The Top 10 Open Questions in Computational Biology, 2026 Edition

AlphaFold2 deposited predicted structures for more than 200 million proteins. Perturb-seq brought genome-scale functional screens into routine use. Long-read sequencing assembled the first truly complete human chromosomes. Yet each of these advances revealed how much of the underlying biology we still cannot explain mechanistically.

Below are ten open problems where the measurement tools exist but the answers do not.

From sequence to function: two problems AlphaFold left open

AlphaFold2 solved protein structure prediction for most of the known protein universe (Jumper et al., 2021, Nature, doi:10.1038/s41586-021-03819-2). Structure prediction is, at this point, a largely solved problem for single-chain proteins. What comes next is not.

1. Predicting protein function from structure

Knowing the three-dimensional shape of a protein at atomic resolution does not tell you which ligand it binds, which reaction it catalyses, or which other proteins it recruits under which cellular conditions. Functional annotation methods still depend heavily on sequence homology to proteins whose function has been characterized experimentally. For the large fraction of deposited AlphaFold structures with no characterized homologues, functional annotation remains inference at best and guesswork at worst.

Thousands of disease-associated missense variants now sit inside deposited structures with no functional interpretation. A method that maps a structure, or a mutation-induced structural perturbation, to a quantitative prediction of binding affinity or catalytic rate would transform both basic biology and drug discovery.

2. Protein complex assembly and transient interactions

AlphaFold-Multimer extended structure prediction to protein complexes, but it is trained primarily on stable, well-characterized interfaces. Transient, condition-dependent, and intrinsically disordered interactions are systematically underrepresented. Many signaling pathways depend on interactions weak enough to be reversible on a timescale of seconds. These interactions resist structural characterization by co-crystallization or cryo-EM, making them invisible to training data. Predicting whether a given protein pair interacts under a specific cellular condition, and with what affinity, remains a genuinely open problem.

Decoding non-coding variation

More than 90% of disease-associated variants identified by genome-wide association studies fall in non-coding regions of the genome. Two distinct challenges have emerged from this fact.

3. Causal variant identification from GWAS fine-mapping

Statistical fine-mapping methods (SuSiE, FINEMAP, PolyFun) narrow a GWAS locus from hundreds of candidate variants to a shorter credible set, but they frequently disagree on the same data. Colocalization analysis that overlays expression quantitative trait loci (eQTLs) with GWAS hits has improved, but it depends on tissue-specificity and cell-type composition assumptions that are violated in mixed-tissue samples. The result is that for most GWAS loci, we cannot name a single causal variant with confidence, let alone explain its mechanism.

An approach integrating chromatin accessibility, three-dimensional genome conformation, and allele-specific expression simultaneously, without requiring enormous matched cohorts for every tissue and disease, would close this gap across the board.

4. Non-coding RNA function at scale

The human genome encodes tens of thousands of long non-coding RNAs (lncRNAs). Experimental characterization, primarily via CRISPR interference tiling screens, has shown that a subset of these have phenotypic effects on proliferation, differentiation, and gene regulation. For fewer than 5% of catalogued lncRNAs is the molecular mechanism understood.

The problem is that phenotypic effect and molecular mechanism are not the same thing. Knowing that perturbing a lncRNA changes gene expression in nearby loci does not explain how. For protein-coding genes, sequence homology to characterized proteins provides a functional scaffold. Most lncRNAs have no conserved homologues, leaving their mechanism completely opaque.

The single-cell layer: two unresolved problems

Single-cell sequencing has transformed our view of cell-type diversity. Two foundational analysis problems remain open.

5. Cell fate inference and trajectory disagreement

Methods that reconstruct developmental trajectories from single-cell RNA sequencing data include RNA velocity (La Manno et al., 2018, Nature), pseudotime ordering, and fate probability tools such as CellRank (Lange et al., 2022, Nature Methods). Applied to the same dataset, these methods frequently disagree on branching topology and fate assignments. Two progenitor populations can be assigned to opposite fate outcomes depending on which algorithm is used.

Ground truth is rare. Prospective lineage tracing at single-cell resolution remains expensive and technically limited to organisms and tissues amenable to clonal labeling. A method that integrates transcriptional kinetics, chromatin state, and spatial context into a model that can be prospectively validated would resolve this.

6. Multi-omics integration across modalities

Technologies such as CITE-seq, 10x Multiome, and Slide-tags now routinely pair RNA with chromatin accessibility, protein abundance, or spatial location in the same cells. Integrating these modalities into a coherent biological model, especially across samples or batches, remains unsolved. Benchmarking studies consistently find that no single method (Seurat WNN, MOFA+, Mowgli, and others) outperforms others across datasets. Each method makes assumptions about the latent structure relating modalities that may or may not hold in a given experiment. Reliable uncertainty estimates for cross-modal inference are rarely provided.

Networks and causality: two problems at the heart of systems biology

7. Causal regulatory network reconstruction from perturbation data

Large Perturb-seq datasets now document transcriptional responses to thousands of single-gene CRISPR perturbations in the same cells. Reconstructing a causal gene regulatory network from these data is not a solved problem. Causal discovery algorithms (PC, GES, NOTEARS) make assumptions of acyclicity, Gaussian noise, or causal sufficiency that are routinely violated in gene regulation. Networks inferred by different methods on the same perturbation data show low pairwise overlap. The core challenge is distinguishing direct regulatory effects from indirect ones in a system with ubiquitous feedback, latent regulators (transcription factors with low mRNA but high protein activity), and combinatorial logic.

8. Microbiome composition to host phenotype: the causal chain

Shotgun metagenomics has documented associations between gut microbiome composition and dozens of conditions, from inflammatory bowel disease to type 2 diabetes to depression. For most of these associations, the causal chain from microbe to host phenotype is not established. Diet, medication use, host genetics, and early-life exposures are each correlated with both microbiome composition and disease risk, making observational studies difficult to interpret.

Germ-free mouse models and fecal microbiota transplantation experiments establish causality for specific microbe-phenotype pairs, but they do not scale across the full diversity of human microbiome variation. Mendelian randomization using heritable microbiome taxa as genetic instruments has shown promise in recent work but remains limited by the small number of taxa with identifiable heritable genetic determinants.

The design and completeness frontier

9. Protein design for novel enzymatic function

De novo protein design has advanced rapidly. Diffusion-based backbone generators (Watson et al., 2023, Nature) produce foldable structures for arbitrary geometric targets. Sequence design tools such as ProteinMPNN (Dauparas et al., 2022, Science) reliably identify sequences that fold to those backbones. Designing a protein with a specific novel enzymatic activity, meaning a reaction chemistry not found in natural enzymes, remains hard. Most reported designed enzymes achieve catalytic rates two to four orders of magnitude below natural enzymes for the same reaction class. The problem is not structure, it is mechanism: placing a catalytic residue geometry and electrostatic environment in a designed scaffold that actually lowers the activation energy of the target transition state.

10. Pan-genome completeness and structural variation

The Human Pangenome Reference Consortium released a 47-genome draft pangenome reference in 2023 (Liao et al., Nature), resolving many structural variants that are invisible to the GRCh38 single-individual reference. The pangenome represents a small and still biased sample of human genetic diversity. Structural variants, including inversions, tandem duplications, mobile element insertions, and complex rearrangements, remain difficult to genotype at scale even when they are present in the pangenome graph. Accurate short-read genotyping of structural variants from graph references is an active area without a satisfying solution. A population-representative pangenome with complete structural variant catalogs, paired with short-read genotyping methods accurate enough for biobank-scale cohorts, would unlock a layer of variation that has been inaccessible to most genetic studies.

What this means for your research

Before committing to a research direction, search your specific topic, such as "protein function prediction" or "GWAS causal variant", in the research gap finder. The search is free and runs over the 4.5M-paper local library; it returns the papers it found, in one list you can filter, sort and export. One topic unlock, paid in credits (see pricing), fills the literature comparison table and writes the research gaps from its rows, with a research proposal included.

You can Pre-Check a manuscript from its title and abstract for an acceptance estimate before you submit to a journal. For computational biology specifically, the full AI Review runs a methodology agent calibrated on eLife, Nature Communications and PCI reviews from our 69,000-review corpus.

If your work touches any of the ten areas above, the prior publication detection step of the AI Review checks seven external sources (CrossRef, PubMed, arXiv, bioRxiv, medRxiv, Europe PMC and Unpaywall) plus our local library for an earlier version of your own manuscript.

Frequently asked questions

What distinguishes an "open question" from an incremental research gap?

An open question has two properties: the community agrees on what a satisfying answer would look like, and multiple independent groups have attempted to answer it without convergence. An incremental gap is the absence of a study in a space where the method is known and the result is predictable. Open questions are different in kind.

How does the research gap finder surface open problems in computational biology?

Our research gap finder works in two steps. A free search of any subfield, such as "single-cell trajectory inference" or "microbiome causality", runs over the 4.5M-paper local library and returns the papers it found. One topic unlock, paid in credits, then fills the literature comparison table and writes the research gaps from its rows, with a research proposal included.

Are all ten problems suitable for a PhD project?

No. Problems 3 (GWAS causal variants), 5 (trajectory inference disagreement), and 7 (regulatory network benchmarking) have subproblems accessible to a well-scoped PhD, particularly method comparison on public datasets, benchmark construction, or extending existing causal inference frameworks. Problems 1 (function prediction) and 9 (enzyme design) require wet-lab experimental validation and are typically beyond a purely computational project without experimental collaborators.

Can I use Science AI Journal to review a paper on any of these topics?

Yes. The review covers all ten disciplines we support, including life sciences, computer science, mathematics, and environmental science. The methodology agent's rubric is specifically calibrated on eLife, Nature Communications, PCI Ecology, PCI Evolutionary Biology, and PCI Mathematical and Computational Biology standards. You can run the full AI Review on the PDF, or a Pre-Check on a title and abstract; both are priced in credits (see pricing).

AI Review reports are written by AI agents. They are pre-submission feedback, not an editorial decision.