NeuroTrailblazers banner

Open Problems for Undergraduate Teams

This page is a working map of open problems in connectomics that satisfy two constraints at once: an undergraduate team can make real progress on them within one to two semesters, and progress on them matters to the field’s current scaling effort — the NIH BRAIN Initiative CONNECTS program and its flagship projects, including MouseConnects/HI-MC.

“Open problem” here does not mean “unsolved grand challenge.” It means a question where the answer is genuinely unknown, where the data and tools to attack it are publicly available, and where a competent team’s result — positive or negative — would be worth reporting. Connectomics is unusual among neuroscience subfields in how many such questions it currently has. The field’s data production has outrun its analysis capacity: the fly brain connectome alone (about 140,000 neurons and roughly 50 million synapses) was released with the explicit expectation that most of its science would be done by people outside the labs that built it, and as of early 2026 more than 60,000 users had signed into Codex, its browser-based explorer. Electron-microscopy-based connectomics was named Nature Methods’ Method of the Year 2025 — recognition that the data exists; the open question is what the community does with it and how the pipelines scale next.

Why now, and why undergraduates

Three structural facts make this the right moment:

  1. The data is public and browser-accessible. FlyWire (adult female fly brain), the Janelia male CNS connectome (the first complete male fly central nervous system, brain plus ventral nerve cord with an intact neck connective), MICrONS (a cubic millimetre of mouse visual cortex with co-registered functional imaging — roughly 200,000 cells and over half a billion synapses), H01 (a petabyte-scale fragment of human temporal cortex), the larval Drosophila CNS (3,016 neurons, Winding et al. 2023), and the eight developmental C. elegans connectomes of Witvliet et al. (2021) are all downloadable or queryable without institutional credentials. See Datasets for access routes.

  2. The analysis tooling is free and runs on a laptop. CAVEclient, navis, neuprint-python, DotMotif, NetworkX/igraph, and Neuroglancer cover the whole path from raw query to publishable figure. Most of the projects below need no cluster; a Colab notebook is enough.

  3. The scaling effort has known, named bottlenecks. CONNECTS — a roughly $150 million NIH program launched in 2023 with eleven funded projects — exists because going from 1 mm³ (MICrONS) to 10 mm³ (MouseConnects) to a whole mouse brain (~500 mm³, on the order of an exabyte of raw imagery at synaptic resolution) breaks every stage of the current pipeline. Each break is an open problem, and several of them are more limited by careful benchmarking and analysis — undergraduate-shaped work — than by resources only large labs have.

What follows is organized as seven problem areas. For each: what is actually open, why CONNECTS needs it solved, the current state of the art with numbers, and concretely scoped projects with entry points.

Ready to start rather than read? Every problem below is also a card with a five-step on-ramp — read, get the data, reproduce a known number, first measurement, scope the semester — at Open Problems. The data step of every ramp runs through Getting Started with Data, which exists because data access is the step most teams stall on.

Problem 1 — The proofreading budget: automated error detection and “how much is enough?”

What is open. Automated segmentation still makes merge errors (two neurons fused into one object) and split errors (one neuron broken into fragments), and humans currently fix them. FlyWire took a community of hundreds of proofreaders years to finish one fly brain. Complete manual proofreading of MouseConnects’ 10 mm³ is widely acknowledged to be infeasible, and a whole mouse brain is out of the question. Two questions are open. First, can error detection and correction be automated well enough to replace most human effort? Second — less glamorous and at least as important — how much proofreading does a given scientific claim actually require? Nobody has a principled answer to “this motif result is robust once the volume is X% proofread.”

Why CONNECTS needs it. Proofreading is the single largest human cost in the pipeline and the stage that scales worst. Every CONNECTS project’s feasibility case assumes an order-of-magnitude reduction in per-neuron human effort.

State of the art. NEURD (Nature, 2025) decomposes reconstructed neurons into graphs of processes and applies rule-based automated proofreading plus feature extraction, and worked at MICrONS scale. ConnectomeBench (2025) tested whether large multimodal language models can do proofreading subtasks: current models reached 52–82% balanced accuracy on segment-type identification (chance 20–25%) and 75–85% on split-error correction posed as multiple choice (chance 50%), but performed poorly at detecting merge errors — the error class that matters most, because merges silently create false edges in the graph. Supervised merge/split-detection classifiers trained on edit histories are an active area with no dominant solution. On the “how much is enough” question, the honest state of the art is ad hoc: papers report proofreading coverage but rarely show how their conclusions vary with it.

What an undergraduate team can do.

Entry points. CAVEclient + MICrONS/FlyWire materializations; NEURD’s public code; ConnectomeBench’s task definitions; FlyWire edit history via CAVE.

Problem 2 — Synapse detection that generalizes

What is open. Synapse detectors are typically trained per-dataset, on one species, one staining protocol, one microscope. Performance degrades — sometimes badly — under domain shift to a new sample, and quantifying that degradation requires ground truth that mostly does not exist outside a few well-annotated volumes. Detecting the sign (excitatory/inhibitory) and strength (synapse size as a proxy) adds further open layers. Cross-species generalization is explicitly unsolved: 2025 work on generalized synapse detection across invertebrate species is an early attempt, not a solution.

Why CONNECTS needs it. A whole-mouse-brain effort cannot afford to re-annotate ground truth and retrain for every region and every sample. Worse, region-dependent detector bias would masquerade as biology: if the detector finds 10% fewer synapses in hippocampus than cortex for purely image-statistical reasons, every cross-region comparison inherits that artifact.

State of the art. Modern detectors reach high F1 on the domain they were trained on; the reviews accompanying Method of the Year 2025 credit AI-driven reconstruction with an order-of-magnitude error reduction over the past decade, while flagging generalization and validation as the open front. Public ground-truth synapse annotations exist for CREMI (fly), MICrONS subvolumes, and scattered smaller sets.

What an undergraduate team can do.

Entry points. CREMI challenge data; MICrONS synapse tables via CAVEclient; published detector weights (e.g., from the MICrONS and FlyWire pipelines).

Problem 3 — Cell types: matching across individuals, sexes, species, and modalities

What is open. Cell typing in connectomics is done by morphology and connectivity; transcriptomic atlases type cells by gene expression. Whether these taxonomies align — and how types correspond across individual animals, sexes, and species — is largely open. The 2025 male fly CNS release made this concrete: comparing male and female fly brains identified 262 sex-specific and 114 sexually dimorphic cell types, about 4.8% of central-brain types. That comparison was possible because fly cell types are stereotyped; in mouse cortex, where types are statistical rather than stereotyped, cross-dataset matching is much harder and mostly unsolved.

Why CONNECTS needs it. MouseConnects will produce hippocampal neurons by the hundred thousand. Without reliable automated typing, downstream analysis (connectivity by type, comparison to cortical data, links to Alzheimer’s-relevant cell populations) stalls. And because CONNECTS spans multiple projects and species (mouse, macaque), cross-dataset type matching is the only way their results compose into one body of knowledge.

State of the art. NBLAST and its successors match neurons by morphology within a species; connectivity-based typing (clustering neurons by their partners) worked well in the fly hemibrain and FlyWire; MICrONS papers demonstrated coarse type prediction from EM features. Cross-modality (EM ↔ transcriptomics) correspondence is established for major classes, sketchy for subtypes.

What an undergraduate team can do.

Entry points. navis + NBLAST; FlyWire Codex annotations; neuprint for hemibrain/male CNS; MICrONS cell-type tables; Allen Institute transcriptomic atlases for the cross-modality angle.

Problem 4 — Graph algorithms and statistics at connectome scale

What is open. Connectome graphs are now large enough (10⁵ nodes, 10⁷–10⁸ edges) that exact versions of the field’s favourite analyses are intractable, and the statistical foundations are shakier than the literature admits. Three open strands. (a) Combinatorial optimization: ordering the fly connectome to maximize feedforward flow is a minimum-feedback-arc-set problem; FlyWire’s open Minimum Feedback Challenge stands at 35,463,823 of 41,912,141 forward-pointing synapse-weights (~85%), set in January 2026, with no known optimality bound — beating it earns a leaderboard spot and is pure algorithms work. (b) Null models: motif significance depends entirely on the null (degree-preserving vs Erdős–Rényi vs geometry-aware), and there is no consensus null that respects spatial embedding, multi-synapse edges, and cell types simultaneously — see Motif analysis. (c) Stability: community-detection and embedding results on connectomes are rarely reported with stability analysis across algorithm seeds, resolution parameters, or edit-history versions.

Why CONNECTS needs it. Every one of these gets an order of magnitude harder at 10 mm³. If the analysis stack’s conclusions are not stable at fly scale, they will be meaningless at mouse-hippocampus scale. Scalable approximate algorithms with error bounds, and null models that survive peer scrutiny, are prerequisites for the science CONNECTS is funding the data collection for.

What an undergraduate team can do. This is the cheapest entry point on the page — a math/CS team needs no biology background to start, though the interpretation payoff grows with it.

Entry points. Codex data downloads; the MFAS challenge page; DotMotif; graph-tool/igraph for nulls at this scale (NetworkX becomes the bottleneck around 10⁵ nodes — discovering exactly where is itself useful telemetry).

Problem 5 — Structure to function: prediction and its limits

What is open. The founding bet of connectomics is that wiring constrains function. It is now testable, and the early answers are mixed in interesting ways. Connectome-constrained models of the fly visual system (Lappalainen et al., Nature 2024) predicted neural responses of unrecorded cell types from wiring plus task training — a landmark positive. But 2025 work in Cell on MICrONS-style functional connectomics found that connectomic predictions are accurate for some response properties (orientation tuning) and surprisingly poor for others (receptive field size), with infrequent strong connections carrying disproportionate weight; and Nature Neuroscience work on connectome-constrained recurrent networks concluded the connectome alone is generally insufficient — pairing it with sparse recordings is what makes activity in unrecorded neurons predictable. Mapping where the predictive boundary lies — which functional properties wiring determines, which it merely biases, and which it leaves free — is the central open scientific question of the field.

Why CONNECTS needs it. It is the justification for the whole program. If function cannot be usefully constrained by structure, a mouse connectome is an expensive atlas; if it can, it is a foundation for mechanistic neuroscience and disease modelling. Every negative result that sharpens the boundary changes how the next billion dollars of acquisition is prioritized (e.g., how much co-registered functional data CONNECTS projects must collect).

State of the art. Besides the above: ZAPBench (ICLR 2025) provides a whole-brain cellular-resolution activity-forecasting benchmark in larval zebrafish, built to meet the forthcoming zebrafish connectome; current baselines show video-based models beating trace-based ones, and the connectome-conditioned frontier is essentially unexplored.

What an undergraduate team can do.

Entry points. ZAPBench data and code; the flyvis package (Lappalainen et al.’s released model); MICrONS functional data via the public releases; NeuroAI bridge.

Problem 6 — Data logistics at exabyte scale

What is open. A whole mouse brain at synaptic resolution is on the order of an exabyte of raw imagery; a human brain estimate runs to zettabytes. Storage, compression, formats, and access patterns at that scale are unsolved in practice: what lossy compression is scientifically safe (does 10× compression change segmentation output? synapse detection output?), how should petabyte volumes be chunked and served for the access patterns proofreaders and analysts actually have, and how can acquisition QA (detecting folds, debris, focus drift, misalignment) be automated well enough that errors are caught in hours rather than after months of downstream compute?

Why CONNECTS needs it. Several CONNECTS awards are explicitly for scalable acquisition and “intelligent image acquisition and reconstruction.” Storage costs are a first-order line item in every proposal; safe lossy compression at even 5× changes project budgets by millions. Acquisition QA failures are the most expensive kind: re-imaging destroyed tissue is impossible.

State of the art. The stack that got the field to petabyte scale — precomputed/sharded formats, OME-Zarr, CloudVolume, CAVE (published in Nature Methods, 2025), Neuroglancer — is public and battle-tested at 1 mm³. The exabyte-scale question marks (compression safety, tiered storage economics, QA automation) are documented as open in the field’s own scaling reviews.

What an undergraduate team can do. This is the natural home for a pure CS/data-engineering team.

Entry points. CloudVolume + public MICrONS/H01/FlyWire image volumes; OME-Zarr tooling; Data formats and Reconstruction pipeline.

Problem 7 — The human side: annotation science and distributed proofreading

What is open. Connectomics has quietly run some of the largest distributed scientific-annotation efforts ever — Eyewire’s hundreds of thousands of registered players, FlyWire’s citizen-scientist community that out-proofread professionals in places — but the science of that work is thin. Open questions: what actually predicts annotator accuracy and how fast do learners reach expert agreement; which errors do novices systematically make (and can training target them); how should consensus be aggregated when annotators disagree; what task designs keep volunteers engaged without degrading quality? These are tractable human-subjects/HCI questions with direct pipeline impact.

Why CONNECTS needs it. Even heavily automated pipelines keep humans in the loop for the hard residue, and MouseConnects-scale efforts plan on distributed communities — including, explicitly, NeuroTrailblazers-trained students. Every percentage point of annotator accuracy and retention changes the labour budget of a five-year project. This is also the one problem area where an education-, psychology-, or HCI-leaning team has the home advantage.

What an undergraduate team can do.

Entry points. FlyWire/Eyewire published community analyses; CAVE edit logs; this program’s own worksheets as standardized training instruments.

Choosing: a decision protocol

Four questions, in order, will land a team on the right problem:

  1. What is the team’s strongest existing skill? Algorithms/math → Problem 4. ML engineering → Problems 1, 2, 5. Biology/microscopy → Problems 2, 3. Systems/data engineering → Problem 6. Psychology/education/HCI → Problem 7. Choosing a problem that needs a skill nobody has is the most common failure mode; the biology can be learned alongside a skill-matched problem, but not instead of one.
  2. Is the data truly in hand in week one? Every project above is scoped to public data. Verify by downloading a working subset before committing — “the data exists” and “we have the data loading in a notebook” are separated by weeks of format wrangling (Datasets shortcuts most of it).
  3. What is the minimum reportable result? Scope so that a negative or partial answer is still a deliverable: a benchmark, a sensitivity table, a documented failure taxonomy. Projects whose only success mode is “beat the state of the art” are lottery tickets; projects whose worst case is a careful measurement are investments.
  4. Who is the customer? For each problem above the customer is named — a pipeline team, the analysis community, a challenge leaderboard, a planning document. Write the abstract of the final report in week two, addressed to that customer, and let it steer scope.

Worked example: scoping the null-model sensitivity project

To make the scoping concrete, here is Problem 4’s second project taken from idea to plan, with the reasoning shown.

Claim selection. Start from a published, specific, quantitative claim — for example, over-representation of reciprocal connectivity among excitatory neurons, reported in both fly and cortical data. Reciprocity is a two-node motif, so counting is trivial even at full-connectome scale; the entire difficulty, and the entire scientific content, lives in the null model. That asymmetry — cheap statistic, contested null — is exactly what makes it a good undergraduate target. A team tempted instead by, say, five-node motifs would spend the semester on subgraph-isomorphism scaling (a Problem-4a computational project — legitimate, but a different one) and never reach the statistics.

Design. Fix one dataset snapshot and record its version identifier — CAVE materializations make analyses reproducible only if the version is pinned, which is why Provenance and versioning is a week-one reading. Compute observed reciprocity for one well-defined cell population. Then recompute expectation and z-score under at least four nulls of increasing biological fidelity: Erdős–Rényi (edge count preserved); configuration model (in/out degree sequences preserved); a distance-dependent null (connection probability as the empirical function of inter-soma distance, preserving spatial embedding); and a type-aware null (block structure by cell type preserved). Each null answers a different question, and writing down which question before running it is the actual intellectual work: significance under Erdős–Rényi but not under the distance null means “neurons are near each other,” not “the circuit seeks reciprocity.”

Anticipated result shapes. All-nulls-significant strengthens the published claim and the deliverable is a robustness certificate. Significance that disappears under the distance or type null is the more interesting outcome — a documented case that a standard claim is a geometry or composition artifact. Either way the minimum reportable result (question 3 above) is the sensitivity table itself, which is why this project passes the protocol.

Timeline honesty. Weeks 1–3: data access, version pinning, population definition (slower than expected — always). Weeks 4–6: observed statistics and the two easy nulls. Weeks 7–10: the distance-dependent null, which requires fitting the empirical distance-probability curve and is where the project’s real effort concentrates. Weeks 11–13: type-aware null, sensitivity table, write-up. A second semester extends the same machinery to more claims, the male CNS comparison (Problem 3 crossover), or proofreading-stage sensitivity (Problem 1 crossover).

Self-check

Answer before reading the answers.

  1. Why is merge-error detection considered harder-to-automate and higher-stakes than split-error detection?
  2. A team proposes: “we will build a better segmentation model than Google’s flood-filling networks.” Which decision-protocol questions does this fail?
  3. Your motif is significant under a configuration-model null but not under a distance-dependent null. What is the correct interpretation?
  4. Name two ways the proofreading problem (Problem 1) and the structure-function problem (Problem 5) are coupled.

Answers.

  1. Splits are conspicuous — a fragment with no soma, an axon ending mid-volume — and correcting one candidate pair is a well-posed local question, which is why models already do fairly well on it (75–85% in ConnectomeBench’s forced-choice setup). A merge produces a plausible-looking chimera; detecting it requires noticing that a single object’s morphology or connectivity is internally inconsistent, a global judgement. Stakes: an undetected merge fabricates edges between unrelated neurons, corrupting the graph silently; an undetected split mostly loses edges, a more visible and more recoverable error.
  2. Primarily question 1 (skill match against industrial teams with training infrastructure no undergraduate team can replicate) and question 3 (the only success mode is beating a heavily resourced state of the art; a near-miss yields no deliverable). Re-scoped as benchmarking existing models under domain shift (Problem 2), the same interest passes the protocol.
  3. The motif occurs more often than random wiring with matched degrees would produce, but not more often than expected given how close the neurons are to each other: spatial proximity, not motif-specific selection, is the parsimonious explanation. It is a claim about geometry, not circuit computation — and it is a finding, not a failure.
  4. (a) Structure-function models consume the graph that proofreading produces; robustness-versus-proofreading curves determine how much reconstruction error those models’ predictions can tolerate. (b) Lesion/perturbation studies on connectome-constrained models identify which structural features predictions depend on, which tells proofreading campaigns where accuracy actually matters — e.g., the Cell 2025 finding that infrequent strong connections carry disproportionate functional weight implies proofreading effort should prioritize exactly the edges most affected by merge errors.

What this page does not cover

Key references