2.1. A Foundation Model Meets a Lab-in-the-Loop

Foundation models are deep neural networks pretrained on expansive data under self-supervised objectives, acquiring representations that transfer to a broad range of downstream tasks [Reference LeCun, Bengio and Hinton32,Reference Moor, Banerjee and Abad33,Reference Bommasani, Hudson and Adeli34], and they have begun to reshape the modeling of nucleic acids and small molecules. Genomic language models such as DNABERT learn contextual representations of nucleotide sequence that transfer to regulatory-element and genetic-variant identification [Reference Ji, Zhou, Liu and Davuluri35,Reference Dalla-Torre, Gonzalez and Mendoza-Revilla36], whereas RNA-focused models such as 3UTRBERT capture the sequence determinants of mRNA subcellular localization and RNA-protein interactions within untranslated regions [Reference Yang, Li and Pang37,Reference Chu, Yu and Li38]; at the coding-sequence level, CodonBERT is pretrained on more than 10 million codon-tokenized mRNA coding sequences from diverse organisms and can be fine-tuned to predict properties such as protein expression, mRNA stability, and degradation, thereby supporting the evaluation and selection of coding sequences for mRNA vaccine and therapeutic design [Reference Li, Moayedpour and Li39]; in parallel, molecular foundation models such as Uni-Mol learn three-dimensional representations of chemical structure that support property prediction [Reference Zhou40]. Collectively, these advances support the feasibility of pretrained prior transfer within each modality. No existing model, however, jointly encodes the RNA payload together with the chemistry underlying its delivery. Bridging these hitherto separate priors is a central premise of the platform we propose.

A pretrained prior alone, however, is insufficient. The properties most critical to an RNA therapeutic, notably its in vivo efficacy and delivery performance, are difficult to learn from static datasets alone because they depend on biological context that no fixed corpus fully captures. This limitation motivates the lab-in-the-loop paradigm. A lab-in-the-loop architecture shifts the model from passive fitting to active exploration: the model proposes candidates, automated experiments evaluate them, and standardized results return to recalibrate the model [Reference Burger, Maffettone and Gusev24–Reference Xu, Cui and Pang26,Reference Yang, Wu and Arnold30] (Fig. 1E). The two components are complementary. Without the loop, a foundation model extrapolates unreliably to out-of-distribution designs, precisely the novel regions of greatest therapeutic interest; without a foundation model, the loop must explore afresh at every cycle, expending scarce experimental throughput on a sample-inefficient search. The foundation model thus supplies a transferable prior, and the loop supplies the continual, domain-specific supervision that keeps it calibrated.

2.2. Expected Characteristics and Architecture

Realizing this lab-in-the-loop vision places specific demands on the underlying model architecture. The platform should integrate several core capabilities within a shared representation: the modeling of sequence, the joint encoding of sequence and structure, an awareness of chemical modification, and a description of the delivery vehicle. Concretely, we envision modality-specific encoders that project their respective inputs into a shared latent space for further fusion of features. Transformer architectures offer a natural computational backbone for this task, because their attention mechanisms relate distant tokens directly and are therefore well suited to RNA, in which long-range base pairing, interactions among motifs, and the cross-modal coupling between a sequence and its delivery molecule are all consequential.

The platform should be generative as well as predictive: the models that score a candidate must also be able to propose and sample sequences, modification patterns, and delivery formulations conditioned on a target property profile. This generative capacity renders the loop proactive rather than merely evaluative, and it mirrors the trajectory of protein design, in which models progressed within a few years from scoring natural sequences to generating functional ones de novo [Reference Madani, Krause and Greene28,Reference Watson, Juergens and Bennett29,Reference Ingraham, Baranov and Costello41]. For RNA, the generative target is broader still, as a useful proposal comprises a sequence together with its chemistry and its carrier rather than a sequence alone. Such proposals require the model to respect the hard constraints of synthesizability and manufacturability throughout exploration.

2.3. The Lab-in-the-Loop Workflow

The platform would supplement hypothesis-driven, small-scale iteration with data-driven, model-guided exploration. A single turn of the loop proceeds through five linked stages. The model first proposes candidate RNAs, selected not at random but by expected information gain or predictive uncertainty; automated synthesis then assembles the corresponding library, and high-throughput in vitro and in vivo assays measure their function. The resulting data are standardized, quality-controlled, and normalized to align with the pretraining corpus, after which the model is updated, and the cycle repeats [Reference Burger, Maffettone and Gusev24–Reference Xu, Cui and Pang26,Reference Yang, Wu and Arnold30,Reference Biswas, Khimulya, Alley, Esvelt and Church31] (Fig. 2A).

Figure 2.

(A) A single turn of the lab-in-the-loop cycle. The foundation model is first pretrained on multimodal input (sequences, structures, modifications, and delivery molecules), yielding a prior that supports property prediction over proposed candidates. Candidates selected by the model using predicted performance and uncertainty are advanced to wet-lab experiments, whose functional outcomes are standardized and returned to recalibrate the model. The updated model proposes the next set of candidates, and the cycle repeats, with the foundation model serving as the persistent memory that links successive rounds. (B) Opportunities enabled by the platform. Delivery optimization comprises two distinct tasks: ionizable lipid discovery, which searches the combinatorial chemical space of the ionizable lipid, and formulation optimization across component ratios and process parameters. In parallel, the mRNA construct is optimized across its 5′ UTR, coding sequence (CDS), and 3′ UTR, which are conventionally tuned separately but are jointly optimized here. Coupling delivery optimization with sequence co-design yields joint sequence-carrier design, in which payload and delivery vehicle are optimized within a shared representation rather than sequentially.

Figure 2. Long description

Panel A shows a circular flowchart labeled Lab in the loop. Starting at the left, multimodal inputs including mRNA sequences, structures, and delivery molecules enter a Pre-training phase represented by a neural network icon. This leads to Property prediction, which outputs proposed candidates. These candidates move to a Wet lab experiment. The Experiment outcome is visualized as a line graph of Fluorescence with multiple peaks. This data flows into Re-calibration, which feeds back into the initial input stage to repeat the cycle.

Panel B illustrates the platform components. On the left, a dashed box labeled Ionizable lipid discovery contains a lipid molecule icon, while a sub-box labeled Formulation optimization shows various chemical components. In the center, three horizontal capsules represent the mRNA construct: 5 prime UTR in green, CDS in blue, and 3 prime UTR in orange. Arrows from these three segments point down to a central Co-design node. A final horizontal line connects the lipid discovery and mRNA co-design sections to a label for joint sequence-carrier design, which points to a cluster of lipid nanoparticles on the far right.

The pivotal requirement is that the model emit probabilistic outputs rather than point predictions. Only a calibrated estimate of uncertainty enables the loop to address its central question, which candidates should be measured next, and thereby allocate limited experimental capacity where it most efficiently reduces uncertainty. This requirement binds the model to the loop because the model’s epistemic uncertainty constitutes the loop’s experimental-design objective. We anticipate that this cycle of proposal, measurement, and model update could become as central to RNA therapeutics as the design-build-test-learn cycle has become for synthetic biology.

3.1. Optimization of Sequence

Among the capabilities the platform enables, sequence optimization is the most immediate, and it can be pursued across several sequence scales. The model proposes mRNA candidates by jointly optimizing the 5′ and 3′ untranslated regions and the coding sequence, either as individual components or within the full-length construct (Fig. 2B), while the self-driving laboratory synthesizes and characterizes them automatically, returning quantitative measurements that feed directly into the next design round [Reference Leppek, Byeon and Kladwang13,Reference Chu, Yu and Li38,Reference Sample, Wang and Reid42].

The principal advantage of this approach lies in joint optimization. Untranslated regions, codon choice, and secondary structure are conventionally tuned by separate heuristics, although they trade off against one another in practice [Reference Wan, Qu and Zhang11,Reference Presnyak, Alhusaini and Chen12,Reference Zhang, Zhang and Lin43]; by learning sequence-to-function relationships directly from assay data, the model navigates these coupled axes simultaneously rather than sequentially [Reference Sample, Wang and Reid42,Reference Angenent-Mari, Garruss, Soenksen, Church and Collins44,Reference Valeri, Collins and Ramesh45]. Experimentally, massively parallel reporter assays evaluate the proposed variants in parallel and return functional measurements within days, closing the design-build-test-learn loop [Reference Patwardhan, Hiatt and Witten46].

However, sequence optimization is incomplete if treated as a search over nucleotide identity alone. A therapeutic RNA must simultaneously fold into a structure compatible with its function, resist degradation during storage and in circulation, and avoid triggering innate immune sensing; these three properties are governed by overlapping design variables rather than by independent ones. They also compete: for instance, pairing a nucleotide protects it from hydrolysis but may impede ribosomal scanning [Reference Wayment-Steele, Kim and Choe47]; the chemical modifications introduced to suppress innate sensing simultaneously alter base-pairing thermodynamics and translational fidelity [Reference Karikó, Buckstein, Ni and Weissman48], as the +1 ribosomal frameshifting recently reported for N¹-methylpseudouridine illustrates [Reference Mulroney, Pöyry and Yam-Puc49]. Structure, stability, and immunogenicity must therefore be optimized jointly, and each enters the loop twice: as information the model conditions on, and as a measurement the loop returns.

As model input, all three properties are supplied as conditioning information, but they differ in how reliably that information can be obtained. Structure is the case where a prediction is readily available and most easily over-trusted, and we caution against conditioning on a single predicted minimum-free-energy structure. A long mRNA does not adopt one conformation but populates an ensemble, of which the minimum-free-energy structure may represent only a small fraction. More seriously, the nearest-neighbor parameters underlying standard folding algorithms were calibrated on unmodified nucleotides and do not describe the modified bases used therapeutically [Reference Turner and Mathews50], so the predicted structure is not merely incomplete but systematically miscalibrated. We therefore favor conditioning on base-pairing probability matrices, which retain the ensemble [Reference Lorenz, Bernhart and Höner zu Siederdissen51], and on experimentally measured reactivity profiles from high-throughput chemical probing [Reference Siegfried, Busan, Rice, Nelson and Weeks52], supplied directly as per-nucleotide features rather than converted into pseudo-energies for a thermodynamic folding step, which would reintroduce the same parameter dependence. Stability and immunogenicity admit no comparable predictor, and their determinants enter as explicit features instead: for stability, the modification pattern and the ensemble-averaged fraction of unpaired nucleotides, which sets the rate of hydrolysis [Reference Wayment-Steele, Kim and Choe47]; for immunogenicity, the identity and position of modified nucleotides together with process metadata, which the sequence alone does not determine. Across all three, the principle is the same: predicted or derived quantities should be supplied as soft constraints carrying their own uncertainty, not as ground truth, consistent with the calibration requirements set out above.

As loop measurements, each property has a plate-scale readout: accelerated-degradation and integrity assays for stability, chemical probing for structure, and interferon-stimulated-gene reporters and cytokine panels for immunogenicity [Reference Leppek, Byeon and Kladwang13,Reference Nelson, Sorensen and Mintri53]. These belong to the densely queried layer of the assay hierarchy described below, with human reactogenicity remaining an escalated endpoint that in vitro immunogenicity predicts only imperfectly. Because the three properties trade off, they should enter the acquisition function as competing objectives on the Pareto front (Fig. 3C) rather than collapsing into a single expression score.

Figure 3.

(A) Multimodal, multilevel representation. Four input modalities, biological sequence, chemical modification, structure and delivery vehicle, are tokenized at multiple levels of granularity: sequence is resolved from single nucleotides through k-mers and motifs to the entire sequence; structure is represented at both secondary and three-dimensional levels; and the delivery vehicle is encoded as SMILES strings and chemical graphs. In the proposed architecture, these representations are integrated early through multilevel cross-attention, yielding a shared hierarchical representation in which a sequence and its carrier are reasoned over jointly. (B) Active-learning loop. A model trained on the current training pool estimates its predictive uncertainty over an unlabeled pool of candidates; the most informative candidates are selected for experimental annotation, and the resulting labels are added back to the training pool. Iterating this cycle concentrates experimental effort where it most reduces the model’s uncertainty. (C) Multi-objective selection. Since an RNA drug must satisfy multiple competing properties simultaneously, candidate selection is inherently multi-objective. Rather than maximizing a single scalar, the loop traces the Pareto front (red), the set of non-dominated candidates for which no property can be improved without degrading another, here shown for two representative properties.

Figure 3. Long description

Panel A, Multimodal multilevel representation. At the top, four Multimodal inputs are listed: Sequence, Chemical modification, Structure, and Delivery vehicle. These flow down into Multilevel tokenization boxes. Sequence breaks down into entire sequence, k-mer/motif, and single nucleotide. Structure breaks down into 3 D structure and secondary structure. Delivery vehicle breaks down into SMILES strings, chemical graphs, and formulation features. All four streams converge via arrows into a central box labeled Multilevel cross-attention, which leads to a final Hierarchical representation at the bottom.

Panel B, Active-learning loop. A clockwise cycle starts at a Training pool box. An arrow labeled Model training points to a neural network diagram. From the network, an arrow labeled Estimate uncertainty points to an Unlabelled pool. From there, an arrow labeled Selected candidates points to Annotated labels, and a final arrow labeled Add to training pool completes the cycle.

Panel C, Multi-objective selection. A scatter plot with Property B on the x-axis and Property A on the y-axis. Grey dots represent various candidates. A series of red dots connected by a dashed line forms a descending curve from the top-left to the bottom-right, labeled as the Pareto front. This front represents the optimal trade-off between the two properties.

3.2. Co-Design of the Delivery Vehicle and Sequence

4.1. Data: Pretraining Data and Loop Data

The central modeling obstacle is the scarcity and dispersion of paired measurements linking RNA sequence and formulation to function. Substantial static resources exist, including encyclopedias of genomic elements and structure repositories [79,80,Reference Kalvari, Nawrocki and Ontiveros-Palacios81], yet they seldom pair a sequence with the therapeutic readouts of interest and remain fragmented across RNA classes and species. Integration of data across RNA modalities and across organisms will therefore be essential if the model is to acquire the breadth from which transferable priors emerge [Reference Nguyen, Poli and Durrant23,Reference Dalla-Torre, Gonzalez and Mendoza-Revilla36].

A complementary obstacle arises on the experimental side. Data generated over successive rounds of experimentation must be standardized, quality-controlled, and normalized before they can recalibrate the model, and the associated pipelines must harmonize newly acquired measurements with the pretraining corpus through shared metadata, controls, and normalization [Reference Xu, Cui and Pang26]. This harmonization, however, is not a neutral preprocessing step. A central difficulty is distinguishing the correction of technical batch effects from genuine distribution shift. Successive rounds of active learning deliberately query underexplored regions of chemical or sequence space, so each new batch is expected to differ from the pretraining corpus not only technically but biologically. Batch-correction methods that assume a shared underlying distribution across batches (global scaling, quantile normalization, or purely data-driven surrogate-variable approaches) risk absorbing this genuine novelty as if it were technical variation, over-normalizing away precisely the signal the loop is designed to discover. This risk is greatest exactly for the novel, sparsely sampled points that active learning tends to select, since these are the easiest to distort during normalization. Mitigating this requires anchoring technical variation to explicit references rather than inferring it from the data alone. Bridging or spike-in controls carried across every round allow batch effects to be estimated without conflating them with biological shift, and hierarchical or mixed-effects formulations can model technical and biological variance jointly rather than subtracting the former. Framed in terms of uncertainty, the goal is to attribute technical variability to aleatoric noise while preserving out-of-distribution novelty as epistemic uncertainty, so that the acquisition strategy pursues genuine discovery rather than chasing technical noise. Where paired data are scarce, synthetic data generated by mechanistic or generative models can serve as a complementary signal, provided that their biases are characterized and that empirical measurements remain the final arbiter.

A subtle but consequential manifestation of this problem is the batch effect between newly generated data and the historical corpus. Measurements produced by different laboratories, instruments, and protocols carry systematic offsets that, if left uncorrected, may be misread by the model as biological signal. The platform should therefore treat experimental metadata, including reagent lot, assay date, operator, and instrument, as first-class inputs and incorporate spike-in controls and shared reference standards in every batch, so that the system recalibrates against a fixed anchor rather than drifting with each round of measurement.

4.2. Computational Components for the Foundation Model

The foundation model is shaped above all by how it represents scale, modality, and prior knowledge. Because RNA is organized across many scales, a tokenization that ascends from nucleotides through k-mers and motifs to structural elements, chemical modifications and ultimately the full-length construct can capture this hierarchy [Reference Nguyen, Poli and Durrant23,Reference Dalla-Torre, Gonzalez and Mendoza-Revilla36]. Integration across modalities is equally essential: biological sequence, chemical modification, delivery molecules (representable as chemical graphs or string encodings), and structures must occupy a shared representation [Reference Zhou40] (Fig. 3A). The strategy of modality fusion, however, remains an open design question. Early fusion, integrating modalities at an early representation stage so that the model reasons about a sequence and its carrier in concert rather than in series, is attractive when the modalities are mechanistically coupled, as sequence and delivery vehicle plausibly are, since their interactions can be modelled directly instead of at the output. However, early fusion increases the risk that a data-rich modality dominates a data-scarce one, and it can be brittle when one modality is missing at inference. Late fusion, which combines modalities only at or near the output stage, is more robust to missing modalities and to imbalanced data, while hybrid schemes that share representations only at selected layers offer intermediate trade-offs. We propose early fusion as the primary hypothesis (Fig. 3A) but benchmark it directly against late and hybrid alternatives, using predictive performance on downstream therapeutic-property tasks under matched parameter scale as the primary comparison criterion, with robustness to missing modalities and to modality-scale imbalance as secondary criteria. Finally, mechanistic prior knowledge can constrain learning: thermodynamic base-pairing rules, curated RNA-motif databases and the structure–activity relationships of delivery vehicles can be introduced as inductive biases that ground an otherwise purely data-driven model in mechanism [Reference Wan, Qu and Zhang11,Reference Sato, Akiyama and Sakakibara82].

4.3. Computational Components for the Loop

The loop demands its own computational machinery, which is what distinguishes the platform from a conventional foundation model. At its core lie active learning, Bayesian experimental design and uncertainty-guided selection (Fig. 3B), which determine which candidates to measure, while batch acquisition proposes a set of candidates at once, matching the throughput of parallel assays rather than evaluating a single candidate per round [Reference Yang, Wu and Arnold30,Reference Shields, Stevens and Li62]. Because an RNA drug must satisfy efficacy, stability, safety, and manufacturability simultaneously, this selection is inherently multi-objective, requiring the loop to trace a Pareto front rather than to maximize a single scalar objective (Fig. 3C). These procedures rest on uncertainty quantification, which converts the model’s epistemic uncertainty into a decision rule for experimentation and thereby binds the model to the loop once more. They also depend on automation infrastructure, namely, self-driving synthesis and assay platforms, that supplies the throughput without which the loop stalls [Reference Burger, Maffettone and Gusev24,Reference Boiko, MacKnight, Kline and Gomes25,Reference Xu, Cui and Pang26]. A further principle is to match the granularity of acquisition to the cost of each assay: inexpensive, rapid readouts such as reporter expression are queried densely to map coarse structure, whereas costly, slow readouts such as in vivo biodistribution are reserved for the few candidates that the cheaper layers cannot resolve. Making this escalation decision well requires two things. First, the uncertainty reported at each layer must be trustworthy: when the model predicts a reporter-expression value with, say, 90% confidence, that prediction should be correct 90% of the time, which can be verified on held-out candidates and corrected if it is not. Second, a candidate should be escalated to an expensive in vivo assay not simply because its cheap readout is uncertain, but because measuring it is expected to resolve real uncertainty about the in vivo outcome. This requires knowing how well the cheap readout actually predicts the expensive one: a reporter-expression value that is confidently predicted but only loosely related to biodistribution should not trigger a costly in vivo experiment. The loop therefore escalates a candidate when the expected gain in knowledge about the high-fidelity endpoint, relative to the cost of the assay, is greatest. The loop thereby becomes a hierarchy of nested cycles rather than a single feedback line, with information propagating upward from high-throughput surrogates to low-throughput, high-fidelity endpoints, and with the foundation model serving as the shared memory that binds the layers together.

4.4. Agentic Reasoning as the Proposal Layer

The loop’s proposal step need not be a purely numerical optimizer. A parallel line of work reframes candidate generation as an agentic process, in which large language models are organized into cooperating specialist agents that reason over the scientific literature, formulate hypotheses, and critique one another before committing to an experiment. Co-Scientist exemplifies this paradigm: a coalition of generation, reflection, ranking, evolution, and meta-review agents advances hypotheses through tournament-style scientific debate, with an Elo-based ranking and additional test-time computation devoted to the most promising proposals. The system has produced experimentally supported hypotheses in drug repurposing, novel-target discovery, and the mechanism of antimicrobial resistance, in one case recapitulating an unpublished experimental finding [Reference Gottweis, Weng and Daryin27]. Such systems complement the foundation-model-plus-loop architecture we advocate. Their strength is breadth and interpretability, the ability to traverse a large hypothesis space and to attach an explicit, literature-grounded rationale to each proposal. Their weakness is that language-model reasoning alone is not calibrated against wet-laboratory outcomes and is prone to fluent but unfounded claims. The foundation model supplies precisely what the agents lack, a quantitative, uncertainty-aware estimate of how a candidate is likely to behave, and the self-driving laboratory supplies the empirical verdict that neither the agents nor the model can substitute.

Adapted to RNA therapeutics, an agentic layer could sit at the front of each loop iteration and take on the tasks least amenable to a numerical objective. It could translate a therapeutic goal into concrete design hypotheses, for instance proposing a chemical-modification pattern to evade innate immune sensing, a 5′ UTR motif to raise translation, or an ionizable-lipid scaffold for a target tissue, together with the mechanistic argument for each. It could reconcile these proposals against prior loop feedback and the published literature, flagging where a suggestion contradicts an earlier measurement, and it could nominate the experiments expected to be most discriminating, which the foundation model then scores and ranks by calibrated uncertainty before the self-driving laboratory executes them. Realizing this safely requires the same guardrails we develop below: every agent-generated hypothesis must remain experimentally falsifiable, be traceable to the evidence that motivated it, and be gated by calibrated uncertainty rather than by the rhetorical confidence of the language model that proposed it. Framed this way, agentic reasoning and the calibrated foundation model occupy distinct and complementary roles, one supplying interpretable hypothesis breadth and the other the quantitative ranking on which the loop’s acquisition decisions ultimately depend.

5.1. Data and Automation Infrastructure

The computational loop can advance only as quickly as the automated laboratory can supply reliable data. Self-driving laboratories meet this demand but remain costly and unevenly accessible [Reference Burger, Maffettone and Gusev24,Reference Boiko, MacKnight, Kline and Gomes25,Reference Xu, Cui and Pang26]; paired sequence-to-function data are still scarce; and the computational expense of pretraining large models constrains who can participate. Low-resource training and inference techniques, together with shared infrastructure, will therefore be essential to broaden access, particularly for academic laboratories operating under constrained computational budgets.

5.2. Evaluation Methodology

The field of RNA therapeutics lacks standardized benchmarks, directly limiting method comparison and model selection. Feedback drawn from different layers of the loop carries different degrees of reliability, since an in vitro reporter readout, an in vivo biodistribution measurement, and a clinical endpoint are not interchangeable; the translation gaps between them oblige the loop to weight its feedback accordingly [Reference Sahay, Querbes and Alabi75,Reference Bhatia, Chen, Dobrovolskaia and Lammers83]. Benchmarks for this setting should therefore reward what the loop actually optimizes, not agreement with fixed labels. Four metrics are well suited. Prospective hit rate, the fraction of proposed candidates that succeed when tested in the next round, is the quantity the loop most directly aims to improve, and cannot be inflated by fitting existing data. Out-of-distribution generalization measures performance on chemistries or sequences absent from training, which matters because active learning deliberately explores such regions. Calibration error checks whether the model’s stated confidence matches its actual accuracy, a prerequisite for the uncertainty-guided selection described above. Pareto hypervolume measures how well the proposed candidates cover the trade-offs among efficacy, stability, safety, and manufacturability, rather than collapsing them into one score. A further caution concerns evaluation metrics anchored to established design heuristics, though it requires a distinction we did not draw sharply enough. Some accumulated design knowledge rests on well-characterized biophysics: the apparent pKa window within which an LNP remains near-neutral in circulation yet protonates in the acidifying endosome [Reference Jayaraman, Ansell and Mui84,Reference Hassett, Benenato and Jacquinet85], and the molecular geometry favoring non-bilayer phases that destabilize the endosomal membrane [Reference Hafez, Maurer and Cullis86,Reference Cullis and Hope87], are mechanistic constraints on ionization, packing, and self-assembly, supported by structure–activity data across many chemistries [Reference Jayaraman, Ansell and Mui84,Reference Whitehead, Dorkin and Vegas88]. They belong in the platform as inductive biases and should be evaluated as critical quality attributes in their own right. Our caution is narrower and concerns the more specific rules that accompany them, such as a preferred tail length or head-group class [Reference Whitehead, Dorkin and Vegas88], which largely describe the regions of chemical space that existing libraries happened to cover [Reference Jayaraman, Ansell and Mui84,Reference Whitehead, Dorkin and Vegas88]. Rewarding agreement with these penalizes a model precisely when it proposes a design outside that region [Reference Hajj, Ball and Deluty89,Reference Brookes, Park and Listgarten90]. The two cases are empirically separable: a mechanistic constraint should continue to govern outcomes as the loop moves into unexplored space, whereas a sampling artifact should not. We therefore advocate encoding mechanistic constraints as priors, treating descriptive regularities as testable hypotheses rather than scoring rules, and grounding evaluation in prospective outcomes.

5.3. Interpretability and Hallucination Risk

In a therapeutic context, the platform’s interpretability and trustworthiness are essential. Any candidate the platform proposes must be experimentally falsifiable, and the model must be able to acknowledge the limits of its knowledge; in a patient-safety context, this explicit communication of uncertainty constitutes a hard constraint [Reference Moor, Banerjee and Abad33,Reference Truhn, Reis-Filho and Kather91,Reference Thirunavukarasu, Ting and Elangovan92]. Large language models illustrate both the promise and the peril, as adapted models can match domain experts on certain tasks yet remain prone to fluent and confident error [Reference Van Veen93]. We accordingly require that the platform’s outputs be traceable to training data and loop feedback, remain consistent with the specified context, and be accompanied by calibrated uncertainty. To be actionable rather than aspirational, these principles must be instantiated as auditable safeguards. Traceability becomes provenance reporting: each proposed candidate carries a record of the training data and prior measurements that most influenced it, so the proposal can be inspected rather than trusted. Calibrated uncertainty becomes an explicit threshold, above which a prediction is not advanced automatically but flagged for human review or additional measurement. This defines an abstention criterion: the platform declines to make a confident recommendation when its uncertainty is high or a candidate falls outside its training space, an explicit declaration of uncertainty that is safer than a confident error. Finally, escalation between assay layers is governed by predefined decision criteria, pre-registered criteria on efficacy, safety, and uncertainty that a candidate must meet before advancing to a costlier assay, so that each escalation can be audited against the criterion that authorized it.

5.4. Feedback Loop Bias

A risk that is especially acute in a closed loop is the distribution shift that the loop induces in itself. Should the model propose only candidates in regions where it is already confident, the loop will reinforce that confidence, starve the model of informative failures, and systematically overlook novel designs; this is the dynamic analogue of an evaluation metric that penalizes novelty. Mitigating this danger calls for deliberate exploration, uncertainty-seeking acquisition and the periodic inclusion of diverse or adversarial candidates, so that the loop broadens the model’s competence rather than narrowing it.

5.5. Open Science and Ethical Considerations

Where privacy, security and intellectual-property constraints permit, models and curated data should be openly accessible, accompanied by clear statements of their capabilities, limitations and intended use [Reference Bommasani, Kapoor and Klyman94]. The translation of AI-designed RNA into the clinic raises questions of intellectual property, regulatory oversight and the validation of predictions on clinical cohorts before use in patient care; the recent progress of computationally designed drugs into clinical testing renders these questions concrete rather than hypothetical [Reference Ren, Aliper and Chen95]. Patient data privacy and representativeness across populations must likewise be safeguarded, lest the platform encode or amplify existing inequities.

5.6. A Future of Collective Innovation

The development of a lab-in-the-loop AI platform will require a collaborative ecosystem of RNA biologists, chemists, AI researchers, automation engineers and clinical scientists, each contributing data, models and validation. The foundation model distills what constitutes a good RNA design into transferable knowledge, while the lab-in-the-loop keeps that knowledge anchored to the evolving reality of the wet laboratory; together they can carry RNA therapeutics from a craft of trial and error toward a more predictable and scalable engineering discipline. The most compelling applications are already in view: personalized RNA neoantigen vaccines tailored to the individual patient [Reference Sahin, Derhovanessian and Miller96,Reference Rojas, Sethna and Soares97,Reference Villanueva98], therapies for rare diseases underserved by conventional development models [Reference Kim, Hu and Moufawad el Achkar99], and rapid-response vaccine platforms that can be reprogrammed against emerging pathogens on compressed timelines [Reference Chaudhary, Weissman and Whitehead2,100]. These prospects are nonetheless bounded. The advantage for rare diseases is clearest where a well-defined molecular target exists, but patient numbers are too small to justify conventional development economics [Reference Kim, Hu and Moufawad el Achkar99]; it is more limited where the underlying mechanism is poorly characterized or where data remain too sparse for a model to generalize. Likewise, the promise of rapid-response platforms applies principally to the design phase: sequence and formulation can be reprogrammed quickly, but this compresses only one segment of the path from pathogen to approved product [Reference Saville, Cramer and Downham101]. A realistic appraisal must therefore separate what the platform accelerates from what it does not. The loop can shorten the design-test cycle and improve the quality of candidates entering development, but manufacturing, regulatory review, and clinical validation follow their own timelines and are largely independent of it [Reference Youssef, Hitti, Puppin Chaves Fulber and Kamen102]. Good manufacturing practice production and scale-up, the evidentiary requirements of regulatory approval, and the multi-year safety and efficacy testing of human trials remain rate-limiting regardless of how efficient candidate generation becomes. The platform’s contribution is to ensure that the candidates entering these stages are better characterized and more likely to succeed, not to bypass the stages themselves. Realizing this future will require a shared scientific community in which models learn from, and experiments continually test, each successive design.