Most human long noncoding RNAs have no known function, and the reason is not that we cannot perturb them but that we cannot afford to characterize them one at a time. A single arrayed study that profiled 285 lncRNAs required roughly 2,000 separate antisense oligonucleotide transfections and about 1,000 deep-sequencing libraries for molecular phenotyping (Functional annotation of human long noncoding RNAs via molecular phenotyping), while the catalog of unannotated transcription keeps growing — about 72% of mapped transcription start sites could not be assigned to any known gene (Discovery of widespread transcription initiation at microsatellites predictable by sequence-based deep neural network). In parallel, an unrelated community solved a structurally similar problem: DNA data storage systems retrieve individual files from a pool of billions of molecules without reading the whole pool, using short address sequences attached to each molecule (Magnetic DNA random access memory with nanopore readouts and exponentially-scaled combinatorial addressing).
This proposal connects the two. The idea: put a storage-style random-access address directly onto each perturbation molecule — an antisense oligonucleotide (a short synthetic strand that binds a target RNA and recruits its degradation) — so that one reusable master pool of thousands of perturbation agents can be queried non-destructively, and only the selected sub-pool is spent on expensive molecular phenotyping. The technical core is a joint constrained-coding problem that neither community has formulated: the address must satisfy the storage world's rules (balanced base composition, no homopolymer runs, no cross-hybridization between addresses) while simultaneously not destroying the molecule's biological activity (no folding against the target-binding region, no off-target recruitment of RNA-degrading enzymes, no interference with chemical stabilization). The research file shows no existing work bridging molecular random-access architectures and functional annotation screens; the closest works either address inert digital payloads or place constraint-compliant barcodes on functional molecules for security attestation only.
Core Claim
A single coding construction can make perturbation pools randomly addressable without sacrificing biological function, by applying an unequal constraint budget across the molecule: strict algebraic constrained coding on the synthetic address region, and biology-dictated, error-tolerant sequence on the functional region. Specifically, each ASO in the pool carries (i) a gapmer target-binding region whose sequence is fixed by the transcript it silences, (ii) an inert linker, and (iii) an address barcode drawn from an algebraic cross-bifix-free code — meaning no prefix of one address matches a suffix of another, which prevents staggered mis-hybridization during retrieval — extended with the additional constraint that the barcode must not form a hairpin or duplex with the binding region of any molecule in the pool.
The budget for the address overhead follows the existing theory. For $n$ addresses at error-correcting distance $d$, algebraic balanced primer codes achieve redundancy
$$R(n,d) = (d+1)\log_4(n+1) \ \text{nucleotides,} \quad \text{vs. the sphere-packing lower bound } \lfloor (d-1)/2 \rfloor,$$
as established in Efficient and Explicit Balanced Primer Codes, which constructs GC-balanced, mutually uncorrelated, primer-dimer-avoiding address families by a half-flip operation on cyclic shifts — no stochastic search, linear-time encoding. Our contribution is not a new bound but a new constraint class: the barcode interacts with a functional payload that cannot be redesigned to accommodate it. This directly addresses the open problem left in Mutually Uncorrelated Primers for DNA-Based Data Storage, which formalized $\kappa$-weakly mutually uncorrelated codes and the decoupling of GC balance from error correction but explicitly left intramolecular secondary-structure constraints and all wet-lab validation open.
The unequal budget also resolves, in this setting, a genuine tension in the coding literature. Embracing errors is more effective than avoiding them through constrained coding for DNA data storage argues that strict constrained coding wastes rate and that downstream error correction on noisy reads does better; classical pipelines such as the variable-to-variable encoder of Efficient and low-complexity variable-to-variable length coding for DNA storage (which reaches 1.988 bits per nucleotide, 99.9% of its theoretical maximum at homopolymer limit $h=4$) enforce constraints up front. In a functional pool the split is natural rather than ideological: the address region is fully synthetic, so avoidance is cheap and mis-retrieval is silent and wasteful; the functional region is dictated by the target transcript, so avoidance is impossible and errors must be tolerated — a missed knockdown is detectable, but a mis-targeted one wastes a phenotyping slot.
Research Program
Three components, in dependency order. First, a code-construction study: extend the cyclic balanced-primer machinery of Efficient and Explicit Balanced Primer Codes with a pooling constraint — no barcode may stably hybridize or fold with any payload in the pool — and quantify what this added constraint costs in rate and in achievable pool size for pools of $10^2$ to $10^5$ agents. The deterministic resampling loop of Synthesis Tamper-evident Attestation and Molecular Provenance (STAMP), which satisfied homopolymer and GC constraints on 100% of 1,918 real plasmids in fewer than 10 attempts each, supplies a practical fallback encoder; the difference is that STAMP's loop guarantees per-molecule constraints only, with no cross-molecule distance or cross-hybridization guarantee, which is exactly what a retrieval pool needs and what we will add. Interpretability evidence from CircRNA identification and feature interpretability analysis — where intron-flanking hairpins emerged as the dominant determinant of circularization, and where a structure-aware model reached 0.8314 test accuracy versus 0.605 for plain BLAST alignment — tells us hairpin screening must be structural, not sequence-matching alone.
Second, a physical retrieval architecture adapted from Magnetic DNA random access memory with nanopore readouts and exponentially-scaled combinatorial addressing: the master ASO pool is conjugated to magnetic agarose beads via click chemistry so retrieval amplifies and releases copies while the master template stays intact, and combinatorial barcode addresses built from a small set of physical primers expand the address space exponentially. MDRAM proved this works for digital files — 17,600-fold enrichment of a targeted oligonucleotide, decoding from as little as 3x read coverage where bulk frameworks need 30x or more, and stable performance across five sequential random-access cycles on one bead batch — but it never evaluated how sequence constraints on biologically active payloads affect conjugation, amplification, or retrieval fidelity. That is precisely our experimental territory.
Third, the biological interface and the triage workflow that motivates everything: because molecular phenotyping is nearly twice as sensitive as cellular phenotyping — only about 30% of lncRNA knockdowns in Functional annotation of human long noncoding RNAs via molecular phenotyping produced visible cellular defects while nearly 60% shifted the transcriptome — the expensive CAGE readout (Cap Analysis of Gene Expression, a method that counts transcription start sites genome-wide) is the right assay but the wrong thing to spend on every candidate. The pool enables an iterative loop: run cheap whole-pool imaging, retrieve only the sub-pool worth deep phenotyping, transfect that sub-pool, and keep the master intact for the next round. Where RNA guides rather than ASOs are wanted, the chemistry constraints are sharper: Chemically modified guide RNAs enhance CRISPR-Cas13 knockdown in human cells shows the 3' terminus of Cas13 crRNAs is the exonuclease-vulnerable site and that 2'-O-methylation there yields up to 80% knockdown in HEK293FT cells and 60–65% in primary T cells versus roughly 40% unmodified — so any address appended to an RNA guide must avoid or be co-stabilized at that terminus, a placement constraint our code construction must encode.
Baselines
The naive reference point is the arrayed screen as executed by FANTOM6 (Functional annotation of human long noncoding RNAs via molecular phenotyping): 285 lncRNAs, roughly 2,000 separate ASO transfections, about 5,000 imaging samples, and around 1,000 CAGE libraries, yielding molecular phenotypes for 154 targets. This is the strongest competitor on annotation quality — every perturbation is individually tracked — and we will compare directly on cost per transcript with a molecular phenotype, using the same HDF cell system and the same ASO target set so the comparison is a like-for-like accounting of reagents, libraries, and sequencing lane-hours rather than a methodological confound. Beating it means annotating the same 285-target cohort at strictly lower cost per annotated transcript while reproducing the ~60% molecular-phenotype hit rate and a known positive control: the 40–45% wound-closure reduction upon ZNF213-AS1 knockdown reported in that paper.
The strongest existing method on the retrieval question is the MDRAM architecture itself (Magnetic DNA random access memory with nanopore readouts and exponentially-scaled combinatorial addressing), which reports 17,600-fold enrichment of the targeted oligonucleotide, successful decoding at 3x read coverage, and five sequential non-destructive accesses on one bead batch. We will run its combinatorial addressing and bead chemistry unchanged as our retrieval engine, on the same pool sizes and sequencer, so the comparison isolates our one new ingredient — joint constraint-aware barcode placement on functional payloads — rather than hardware. Beating it means matching its enrichment and reusability while additionally preserving knockdown activity of the retrieved agents and holding cross-talk below what naive CBA placement on ASOs produces.
The pooled-screen baseline is the high-content RNAi approach of A high-content RNAi screen reveals multiple roles for long noncoding RNAs in cell division, which screened 2,231 lncRNAs in pooled format and found, for example, that linc00899 depletion extends median mitotic duration from about 40 minutes to over 100 minutes. It demonstrates that pooled screens scale to thousands of targets, but its readout requires processing the whole pool destructively; we compare against it on pool durability across repeated query cycles and on total sequencing spend per round of phenotyping.
Finally, an ablation isolates the coding ingredient: the identical pool architecture and retrieval protocol, but with addresses drawn by random generation plus per-sequence filtering (GC, homopolymer, BLAST-style screening) instead of the algebraic cross-bifix-free construction. The prediction, grounded in Mutually Uncorrelated Primers for DNA-Based Data Storage, which shows unconstrained heuristic selection produces primer-dimer formation and cross-talk during amplification, is that mis-retrieval grows with pool size for the heuristic and stays flat for the code.
Why not just…?
Why not just run arrayed screens in 384-well plates? This is the real threat, because arrayed screens already exist at scale and require no new coding theory. The evidence against it as a scaling path is that FANTOM6's arrayed molecular phenotyping of just 285 transcripts already consumed about 1,000 CAGE libraries, and its authors state that scaling across the tens of thousands of uncharacterized transcripts is logistically and economically constrained (Functional annotation of human long noncoding RNAs via molecular phenotyping). But plate miniaturization keeps improving, so the decisive experiment rather than an argument: the cost-per-annotated-transcript comparison above, including a projected per-transcript cost curve as cohort size grows from 285 to 10,000, where reagent amortization from a reusable master pool should bend the curve and arrayed handling should not. If the crossover point sits beyond 10,000 targets, we will say so plainly.
Why not just deep-sequence the entire pooled library? Because the expensive resource is not reading the pool, it is the phenotyping readout, and full-pool processing also destroys the analyte. Full-pool sequencing introduces read latency and computational overhead when only a subset is needed, and repeated unindexed PCR exhausts and degrades the master template, whereas MDRAM demonstrated targeted retrieval at 3x coverage versus 30x or more for bulk frameworks and a 10x improvement in bases-read-per-bit (Magnetic DNA random access memory with nanopore readouts and exponentially-scaled combinatorial addressing; the same latency argument is made by Tabatabaei Yazdi et al. and Meng et al.). The honest caveat is that guide-identity sequencing is already cheap; our claim rests instead on the triage structure — sequencing the whole pool tells you which perturbations were present, not which deserve a 1,000-library CAGE investment — and on master-pool reuse across rounds, which whole-pool amplification precludes. The durability experiment across five retrieval cycles tests this directly.
Why not just generate random barcodes and filter them heuristically? For pools of a few hundred sequences, this plausibly works, and we should concede it. At scale it fails on two documented axes: stochastic search becomes computationally prohibitive and offers no distance guarantee (Efficient and Explicit Balanced Primer Codes), and heuristic sequence filters miss structural constraints — plain BLAST reached 0.605 accuracy where a structure-aware model reached 0.8314 on the same circRNA discrimination task (CircRNA identification and feature interpretability analysis). Since the barcodes in our design must additionally coexist with functional payloads, the heuristic approach must screen barcode-payload folding across all pairs in the pool, which is quadratic and unsampled by any existing filter. The ablation above settles it empirically: if heuristic barcodes show cross-talk indistinguishable from the algebraic code at $10^4$ pool size, the coding contribution is overstated and we will report that.
Proposed Experiments and Decisive Checks
The computational study comes first and is fully specified: construct joint-constraint address families for the 285 FANTOM6 targets (real ASO sequences, so payload constraints are grounded) and for synthetic cohorts up to $10^5$; report achievable redundancy against the $(d+1)\log_4(n+1)$ baseline and the $\lfloor(d-1)/2\rfloor$ bound, the fraction of barcode-payload pairs with predicted stable structure, and encoding runtime against STAMP's resampling loop. The in vitro study then synthesizes the 285-agent pool, conjugates it to beads, and measures retrieval enrichment (benchmark: MDRAM's 17,600-fold), mis-retrieval rate with algebraic versus heuristic addresses at increasing pool sizes, and master-pool integrity across five cycles. The cellular study has two gates: first, addressed versus unaddressed ASO knockdown of a measurable target in HDFs, where success means retaining activity within roughly 10% of the unbarcoded control and reproducing the ZNF213-AS1 wound-closure phenotype; second, the full triage loop — whole-pool imaging, retrieval of the responsive sub-pool, CAGE on that sub-pool only — scored on cost per transcript with a molecular phenotype against the arrayed baseline's roughly 1,000 libraries for 154 characterized targets. The three decisive checks, in order of importance: barcode placement must not cost biological activity; joint constraint design must suppress cross-talk relative to heuristic filtering as pool size grows; and the master pool must survive repeated retrieval well enough that the second query round costs less than the first. Synthesis cost, the standing bottleneck documented in High-throughput DNA synthesis for data storage, is itself an argument for reuse, since a durable master pool amortizes the dominant expense. As a longer-horizon extension, the rewritable nick-based metadata layer of Rewritable two-dimensional DNA-based data storage with machine learning reconstruction could track batch, passage, and exposure history physically on the pool — but that is deliberately outside this proposal's scope, which ends at demonstrating that addressable, reusable, function-preserving perturbation pools are buildable and cheaper per annotation than the arrayed alternative.
References
- Functional annotation of human long noncoding RNAs via molecular phenotyping
- Discovery of widespread transcription initiation at microsatellites predictable by sequence-based deep neural network
- Magnetic DNA random access memory with nanopore readouts and exponentially-scaled combinatorial addressing
- Efficient and Explicit Balanced Primer Codes
- Mutually Uncorrelated Primers for DNA-Based Data Storage
- Embracing errors is more effective than avoiding them through constrained coding for DNA data storage
- Efficient and low-complexity variable-to-variable length coding for DNA storage
- Synthesis Tamper-evident Attestation and Molecular Provenance (STAMP): Cryptographic Molecular Barcoding for DNA Synthesizers
- CircRNA identification and feature interpretability analysis
- Chemically modified guide RNAs enhance CRISPR-Cas13 knockdown in human cells
- A high-content RNAi screen reveals multiple roles for long noncoding RNAs in cell division
- DNA-Based Storage: Trends and Methods
- High-throughput DNA synthesis for data storage
- Rewritable two-dimensional DNA-based data storage with machine learning reconstruction