take five to ten days to reach steady state. Full factorial screening at that scale is not slow — it is commercially impossible.

This is the gap computational biology has been filling in algae research over roughly the last decade. Not replacing the wet lab — narrowing an impossibly large search space down to the handful of conditions, or genetic edits, that are actually worth the cost of testing. Recall the Design-Build-Test-Learn cycle from Week 77–80's synthetic biology survey: computation has mostly compressed the Design and Learn steps, while Build and Test still happen in a real reactor with real cells. Nothing in this module replaces a photobioreactor. Everything in it changes which experiment gets run in that photobioreactor first.

Part 2 of 5 · Genome-scale metabolic models

What a Cell's Spreadsheet Looks Like

A genome-scale metabolic model turns an organism's entire known biochemistry into a system of equations — and flux balance analysis asks those equations what the cell would do if you changed one input.

Every catalogued metabolic reaction in the organism is written down with its stoichiometry: which molecules go in, which come out, in what ratio. Flux balance analysis (FBA) then searches for the combination of reaction rates — the "flux distribution" — that maximizes an objective, usually growth rate, subject to mass balance and whatever nutrient uptake limits are measured or assumed. It is linear optimization applied to biochemistry, and it has been run on microalgae for longer than most people in the field realize.

The first such model for an alga came from Boyle and Morgan in 2009: a reconstruction of Chlamydomonas reinhardtii primary metabolism with 484 reactions and 458 metabolites. It was already enough to show, on a spreadsheet, that autotrophic growth yields roughly 28.9 grams of biomass per mole of carbon fixed, against roughly 15 grams per mole when the same species runs heterotrophically on acetate in the dark — close to a two-fold carbon efficiency penalty for going heterotrophic, predicted before anyone needed to run a fermenter trial to find out.

The model that the field actually uses today is the 2015 refinement, iCre1355, built by Saheed Imam and colleagues. It scaled the reconstruction to 1,355 genes and roughly 2,400 metabolites and reactions spanning the nucleus, chloroplast, and mitochondria — and when tested against real chemostat data, the model's growth-rate predictions matched measured rates with an R² of 0.83, using only three measured nutrient uptake rates (carbon, nitrogen, phosphate) as input. That is a model explaining 83% of the variation in real growth from three numbers. It also correctly predicted the point at which nitrogen starvation halts growth and triggers triacylglycerol (lipid) accumulation — the exact mechanism behind most algal biofuel strain strategies.

For a company evaluating a process change, this is the practical value: a model like iCre1355 lets you ask "what happens to lipid yield if nitrogen is cut by 30%" without burning three weeks of reactor time to find out. It will not give you a final answer — outdoor productivity always diverges from model predictions, a theme this module returns to — but it gives you a ranked list of which questions are worth asking the reactor at all.

Diagram · The model-to-decision loop

Part 3 of 5 · Machine learning for monitoring and control

From Manual Sampling to Continuous Prediction

Every raceway and photobioreactor generates more sensor data per hour than a single technician can meaningfully act on — which is exactly the gap machine learning has been built to fill.

The traditional way to know how much biomass is in a culture is to pull a sample, measure optical density or dry weight, and wait — a process that takes hours and always describes a culture that has already moved on. A "soft sensor" replaces that lag with a model trained on continuous streams of light, temperature, pH, dissolved oxygen, and turbidity data, predicting biomass concentration and productivity in something close to real time. Igou and colleagues built one of the more concrete demonstrations of this for open systems: a deep learning model trained directly on real-time sensor profiles from open raceway ponds, predicting productivity from the sensor stream itself rather than waiting on lab assays.

Across the published literature, the toolkit in active use spans support vector machines, genetic algorithms, decision trees, random forests, artificial neural networks, and deep learning — applied to a fairly specific set of problems: species classification and contamination detection from microscope images, cultivation-condition optimization, harvesting-timing prediction, and productivity forecasting. None of these is exotic machine learning by 2026 standards; what's notable is how recently algae cultivation became a serious application area for it. One review framed the stakes in market terms: global microalgae-based product revenue was projected to grow from roughly $32.6 billion in 2017 to $53.43 billion by 2026 — and that growth is one of the reasons automation has become commercially urgent rather than academically interesting.

Computer vision deserves its own mention because it changes a specific operational bottleneck directly relevant to Week 62–65's harvesting economics. A convolutional neural network trained on microscope images can classify algal species and flag contaminating organisms or grazers faster, and more consistently, than a technician scanning a hemocytometer by eye. Catching a rotifer bloom or a competing diatom strain a day earlier doesn't just save the current batch — it changes which harvesting train is worth running on the rest of it.

Layer 1

Model

Genome-scale metabolic models and flux balance analysis. Simulates whole-cell biochemistry to predict phenotype from genotype and environment before a flask is touched.

Layer 2

Monitor

ML soft sensors and computer vision. Converts continuous sensor and imaging streams into real-time biomass, productivity, and contamination calls.

Layer 3

Design

AlphaFold- and ProteinMPNN-style protein design tools. Predicts and generates protein structures and sequences to guide which genetic edit is worth attempting.

Part 4 of 5 · AI-designed proteins

A Structure Without a Crystal

For sixty years, knowing exactly how a protein folds meant growing a crystal of it and bouncing X-rays through it for weeks. AlphaFold changed the unit of measurement from weeks to minutes.

AlphaFold, developed by Google DeepMind, predicts a protein's three-dimensional structure from its amino acid sequence alone, trained on the accumulated structures in the Protein Data Bank. The AlphaFold Database, run jointly with EMBL-EBI, now hosts over 200 million predicted structures — open access, covering essentially the cataloged span of UniProt. In the field's blind benchmark competition, CASP13, the original AlphaFold produced high-accuracy structures for 24 of 43 of the hardest test domains, against 14 of 43 for the next-best method. That is not an incremental improvement on prior tools; it is the reason structural biology now treats "predict the fold" as a largely solved first step.

For enzyme engineering specifically, this matters because the predicted structure shows the shape of an active site, which lets researchers decide which amino acids to mutate to shift substrate specificity or boost catalytic activity — directly relevant to engineering lipid-pathway enzymes for higher EPA or DHA yield, or carotenoid-pathway enzymes for astaxanthin titer, both covered in Week 72–80's strain selection and synthetic biology survey. AlphaFold3, released in 2024, extended this further: rather than predicting an isolated protein, it can model how that protein's active site holds an actual ligand, cofactor, or nucleic acid — the difference between seeing an empty lock and seeing the lock with the right key already in it.

A complementary tool, ProteinMPNN, runs the problem in the opposite direction. Instead of predicting structure from a known sequence, it generates new sequences likely to fold into a chosen target shape — letting researchers design a novel enzyme outward from a desired active-site geometry, rather than mutating an existing enzyme by trial and error one residue at a time. Combined, AlphaFold and ProteinMPNN form a rough design loop: predict what a candidate enzyme would look like, generate sequence variants likely to hit a better-shaped active site, predict those structures, and narrow down to a shortlist worth synthesizing.

The catch that matters

A predicted structure is one frozen snapshot. Real enzymes flex between multiple shapes during catalysis, and that flexibility is frequently where the actual catalytic improvement lives. Researchers using AlphaFold structures for docking-based virtual screening have reported lower hit rates than docking against structures sampled across that conformational range — the model narrows the search space; it does not finish the job. Anyone proposing to skip wet-lab structural validation because "the model is accurate enough now" is skipping the part where the prediction gets tested against a moving target.

Part 5 of 5 · Limits, synthesis, and the commercial signal

Where the Model Stops and the Reactor Starts

Every tool in this module narrows a search space. None of them grows a single cell of algae, and the gap between a confident prediction and a validated outdoor result is where most of the field's overpromising lives.

Data scarcity outside a handful of species

Genome-scale models and large training datasets concentrate on the best-characterized organisms — Chlamydomonas reinhardtii, Nannochloropsis — because they have the multi-omics data needed to build a model at all. Commercially relevant strains like Haematococcus or Schizochytrium are comparatively undermodeled, which means the computational layer is least mature exactly where SustaBloom's product strains sit.

The lab-to-field gap

Models and ML systems trained on controlled bioreactor data routinely underperform outdoors, where light, temperature, and contamination vary in ways the training data never saw — the same uncontrolled-environment problem covered in Week 51–54's raceway pond economics, now showing up as a model limitation rather than a cultivation limitation.

Conformational blindness in structure prediction

A single AlphaFold structure captures one shape. Enzymes move through several during a catalytic cycle, and docking decisions made against only the static fold can mislead a redesign effort — a documented, not hypothetical, failure mode in the published literature.

Regulators don't accept model output as data

FSSAI, EFSA, and the FDA require wet-lab and field validation regardless of how confident a computational prediction is. A genome-scale model can guide which strain to pursue; it cannot substitute for the safety dossier that Week 96–99's regulatory survey describes.

Talent and compute cost are real, and scarce

Building or validating a genome-scale model, or training a computer-vision contamination classifier, requires bioinformatics skill that remains genuinely rare in algae-specific biotech outside a small number of large research institutes — this is not a tool you adopt by buying software.

Small-sample overfitting in cultivation ML papers

A recurring methodological weakness flagged across the review literature: many published cultivation-optimization models report strong fit statistics on training sets of a few dozen experimental runs from a single facility — exactly the kind of result that looks impressive on paper and collapses on a different reactor in a different climate.

On what computation actually buys you

A genome-scale model doesn't grow a single cell of algae. What it does is tell you which of a thousand possible experiments is worth the cost of running first.

Tool / method	What it predicts	Data needed	Maturity for algae	Named example
Genome-scale model + FBA	Growth rate, flux distribution, lipid trigger points	Genome annotation + measured nutrient uptake rates	Established	iCre1355 (C. reinhardtii)
ML soft sensors (ANN / RF / deep learning)	Biomass, productivity from continuous sensor data	Continuous PBR / raceway sensor logs	Growing	Igou et al. raceway productivity model
Computer vision (CNN)	Species ID, contamination, cell density	Labeled microscope image sets	Early-stage	Species-specific academic classifiers
AlphaFold2 / AlphaFold3	3D protein/enzyme structure, ligand binding	Amino acid sequence (+ ligand for AF3)	Mature, not algae-specific	AlphaFold DB — 200M+ structures
ProteinMPNN	New sequences for a target fold	Target backbone structure	Early industrial use	Inverse-folding enzyme design pipelines
Genetic algorithms	Optimal cultivation-parameter combinations	Historical / factorial experiment data	Established	Light / CO2 / nutrient optimization studies

SustaBloom Signal

The Design and Learn steps of any DBTL cycle SustaBloom runs on a strain or process change (Week 77–80) can now be compressed from weeks to days using a genome-scale model or a trained soft sensor. That doesn't replace the photobioreactor trial — it changes which trial gets run first, and that ordering effect is where the real savings sit.

Of everything in this module, monitoring automation — soft sensors, CV-based contamination detection — is the nearest-term, lowest-risk lever. It reduces a real labor and lag cost in today's operations rather than promising a future strain that doesn't exist yet. If SustaBloom adopts one capability from this list first, this is probably it.

Treat every AI-strain-design or AI-optimized-cultivation claim from a vendor or paper as a hypothesis generated on a small dataset at one facility, not as a validated result. The gap between a published R² and field performance at a Tamil Nadu raceway is the same gap that defines every scale-up failure covered in Week 86–90. Ask for out-of-sample, out-of-facility validation before any number from this space enters a financial model.

Test Your Understanding

Four scenarios. Work through each before revealing the answer.

A research partner proposes using AlphaFold3 to redesign a lipid-pathway enzyme in Nannochloropsis to boost EPA output, and wants to skip wet-lab structural validation since "the model is basically as good as crystallography now." Evaluate that claim, and state what you'd actually require before approving lab work based on the predicted structure.

Show answer ↓

The claim overstates what AlphaFold delivers. AlphaFold2 was a genuine step-change — in the CASP13 benchmark, it produced high-accuracy structures for 24 of 43 of the hardest test domains against 14 of 43 for the next-best method — but "competitive with experiment" on a benchmark of static folds is not the same as "sufficient for redesigning a catalytic site." Two specific gaps matter here. First, AlphaFold predicts one frozen conformation, while enzymes move through several shapes during a catalytic cycle; published work has shown that docking against a single static AlphaFold structure can produce lower hit rates than docking against structures sampled across that conformational range, which means a redesign aimed only at the static structure can target the wrong moment in the catalytic cycle. Second, AlphaFold3 extended prediction to ligand and cofactor complexes, which is a real improvement for seeing how the active site holds its substrate, but that prediction is still a hypothesis about binding geometry, not a measurement of it.

What to actually require: treat the AlphaFold3 structure as the starting hypothesis for which residues to mutate, not as the final design. Before committing lab resources, ask for an activity assay on at least the wild-type enzyme to confirm baseline kinetics match what the model assumes, and request that any proposed mutation set be tested across a small panel of variants rather than a single "optimal" design — since the model's confidence in one answer doesn't reflect its uncertainty about enzyme flexibility. If the partner has access to it, a partial experimental structure (even low-resolution cryo-EM) on the wild-type enzyme would be the cheapest way to catch a conformational mismatch before committing to a full CRISPR edit and field trial. The model earns its place by narrowing a large mutation space to a short list — it should not be the only filter between a hypothesis and a strain change.

A vendor pitches an "AI-optimized raceway" system, citing a deep learning model that predicted productivity from sensor data with R² = 0.91 in their published paper, based on 40 cultivation runs at a single facility. What would you want to know before trusting that number for SustaBloom's site in Tamil Nadu?

Show answer ↓

An R² of 0.91 on 40 runs from one facility is exactly the pattern flagged as a recurring weakness in the cultivation-ML literature: strong fit statistics on a small training set collected under one set of environmental conditions, one strain, one season, and one set of sensors. That number tells you the model fits its own training data well. It tells you almost nothing about whether it will generalize to a different climate, a different strain, or a different season — and Tamil Nadu's light, temperature, and contamination profile will differ from whatever facility generated the training data, sometimes substantially.

Before trusting the number, the questions worth asking the vendor are specific: was the model validated on a held-out test set from a different facility or season than the training data, not just a held-out subset of the same 40 runs? What was the R² on that out-of-sample test, if one exists? How does the model perform when an unmodeled disturbance occurs — a contamination event, an equipment failure, an unusually cloudy week — since real raceway operation includes exactly these disruptions and a model trained on clean runs may fail silently rather than flagging its own uncertainty? And practically: will the vendor commit to a pilot period at the Tamil Nadu site with performance benchmarked against simple manual sampling before any contract terms assume the model's predictions are reliable? A genuinely useful soft sensor model should be presented with its out-of-sample validation as the headline number, not the in-sample fit — if a vendor leads with the latter, that itself is informative.

The iCre1355 model predicts that switching Chlamydomonas from autotrophic to heterotrophic cultivation lowers carbon-to-biomass yield from roughly 28.9 to 15 grams per mole of carbon. The production team wants to use heterotrophic cultivation anyway, for faster and more controllable indoor growth. Is the model's yield number reason enough to reject the heterotrophic route? Use both the model output and the practical cultivation economics from Week 59–61.

Show answer ↓

No — and treating the model's yield number as a standalone verdict would be a misuse of what a genome-scale model is for. Carbon-to-biomass yield is one input into a techno-economic analysis, not the whole analysis. The model is correctly telling you that heterotrophic growth on acetate is roughly twice as carbon-inefficient as autotrophic growth — that's real and worth knowing — but it says nothing about volumetric productivity, light dependency, contamination risk, or capital cost, which is where heterotrophic cultivation's actual commercial case comes from. Week 59–61 covers exactly this trade-off: heterotrophic and mixotrophic systems can reach much higher cell densities per unit reactor volume because they aren't light-limited, growth can run in the dark on conventional fermenter infrastructure that's cheaper and better understood than a photobioreactor, and a closed fermenter is far easier to keep free of competing organisms than an open raceway exposed to ambient air and water.

So the correct synthesis is: the carbon-efficiency penalty the model identifies has to be weighed against the cost of the carbon source itself (acetate or glucose, both of which have a real price per ton), against the capital savings from using fermenter infrastructure instead of a photobioreactor, and against the value of more predictable, contamination-resistant production for whatever downstream product depends on consistent supply. If the target product is something like DHA, where heterotrophic Schizochytrium-style production has already proven commercially viable at scale, the carbon penalty the model flags may be entirely worth paying. The model doesn't replace the TEA — it supplies one term in it, and a production decision made on that one term alone, in either direction, would be a misreading of what the model was built to do.

A bioinformatics intern uses a genetic algorithm to find the combination of light intensity, CO2 concentration, and nutrient levels the model predicts will maximize lipid accumulation in Chlorella protothecoides. Field trials at the pilot raceway show productivity roughly 40% below the model's prediction. Give at least three plausible reasons for the gap, and say what you'd change about how the model gets used going forward.

Show answer ↓

Three plausible reasons, all consistent with the limitations covered in this module. First, the genetic algorithm only optimized over the variables it was given — light, CO2, and nutrient levels — and outdoor raceway productivity is also shaped by variables the model never saw, such as dissolved oxygen buildup, hydraulic shear stress from the paddle wheel, or temperature swings between day and night, any of which can suppress lipid accumulation independently of the "optimal" inputs being met. Second, this is the lab-to-field gap directly: a genetic algorithm trained or validated on controlled, indoor conditions has no way to represent the variability of an open pond, where light penetration, evaporation, and contamination pressure change hour to hour in ways the training data never captured. Third, strain behavior itself can drift — a culture maintained through repeated outdoor cycles is not guaranteed to behave identically to the strain characterized in the lab, especially under the stress conditions a "maximize lipid accumulation" optimization typically pushes toward.

Going forward, the model should not be treated as producing a final optimum to implement once. It should be the first iteration of a loop: run the GA-predicted conditions at pilot scale, feed the actual field results — including the 40% shortfall — back into the model as new training data, and let the model's next prediction account for the gap rather than repeating it. This is the same logic as the GEM validation loop earlier in this module: the model proposes, the reactor disposes, and the disposal data is what makes the next round of modeling more useful than the last. A model whose predictions are never corrected by field outcomes isn't being used as a computational tool — it's being used as an excuse to skip validation.