Scientists Just Hacked Evolution’s “Randomness”? University of Tokyo Breakthroughs Hint We Can Forecast Life’s Next Moves

• University of Tokyo (UTokyo) researchers built “Evodictor,” a machine‑learning framework that predicted long‑term gene gains and losses across ~3,000 bacterial genomes—evidence that parts of evolution are predictable at the system level. PubMed
• UTokyo virologists unveiled “CoVFit,” a protein‑language‑model that predicted which SARS‑CoV‑2 spike mutations would boost variant fitness—correctly flagging changes later seen in BA.2.86 descendants (JN.1, KP.2, KP.3). Nature
• UTokyo biologists showed the repeatability of protein evolution: two different enzyme families (AdhE and BdhE) independently evolved near‑identical functional architectures—evidence for structural convergence that obeys discernible rules. Nature
• A UTokyo team also demonstrated lab methods to accelerate bacterial genome evolution (activating “jumping genes”), compressing decades of change into weeks—creating a testbed to validate evolutionary predictions. s.u-tokyo.ac.jp
• Across biology, prediction improves when selection is strong, constraints are known, and data are rich; leading evolutionary theorists argue that “data limits,” not pure randomness, are often the main barrier. Nature
• But prediction has layers: phenotypes can be repeatable while the exact DNA changes are not—hot‑temperature adaptation in seed beetles was highly repeatable for traits, less so for genomes. Nature
• Practical payoffs are already visible: Evodictor suggests ways to rank bacteria by risk of acquiring antibiotic resistance genes; CoVFit supports early warning for high‑risk viral variants; convergent enzymes point to routes for metabolic engineering (e.g., bioethanol). s.u-tokyo.ac.jp

Can evolution be predicted? A science‑based answer

For more than a century, biologists have sparred over whether evolution is a chaotic, one‑off historical process or a rule‑governed phenomenon that can be forecast. A consensus is emerging: we can predict some aspects of evolution—especially over short‑to‑medium time scales, under strong selection, or when physical/biochemical constraints are tight—provided we have enough data. Reviews in Nature Communications and related work lay out how predictive skill rises when we couple time‑series measurements with mechanistic and genomic models. Nature

The University of Tokyo has become a focal point for turning that philosophy into practice. Their teams have combined big‑data phylogenomics, protein AI, and lab evolution to show, case by case, that evolutionary outcomes can be forecast with useful accuracy—not everywhere, not always, but often enough to matter.

Case study #1 — Predicting bacterial genome evolution (Evodictor)

In January 2023, UTokyo researchers reported Evodictor, which learns from the history of bacterial gene gain/loss to forecast the future evolution of metabolic systems. Trained on ~2,894–3,000 bacterial species, the model predicted where and when genes would be added or dropped along a phylogeny—capturing universal constraints that shape bacterial metabolisms. Crucially, they also showed that even “ongoing” intraspecies changes are predictable when analyzed with pan‑genomes. This advances prediction beyond short‑term point mutations into long‑term, system‑level evolution. PubMed

UTokyo’s own summary underscores the practical angle: mapping predictable pathways may help prioritize pathogens by their likelihood of acquiring resistance genes, guiding antimicrobial strategy and synthetic‑biology design. s.u-tokyo.ac.jp

Case study #2 — Forecasting viral evolution with a protein language model (CoVFit)

In May 2025, UTokyo virologists published CoVFit in Nature Communications. Fine‑tuned from ESM‑2, CoVFit learns the relationship between spike‑protein sequences and variant fitness (relative epidemic growth). It not only ranks the fitness of unseen variants, but also pinpoints single amino‑acid substitutions likely to increase spread. Retrospectively and prospectively, CoVFit successfully identified positions (e.g., 346, 455, 456) later found in the fast‑growing BA.2.86 descendant lineages—exactly the kind of early signal public‑health surveillance needs. Nature

Takeaway: when the fitness landscape is constrained by immunology, receptor binding, and epidemiology—and when those constraints are encoded in data—evolutionary forecasting becomes actionable. Nature

Case study #3 — “Life finds the same solutions”: repeatable protein evolution

In September 2025, UTokyo researchers described two enzyme families (AdhE and BdhE) that independently evolved similar two‑step alcohol‑metabolism machinery via distinct gene‑fusion events—structural convergence at the scale of entire protein complexes. Even their inter‑domain interactions converged, despite different overall architectures (helical vs. ring‑like multimers). The work also uncovered a plausible rule: gene neighbors tend to fuse, creating a predictable route from genome organization to biochemical design. This is a rare structural‑level demonstration that evolution can be repeatable—and thus, partly predictable—far beyond single mutations. Nature

Beyond theory, the team noted possible bioethanol applications—reminding us that repeatability doesn’t just explain the past; it points to buildable futures in metabolic engineering. s.u-tokyo.ac.jp

Case study #4 — Speeding evolution to test predictions

Prediction begs validation. In May 2025, a UTokyo group accelerated bacterial genome evolution by activating mobile “jumping genes” (insertion sequences), compressing decades of structural change into roughly ten weeks. This creates a laboratory time machine: a way to stress‑test models like Evodictor under controlled conditions and observe large‑scale genome rearrangements in real time. s.u-tokyo.ac.jp

Where prediction shines—and where it still stumbles

What works now

1. Strong selection + clear constraints. When physics or physiology sets hard limits (e.g., viral entry via spike, metabolic network stoichiometry), models can capture the dominant routes evolution takes. Evodictor and CoVFit thrive here. PubMed
2. Aggregate or system‑level traits. It’s often easier to predict which pathways will change than the exact base pairs that will do the changing—hence success at metabolic gene content and variant fitness. PubMed
3. Convergence and repeatability. UTokyo’s AdhE/BdhE result shows that evolution can independently rediscover similar solutions, revealing rules (e.g., genome neighbors → fusion proteins → substrate channeling). Nature

What limits us

1. Genomic contingency. Even when traits converge, the underlying DNA changes can vary by genetic background and epistasis. In seed beetles, phenotypes adapted to heat in parallel, but genomic paths diverged—so cross‑population genomic predictions were weaker. Nature
2. Data scarcity beats determinism. A major barrier is often insufficient data to resolve selection, environment, and genetic architecture—not intrinsic unpredictability. Building broader, deeper datasets improves forecasts. Nature

Why this matters: from public health to bioengineering

• Pandemic readiness. Tools like CoVFit can triage new variants when only a sequence exists—spotting high‑risk mutations before case curves surge. Nature
• Antibiotic resistance. System‑level predictors (Evodictor) may help rank pathogens by likelihood of acquiring resistance genes, informing surveillance and stewardship. s.u-tokyo.ac.jp
• Green biotech. Understanding repeatable protein architectures can guide metabolic‑pathway design (e.g., efficient two‑step alcohol metabolism). s.u-tokyo.ac.jp

How UTokyo’s results fit the broader field

Global syntheses argue for an “engineering mindset”: define the predictive scope, time scale, and precision you need; assemble the data to match; and quantify uncertainty. Prediction is already improving across medicine, agriculture, and conservation—precisely the areas where UTokyo’s case studies land. Wiley Online Library

At the same time, the best recent experiments caution against overconfidence: phenotypic forecasts can be good, while gene‑by‑gene forecasts remain hard, especially across lineages and warming climates. That nuance sets a realistic ceiling on what “predictable evolution” means today. Nature

FAQ: What “predictable” does—and doesn’t—mean

• Does predictability mean pre‑ordained? No. Models yield probabilities, not certainties. Random mutation order, drift, and epistasis can derail exact‑path predictions, even when the end‑state trait is forecastable. Nature
• Is prediction only short‑term? Not necessarily. Evodictor shows long‑term (phylogenetic) patterns in gene content can be learned; CoVFit extrapolates to unseen variants. But as horizons extend, uncertainty grows—so we quantify and communicate it. PubMed
• What about origins‑of‑life style evolution? UTokyo experiments have even tracked Darwinian evolution in RNA replicators, offering a clean window into the earliest evolutionary rules—another arena where constraints foster predictability. The University of Tokyo

Timeline of the UTokyo predictability push (highlights)

• Mar 18, 2022 — RNA replicators evolve complexity in the lab (Nature Communications). The University of Tokyo
• Jan 12–13, 2023 — Evodictor predicts metabolic gene content evolution across bacteria (Science Advances). s.u-tokyo.ac.jp
• May 15, 2025 — Lab method accelerates bacterial genome evolution to validate predictions (Nucleic Acids Research). s.u-tokyo.ac.jp
• May 13, 2025 — CoVFit predicts fitness of SARS‑CoV‑2 variants and future mutations (Nature Communications). Nature
• Sep 22, 2025 — Repeatability of protein complex evolution demonstrated (Nature Communications). s.u-tokyo.ac.jp

Bottom line

Yes—evolution can be predicted, in meaningful ways. The University of Tokyo’s results show that when we learn the rules—from the structure of protein complexes to the fitness landscape of a virus to the logic of metabolic networks—evolution stops looking like noise and starts looking like a navigable map. The map isn’t perfect, but it’s getting sharper—and it’s already useful.

Sources & further reading

• Konno & Iwasaki. Sci Adv (2023): Machine learning enables prediction of metabolic system evolution in bacteria. UTokyo press explainer. PubMed
• Ito et al. Nat Commun (2025): A protein language model for exploring viral fitness landscapes. EurekAlert press release. Nature
• Konno et al. Nat Commun (2025): Repeatability of protein structural evolution following convergent gene fusions. UTokyo press summary. Nature
• Kanai et al. Nucleic Acids Research (2025): Laboratory evolution of bacterial genome structure via insertion sequences. UTokyo press. s.u-tokyo.ac.jp
• Nosil et al. Nat Commun (2020): Increasing our ability to predict contemporary evolution. Nature
• Rêgo et al. Nat Ecol Evol (2025): Repeatability of evolution vs. genomic predictability under thermal adaptation. Nature