The Future of Biocomputing: Roadmap and Open Problems

The future of biocomputing depends on solving a handful of hard engineering problems: growing larger, vascularized brain organoids that live for months, reading and writing their activity in three dimensions, standardizing the data, and giving researchers remote access. The first applications will likely be drug testing, disease modeling and neuroscience, not general computing.

Key takeaways

Biocomputing with living neurons is real but early. Today’s systems are research tools for studying how neural tissue processes information and adapts.
The organoid intelligence roadmap (Smirnova et al., 2023) sets out the core technical agenda: scale, vascularization, longevity, 3D interfaces, data standards and embedded ethics.
Commercial platforms already exist for experiments: FinalSpark’s remotely accessible Neuroplatform (Jordan et al., 2024) and Cortical Labs’ CL1, announced in March 2025.
Near-term value is most likely in pharmacology, disease modeling and basic research, where human neural tissue is useful even if it computes slowly.
Timelines for anything resembling a general-purpose biological computer are genuinely unknown. Clear milestones are more useful than dates.

Where does biocomputing stand today?

Biocomputing, in the sense we use here, means using living neurons, usually brain organoids or two-dimensional neural cultures, as part of a computing system. The term wetware computing describes the same idea, and organoid intelligence (OI) is the research program focused specifically on three-dimensional human brain organoids. Our complete guide to organoid intelligence covers the foundations.

A handful of results define the current state of the art:

DishBrain. Kagan et al. reported in Neuron (2022) that cultured neurons on a high-density array, receiving structured feedback in a closed loop, improved their performance at a simplified game of Pong compared with control conditions. The study is discussed in depth in our article on how neurons learn in a dish.
Brainoware. Cai et al. reported in Nature Electronics (2023) that a brain organoid used as a reservoir computer could support simple tasks, including speech recognition, with a conventional readout trained on the organoid’s responses. See reservoir computing with brain organoids.
Remote platforms. Jordan et al. (2024) described FinalSpark’s Neuroplatform, a system for running electrophysiology experiments on neural organoids remotely, covered in more detail below.

These are meaningful demonstrations that living networks can be coupled to software and that their responses change with stimulation. They are not yet demonstrations of useful computation that beats, or even approaches, conventional hardware on a practical problem. The honest picture is a field that has proved feasibility and is now working on the infrastructure needed to make progress measurable and repeatable.

What does the organoid intelligence roadmap say?

The most detailed public roadmap is the paper by Smirnova et al. in Frontiers in Science (2023), published alongside the Baltimore declaration (Hartung et al., 2023), a community statement from the first Organoid Intelligence workshop, held online in February 2022. The roadmap identifies a set of technical needs that recur throughout this article: more cells and more complex organoids, perfusion to replace missing blood vessels, long-term culture, three-dimensional electrodes and imaging, machine learning for analysis, and open, standardized data. It pairs these with an embedded ethics program, discussed in our article on the ethics of organoid intelligence.

The table below summarizes the main open problems, where the field stands and what progress would look like.

Challenge	Current state	What progress looks like
Scale	Many organoids under 500 micrometers and fewer than 100,000 cells (Smirnova et al., 2023)	Organoids of around 10 million cells, the roadmap’s proposed target
Vascularization	Most organoids are avascular; diffusion limits healthy tissue to a few hundred micrometers from the surface	Perfusable organoids, engineered vessels or microfluidic substitutes
Longevity	Research platforms report organoid lifetimes of more than 100 days (Jordan et al., 2024)	Stable function over many months, with predictable aging
Reproducibility	Variation between batches, cell lines and labs	Protocols and quality metrics that give comparable organoids and recordings
Interfaces	Mostly flat arrays touching the organoid’s base	3D electrodes and optical methods sampling the whole volume
Data	Mixed formats and metadata across labs	Shared standards such as NWB, public reference datasets
Training	Short-lived, task-specific adaptation	Reliable, lasting learning with clear benchmarks

How big do organoids need to be?

Scale is the most intuitive gap. Smirnova et al. describe many current organoids as below 500 micrometers in diameter with fewer than 100,000 cells, and propose roughly 10 million cells as a target for OI work. For comparison, Azevedo et al. (2009) estimated that the adult human brain contains about 86 billion neurons. Even the roadmap’s target would be several thousand times smaller than a brain.

More cells alone are not the point. What matters for computation is organized connectivity: diverse cell types, layered structure and long-range connections. That is why researchers are also building assembloids, which fuse organoids patterned to different brain regions so they can connect with each other.

Why is vascularization the bottleneck?

Bigger organoids hit a hard physical limit. Without blood vessels, oxygen and nutrients must diffuse in from the surface, and Smirnova et al. note that diffusion can only supply cells within a few hundred micrometers of the surface, so cells in the core begin to die. That is why organoid vascularization is near the top of every roadmap.

Several approaches are being explored:

Host vascularization. Mansour et al. reported in Nature Biotechnology (2018) that human brain organoids transplanted into mouse brains were vascularized by the host and showed functional activity. This shows what blood supply does, but it is not a route for in vitro biocomputing.
Engineered vessels. Cakir et al. reported in Nature Methods (2019) that engineering stem cells to express the transcription factor ETV2 produced a functional vascular-like system inside brain organoids.
Microfluidic perfusion. The OI roadmap proposes microfluidic systems that deliver medium and remove waste in a controlled way, substituting for vasculature.
Slicing. Giandomenico et al. reported in Nature Neuroscience (2019) that culturing organoid slices at an air-liquid interface improved survival and allowed extensive axon tracts to form, sidestepping the diffusion problem by keeping tissue thin.

None of these yet produces a fully perfused, brain-like vascular network in a dish. A robust, reproducible method for perfusing large organoids would be one of the most important milestones the field could reach.

How long can organoids live?

Computing systems need stable hardware. Neural tissue changes continuously: it matures, rewires and eventually degrades. Longevity therefore matters both for running long experiments and for knowing whether a change in behavior reflects learning or simply aging.

Progress here has been real. Gordon et al. reported in Nature Neuroscience (2021) that human cortical organoids maintained in long-term culture showed molecular transitions matching key early postnatal stages, which suggests extended culture lets organoids mature further. Jordan et al. (2024) reported that the Neuroplatform runs experiments on organoids with lifetimes of more than 100 days, supported by automated microfluidic medium exchange. Cortical Labs states on its product page that the CL1’s life support is designed to keep neurons alive for up to six months.

The open problem is not just survival but stability: keeping a network’s baseline activity predictable enough that experiments months apart can be compared.

Can organoids be made reproducible?

Reproducibility is arguably the least glamorous and most important problem. Organoids vary between batches, between cell lines and between labs. Velasco et al. reported in Nature (2019) that individual dorsal forebrain organoids made with their protocol reproducibly generated the range of cell types found in the developing human cerebral cortex, an encouraging result. On the other hand, Bhaduri et al. reported in Nature (2020) that cortical organoids show elevated cellular stress that impairs the specification of neuronal subtypes, a reminder that culture conditions shape what organoids become.

For computing, reproducibility also means functional reproducibility: similar firing rates, similar network bursts, similar responses to the same stimulation protocol. That requires agreed quality metrics, shared protocols and public reference datasets, which brings us to interfaces and data.

How will we read and write activity in three dimensions?

Most organoid electrophysiology today uses a microelectrode array (MEA), a flat grid of electrodes that the organoid sits on. Our explainer on microelectrode arrays covers how they work. High-density microelectrode arrays, built on CMOS chips with thousands of electrodes, have made it possible to record finely detailed activity. Sharf et al. used high-density CMOS arrays and shank electrodes in Nature Communications (2022) to record spontaneous activity from human brain organoids and infer functional connectivity from spike timing.

The limitation is geometry. A flat array only samples the surface touching it, while an organoid is a three-dimensional object. Several groups are building interfaces that reach further:

Shell MEAs. Huang et al. reported in Science Advances (2022) self-folding, optically transparent shells carrying electrodes that wrap around organoids, inspired by EEG caps, and recorded from 400 to 600 micrometer organoids for up to four weeks.
Mesh electronics. Le Floch et al. reported in Advanced Materials (2022) stretchable mesh nanoelectronics that become integrated into organoids as they develop, providing stable long-term recording from single cells inside the tissue.
3D frameworks. Park et al. reported in Science Advances (2021) three-dimensional, multifunctional interfaces designed to enclose cortical spheroids and engineered assembloids for recording and stimulation.
Probes and imaging. The OI roadmap discusses penetrating probes and fast volumetric imaging as complementary ways to sample activity throughout an organoid.

Reading is only half the problem. Biocomputing also needs precise writing: stimulation patterns that reliably drive specific responses and induce synaptic plasticity in controlled ways. Better neural encoding and decoding methods, which translate task inputs into stimulation and recorded activity into outputs, are as important as better electrodes. Upstream of decoding sits spike sorting, which assigns recorded spikes to individual neurons and becomes harder as electrode counts grow.

Why do data standards matter for biocomputing?

A single long organoid experiment can generate enormous amounts of data. Jordan et al. reported that over three years the Neuroplatform was used with more than 1,000 brain organoids and collected more than 18 terabytes of data. Without shared formats, data at that scale is hard to reuse, compare or audit.

Neurodata Without Borders (NWB) is the leading open standard for neurophysiology data. Rübel et al. described in eLife (2022) how NWB provides a common language for recordings and their metadata across modalities and species, so data can be findable, accessible, interoperable and reusable. The OI roadmap names NWB and the Brain Imaging Data Structure (BIDS) among the standards the field should adopt, along with public archives and reference datasets.

For biocomputing specifically, standardized data should capture more than voltages. Useful metadata includes:

Cell line identity, differentiation protocol and organoid age.
Culture conditions, medium changes and any drug exposures.
Electrode layout and recording settings.
Every stimulation event, with the closed-loop logic that produced it.
Consent scope and provenance for the donor cells.

When stimulation protocols and outcomes are recorded consistently, claims such as “this network learned” can be checked by others. That is the difference between a demo and a science.

How will researchers access biocomputers?

Few labs can grow organoids, run electrophysiology, and build machine learning pipelines all at once. Cloud access models split those skills: one team maintains the wet lab and hardware, while researchers anywhere design experiments in software.

FinalSpark’s Neuroplatform

Jordan et al. described the Neuroplatform in Frontiers in Artificial Intelligence (2024). According to the paper, it supports round-the-clock electrophysiology recording and electrical stimulation of neural organoids, uses a microfluidic system for automated medium exchange to reduce disruptions, and exposes an application programming interface for remote research through a Python library or interactive tools such as Jupyter notebooks. The API also controls pumps, cameras and UV light for molecule uncaging, enabling closed-loop experiments that use standard deep learning or reinforcement learning libraries. The authors stated that in 2024 the system was freely available for research purposes and that numerous research groups had begun using it.

Cortical Labs and the CL1

Cortical Labs, the company behind the DishBrain study, announced its CL1 system in March 2025, as reported by IT Brief Australia. Cortical Labs’ product page describes the CL1 as a closed-loop system in which neurons grow on a silicon chip that sends and receives electrical signals, with on-device recording and life support designed to keep neurons alive for up to six months. The company also offers remote access through what it calls the Cortical Cloud. Kagan et al. (2023), writing in Biotechnology Advances, set out a broader framing of the field, which they call synthetic biological intelligence.

Academic groups

Much of the foundational work comes from academic labs, including the groups behind the OI roadmap and the Brainoware study, and the many organoid and electrophysiology labs whose methods this field builds on. Academic research will likely remain central for the open problems above, especially vascularization, reproducibility and interfaces.

Access models compared

Model	How it works	Strengths	Limitations
In-house lab	Grow, record and analyze locally	Full control over cells and protocols	Requires wet lab, hardware and specialist staff
Remote platform	Run experiments via API on someone else’s organoids	Low barrier to entry, scales to many users	Less control over cell lines, culture and hardware choices
Purchased system	Buy a self-contained unit for your own lab	Local control with integrated hardware	Still requires cell culture capability
Open datasets	Reanalyze recordings shared in standard formats	Cheap, reproducible, good for method development	No ability to stimulate or run new experiments

The most likely future is a mix of all four, connected by shared data standards and reproducible protocols.

What will biocomputing be used for first?

The most plausible early applications are ones where human neural tissue is valuable for its biology, even if its computing performance is modest.

Drug discovery and neurotoxicity

Human neurons respond to drugs and toxins in ways animal models sometimes do not. Pamies et al. described in ALTEX (2017) an iPSC-derived brain microphysiological system for studying neurological disease and toxicity. Closed-loop biocomputing platforms add a new kind of readout: not just whether a compound changes firing, but whether it changes a network’s ability to adapt to feedback. That could become a functional assay for compounds that affect learning or plasticity, though such assays still need validation against clinical outcomes.

Disease modeling

Organoids made from patients’ induced pluripotent stem cells carry those patients’ genetics. Lancaster et al. used cerebral organoids to model microcephaly in Nature (2013), and Qian et al. used brain-region-specific organoids to model Zika virus exposure in Cell (2016). Biocomputing tools could extend this by testing how disease-associated variants affect network-level information processing, not only structure and cell composition.

Neuroscience research

Organoid systems let researchers study how networks self-organize and adapt with complete control over inputs. Trujillo et al. reported in Cell Stem Cell (2019) that cortical organoids develop increasingly complex oscillatory activity over months. Closed-loop platforms make it possible to test theories of learning, such as Hebbian learning, spike-timing-dependent plasticity or the free energy principle, directly in human tissue.

Low-power learning research

The motivation that gets the most attention is energy. The International Energy Agency estimated in its 2025 Energy and AI report that data centers used about 415 TWh of electricity in 2024, around 1.5% of global electricity consumption. Brains learn from little data with little power, and understanding how could inform new algorithms and hardware, including neuromorphic computing and spiking neural networks. Whether organoid systems themselves will ever be energy-efficient computers is a separate, open question: no system has yet shown efficient computation on a useful task once life support and electronics are counted. We discuss this in biocomputing vs silicon.

How long will it take?

Honest answer: nobody knows. The field is young, the problems are hard, and progress in biology rarely follows a predictable curve. Any article that gives confident dates for commercial biological computers is guessing.

A more useful way to think about timelines is in terms of what is already happening, what is plausible with steady progress, and what remains speculative.

Horizon	What it includes	Confidence
Already happening	Remote experiment platforms, closed-loop training studies, reservoir computing demonstrations, organoids for disease modeling	Established in published work
Plausible with steady progress	Standardized datasets, better 3D interfaces, longer-lived organoids, validated drug-response assays	Depends on engineering progress and funding
Uncertain	Reliably vascularized large organoids, lasting trained behavior, competitive performance on niche tasks	Significant open research problems
Speculative	General-purpose biological computers, energy advantage over silicon at scale	No current evidence it is achievable

The engineering details of how tissue, electrodes and controllers fit together are covered in bio-silicon hybrid computing, and the hardest open ethical question, whether larger and longer-lived organoids could ever have experiences, is examined in can brain organoids become conscious?.

What would count as real milestones?

Because dates are unreliable, milestones are the better way to track the field. These are the results that would genuinely change the picture:

Reproducible learning across labs. The same protocol, applied to organoids grown in different labs from different cell lines, producing comparable improvement on a defined task, with data shared in a standard format.
Lasting memory. Trained changes in behavior that persist for days or weeks without continued feedback, distinguishable from ordinary drift.
Perfused, large organoids. A reproducible method for keeping organoids of millions of cells healthy throughout their volume.
Whole-volume interfaces. Recording and stimulation that sample the organoid’s interior at scale, stably over months.
Benchmarks. Agreed tasks and metrics, including energy accounting that includes life support, so systems can be compared fairly with each other and with silicon.
Validated applications. A drug or toxicity assay on a biocomputing platform that predicts human outcomes better than existing methods.
Ethics infrastructure. Consent, provenance and review processes that scale with the technology, as called for in the Baltimore declaration.

Notice that most of these are infrastructure milestones rather than headline-grabbing demos. That is typical of fields moving from proof of concept to engineering.

What comes next for the future of biocomputing?

The future of biocomputing will be built less by single breakthroughs than by steady improvement in the parts that make experiments trustworthy: organoids that are bigger, perfused and long-lived; interfaces that reach inside them; data that anyone can reanalyze; and access models that let researchers without wet labs contribute. The first payoffs will most likely come in pharmacology, disease modeling and neuroscience, with computing applications following only if training methods mature.

That is also why software matters as much as biology. Reproducible protocols, closed-loop runtimes, experiment records and NWB-native data are the tools that turn individual demonstrations into a cumulative science. Building that layer is what we are working on at Organoid Programming.

Frequently asked questions

What is the future of biocomputing?

The near future of biocomputing is mostly research infrastructure: larger and longer-lived brain organoids, vascularization, three-dimensional electrode interfaces, standardized data and remote cloud access. The first practical uses are likely to be in drug testing, disease modeling and neuroscience. General-purpose biological computers that compete with silicon remain a long-term, uncertain goal with no agreed timeline.

When will biological computers be available?

Research systems already exist. FinalSpark described a remotely accessible Neuroplatform in 2024, and Cortical Labs announced its CL1 system in March 2025, offering hardware and cloud access to cultured neurons. These are tools for experiments, not replacements for conventional computers. Nobody can credibly date when biocomputers will do commercially useful computation.

What are the biggest open problems in organoid intelligence?

The main open problems are scale (organoids have far fewer cells than brains), vascularization (without blood vessels, organoid cores starve), longevity (cultures must stay healthy for months), reproducibility between batches and labs, interfaces that reach inside 3D tissue, standardized data, and reliable training methods that produce lasting, measurable learning.

How big are brain organoids today and how big need they be?

Smirnova and colleagues (2023) described many current organoids as under 500 micrometers across with fewer than 100,000 cells, and proposed scaling to around 10 million cells for organoid intelligence work. A human brain has about 86 billion neurons, so even that target is a tiny fraction of brain scale.

What will biocomputing be used for first?

The most likely first uses are drug discovery and neurotoxicity testing, disease modeling with patient-derived cells, basic neuroscience on how networks learn, and research into low-power learning algorithms. These uses benefit from human neural tissue even when it computes poorly, which makes them more realistic near-term targets than general computing.

Why does biocomputing need data standards like NWB?

Organoid experiments generate large, complex recordings, and results are hard to compare across labs without shared formats. Neurodata Without Borders (NWB) is an open data standard for neurophysiology that describes recordings and their metadata consistently. Using it lets researchers reanalyze each other’s data, combine datasets and check claims, which a young field especially needs.

Is biocomputing more energy efficient than silicon?

Not yet. Brains learn with very little power, which motivates the idea, but no organoid system has yet shown energy-efficient computation on a useful task once life support, electronics and incubation are counted. Energy efficiency is a long-term research hypothesis, not a demonstrated advantage of current biocomputing systems.

The future of biocomputing: the roadmap, open problems and what comes next

Where does biocomputing stand today?