Lab Newsletter — September 8, 2026: Climbing the Fitness Landscape

On Friday we watched AI design proteins from a blank page, and yesterday we watched it read a genome base by base. Today the verb is different and, in a way, humbler: evolve. Nature already handed us a working parts list — enzymes, binders, fluorescent proteins — that are almost, but not quite, what an experiment needs. The classic way to improve one is directed evolution: mutate, test, keep the winners, repeat. It works so well that Frances Arnold shared a Nobel Prize for it. The question this digest follows today is what happens when you put a machine-learning model inside that loop — when, instead of testing mutations blindly, you predict which ones will help. The mental picture is a fitness landscape: every sequence a point, its activity the elevation, and engineering the act of climbing toward a peak without getting lost.

🧭 The paradigm: learn the map, then climb

The clearest statement of the idea comes from a review by Yang, Wu & Arnold (Nature Methods, 2019): “Protein engineering through machine-learning-guided directed evolution enables the optimization of protein functions.” Why reach for ML at all? Because these models “predict how sequence maps to function in a data-driven manner without requiring a detailed model of the underlying physics or biological pathways” — you don’t need to understand why a mutation helps to learn that it does. And the payoff is speed: such methods “accelerate directed evolution by learning from the properties of characterized variants and using that information to select sequences that are likely to exhibit improved properties.” Test a few, learn the local shape of the landscape, and let the model point you uphill.

🌀 Reading the landscape without a single label

Before you run any assay, evolution has already left you a map: the millions of related sequences that survived selection over deep time. DeepSequence (Riesselman, Ingraham & Marks, Nature Methods, 2018) learned to read it. The insight is that residues don’t act alone — “the functions of proteins and RNAs are defined by the collective interactions of many residues, and yet most statistical models of biological sequences consider sites nearly independently.” Their deep generative model captured those interactions, and, “learned in an unsupervised manner solely on the basis of sequence information,” it “predicted the effects of mutations across a variety of deep mutational scanning experiments substantially better than existing methods based on the same evolutionary data.” A fitness prediction for every possible point mutation — from sequence alone, no wet lab required.

🧩 A representation you can engineer with

DeepSequence models one protein family at a time; the next move was to learn a general language of proteins. UniRep (Alley, Khimulya, Biswas, AlQuraishi & Church, Nature Methods, 2019) did exactly that, applying “deep learning to unlabeled amino-acid sequences to distill the fundamental features of a protein into a statistical representation that is semantically rich and structurally, evolutionarily and biophysically grounded.” Crucially the simplest models built on top of it “are broadly applicable and generalize to unseen regions of sequence space” — and it earns its keep on real tasks, predicting “the stability of natural and de novo designed proteins, and the quantitative function of molecularly diverse mutants,” delivering “two orders of magnitude efficiency improvement in a protein engineering task.” This is the protein-language-model idea in an early, concrete form: learn once from raw sequence, reuse everywhere.

🎯 Twenty-four experiments, ten million candidates

The dream of all this is to spend fewer experiments. Low-N (Biswas, Khimulya, Alley, Esvelt & Church, Nature Methods, 2021) made the number startling. Protein engineering, they note, “is limited by the lack of experimental assays that are consistent with the design goal and sufficiently high throughput to find rare, enhanced variants.” Their answer: a paradigm that “can use as few as 24 functionally assayed mutant sequences to build an accurate virtual fitness landscape and screen ten million sequences via in silico directed evolution.” And it isn’t a one-protein trick — “as demonstrated in two dissimilar proteins, GFP from Aequorea victoria (avGFP) and E. coli strain TEM-1 β-lactamase, top candidates from a single round are diverse and as active as engineered mutants obtained from previous high-throughput efforts.” Twenty-four measurements in, a ten-million-wide search out: that is the automated-discovery dream in miniature.

⚖️ The sober lesson: fuse the two data sources — and keep score

With evolutionary models on one side and assay measurements on the other, which should you trust? Hsu, Nisonoff, Fannjiang & Listgarten (Nature Biotechnology, 2022) asked plainly, noting that fitness models “typically learn from either unlabeled, evolutionarily related sequences or variant sequences with experimentally measured labels.” Their finding is the kind this digest keeps running into: a simple combination wins. They “propose a simple combination approach that is competitive with, and on average outperforms more sophisticated methods” — “ridge regression on site-specific amino acid features combined with one probability density feature from modeling the evolutionary data.” The deeper takeaway is methodological honesty: their “analysis highlights the importance of systematic evaluations and sufficient baselines.” A recurring refrain — prove it, and beat an honest baseline before you claim the win.

🔬 Closing the loop at the bench

Prediction only matters if it changes what comes out of a flask. Wu, Kan, Lewis, Wittmann & Arnold (PNAS, 2019) closed the loop. The motivation is economic: “combinatorial sequence space can be quite expensive to sample experimentally, but machine-learning models trained on tested variants provide a fast method for testing sequence space computationally.” They “validated this approach on a large published empirical fitness landscape for human GB1 binding protein, demonstrating that machine learning-guided directed evolution finds variants with higher fitness than those found by other directed evolution approaches” — then took it to new chemistry, engineering an enzyme that “fixed seven mutations in two rounds of evolution to identify variants for selective catalysis with 93% and 79% ee (enantiomeric excess).” Model proposes, bench disposes, model updates: design–build–test–learn, made real.

🧬 Why it’s our kind of problem

Reading a protein’s function, writing new ones from scratch, and evolving the ones we have are three verbs on the same molecule — and today’s is the one that most looks like a loop. That loop is precisely what a self-driving lab automates and what an AI co-scientist orchestrates: the model nominates a handful of variants, the bench measures them, the landscape sharpens, and the next round climbs higher. Low-N’s “24 assays → ten million in silico” is that flywheel in one sentence. It also feeds the virtual cell: to simulate or re-engineer a pathway, you have to predict how each mutation changes a protein’s behavior — a fitness landscape for every player. And the way these tools travel is our ethos exactly. DeepSequence and UniRep are open source; deep mutational scanning datasets and the GB1 landscape are the public yardsticks everyone is scored against — the same publish-the-model-and-the-test spirit behind the BioImage Model Zoo and BioEngine. A fitness predictor you can call like a service, benchmarked on a shared landscape, and looped straight into an experiment: that’s protein engineering starting to run itself.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.