TL;DR: MIT researchers published FLIGHTED on October 2, 2026 in Cell Systems as a Bayesian way to turn noisy high-throughput protein measurements into probabilistic fitness landscapes that lift standard model accuracy.

A Bayesian fix for noisy fitness data

The central result, described in the Cell Systems fulltext page for FLIGHTED: Inferring fitness landscapes from noisy high-throughput experimental data, is a Bayesian calculation that treats every fitness observation as uncertain rather than exact. The authors call the approach fitness landscape inference generated by high-throughput experimental data, shortened to FLIGHTED. Instead of collapsing a high-throughput experiment into single-point scores, the procedure outputs probabilistic landscapes that retain measurement uncertainty. According to that Cell Systems fulltext page, the probabilistic output is the mechanism for preserving uncertainty from noisy experiments. The same fulltext summary frames the contribution as generating probabilistic fitness landscapes directly from noisy experimental readouts.

The publication identifies the core technical highlight as a Bayesian method that accounts for noise in protein fitness measurements. That emphasis on noise-aware inference distinguishes the work from workflows that feed raw scores into predictors without adjustment. The team presents uncertainty handling as integral to landscape construction, not as a post-processing filter. In their formulation, the landscape itself encodes confidence alongside estimated function. That design choice lets downstream learning algorithms distinguish well-supported observations from weakly supported ones.

Who reported the method and where

The author list on the Cell Systems fulltext page names Vikram Sundar, Boqiang Tu, Lindsey Guan and Kevin Esvelt. The listing assigns Sundar to the Computational and Systems Biology Program at the Massachusetts Institute of Technology in Cambridge, MA, United States of America. It also connects the work to the Department of Biological Engineering at the Massachusetts Institute of Technology. A further affiliation links the project to the Media Lab at the Massachusetts Institute of Technology in Cambridge, MA.

Correspondence details on that Cell Systems fulltext page direct inquiries to Vikram Sundar at [email protected] and to Kevin Esvelt at [email protected]. The same page designates Kevin Esvelt as lead contact. The Media Lab affiliation is paired with the lead-contact role for Esvelt. The Computational and Systems Biology Program affiliation is paired with the corresponding-author role for Sundar.

The journal record in New articles: Cell Systems places first publication on October 02, 2026. That listing groups the paper under new articles published ahead of final issue assignment. It instructs readers to cite authors, year, title, journal and DOI. The October 02, 2026 entry is the newest verified publication fact for this report.

Why protein-design models need better training landscapes

Machine learning for protein design depends on extensive fitness collections produced by high-throughput experiments for both training predictors and benchmarking them, according to the Cell Systems fulltext page. Those collections map sequence variation to measured function across many variants at once. Training uses the maps to learn sequence-function relationships. Benchmarking uses the same type of maps to compare competing predictors.

The problem statement on that Cell Systems fulltext page is that most current models ignore the experimental noise embedded in those large collections. Ignoring that inherent variability degrades predictive performance. The page explicitly links omission of noise modeling to harmed model behavior. In other words, treating noisy readouts as ground truth weakens both learning and evaluation.

The FLIGHTED response keeps dataset scale while changing how confidence is represented. Large high-throughput fitness datasets remain the required input for training and benchmarking. The adjustment is to propagate experimental uncertainty into the landscape used for those two purposes. That framing positions noise accounting as a missing layer in existing protein-design pipelines.

What FLIGHTED actually builds

FLIGHTED expands the acronym as fitness landscape inference generated by high-throughput experimental data on the Cell Systems fulltext page. The method is Bayesian, meaning it formalizes uncertainty in fitness measurements during inference. The output is described as probabilistic fitness landscapes derived from noisy experiments. Those landscapes replace fixed scores with distributions that reflect experimental support.

The New articles: Cell Systems summary restates the idea as a Bayesian method for accounting for uncertainty in protein fitness measurements by generating probabilistic fitness landscapes from noisy high-throughput experiments. That description reinforces that uncertainty is modeled at landscape-construction time. The landscapes are therefore intended as direct substitutes for conventional point-estimate tables. Designers can train on and benchmark against representations that already contain noise information.

The authors state that the construction is broadly transferable across measurement platforms and learning approaches. According to the Cell Systems fulltext page, results indicate application to any high-throughput assay and any machine-learning model. The same page characterizes adoption as straightforward for protein designers seeking to account for noise. The generality claim covers both the experimental side and the modeling side.

Two experimental demonstrations

To show improved performance, the team applied the Bayesian landscapes in two distinct experimental settings detailed on the Cell Systems fulltext page. The first setting covers single-step selection assays, illustrated by phage display. The second setting covers a high-throughput assay that couples activity to base editing. Together the pair spans selection-based enrichment and editing-linked activity readout.

The New articles: Cell Systems entry likewise notes application to selection assays and to a high-throughput base-editing assay. That journal listing ties both demonstrations to improved performance of protein fitness models. The parallel mention confirms that the two cases are the evidentiary basis for the performance claim. No additional demonstration systems are named in the available summaries.

Single-step selection assays such as phage display represent a widely recognized high-throughput format for sampling functional variation. The high-throughput assay tying activity to base editing represents a distinct strategy linking measured activity to editing outcomes. By testing FLIGHTED in a single-step selection context and in a base-editing-linked activity context, the authors probe different sources of experimental noise. The Cell Systems fulltext page presents that two-context test as the demonstration of improved model performance.

Head-to-head comparison with standard models

The evaluation compares standard machine-learning models trained or tested on fitness landscapes with and without FLIGHTED, according to the Cell Systems fulltext page. Landscapes without the Bayesian adjustment serve as the baseline. Landscapes with the Bayesian adjustment serve as the noise-aware alternative. Standard models provide a constant modeling scaffold while the input representation changes.

The reported outcome is that accounting for noise statistically significantly improves model performance. The Cell Systems fulltext page uses that statistical-significance language to characterize the gain. The improvement is therefore presented not as a narrow or incidental shift but as a reliable difference under comparison. The New articles: Cell Systems summary similarly reports improved performance of protein fitness models after FLIGHTED processing.

The comparison design isolates the contribution of uncertainty handling. Because standard models are evaluated in both conditions, performance differences trace to landscape construction rather than to a new predictor architecture. That structure supports the interpretation that noise-aware landscapes enhance existing tools. The authors use the head-to-head result to argue for routine inclusion of uncertainty in fitness-model workflows.

Data size versus model scale

A further finding from the new benchmarking concerns what currently limits protein fitness model performance. Based on that FLIGHTED-based benchmarking, the Cell Systems fulltext page reports that data size, not model scale, is limiting performance. In this account, expanding informative datasets matters more than enlarging model architectures. The statement reframes priorities for groups investing in predictor development.

The data-size conclusion follows directly from benchmarking with noise-aware landscapes. By cleaning the measurement layer through Bayesian inference, the comparison reveals remaining bottlenecks more clearly. The authors interpret the pattern as evidence that dataset quantity and quality constrain gains. Model scaling alone does not overcome those constraints in their tests.

For protein designers, the implication stated on the Cell Systems fulltext page is practical: accounting for noise is straightforward and broadly applicable. Because FLIGHTED is described as compatible with any high-throughput assay and any machine-learning model, teams need not replace preferred assays or predictors. They can instead revise how noisy measurements become training and benchmarking landscapes.

How the journal describes the advance

The New articles: Cell Systems listing by Cell Systems condenses the paper as Sundar et al. presenting FLIGHTED, a Bayesian method of accounting for uncertainty in protein fitness measurements by generating probabilistic fitness landscapes from noisy high-throughput experiments. That editorial summary mirrors the author-supplied summary on the fulltext page. It confirms the method category, the uncertainty target and the landscape output.

The same New articles: Cell Systems entry specifies that FLIGHTED was applied to selection assays and to a high-throughput base-editing assay. It then notes improved performance of protein fitness models. That three-part sequence — method, two application domains, performance gain — matches the longer summary on the Cell Systems fulltext page. The consistency across journal pages anchors the core facts of this report.

First publication on October 02, 2026 places the work among Cell Systems new articles released ahead of issue pagination. The listing offers full-text HTML and PDF access routes. Readers are asked to cite the work by authors, year, title, journal and DOI while it awaits final issue placement.

What changes for training and benchmarking practice

Training and benchmarking of protein-design machine learning both rely on large fitness datasets from high-throughput experiments, as stated on the Cell Systems fulltext page. FLIGHTED intervenes before those uses by converting noisy experimental outputs into probabilistic landscapes. Training can then weight observations by support. Benchmarking can then assess predictors against uncertainty-aware references.

Most existing models omit that experimental-noise layer, with consequent harm to performance noted in the same Cell Systems fulltext summary. The proposed correction does not require a bespoke predictor. Standard machine-learning models show statistically significant improvement when evaluated with FLIGHTED-derived landscapes versus without them. That result suggests that landscape quality, not only algorithm choice, drives outcomes.

The stated generality extends the practice change beyond the two demonstrated systems. Indications on the Cell Systems fulltext page point to use with any high-throughput assay and any machine-learning model. Protein designers are therefore presented with a straightforward route to incorporate noise handling. The two demonstrated cases — single-step selection assays such as phage display and a high-throughput assay tying activity to base editing — serve as templates for wider transfer.

Limits defined by the available record

The available Cell Systems summaries define the scope without providing numerical effect sizes, variant counts or software details. They name single-step selection assays such as phage display and a high-throughput assay tying activity to base editing as the tested cases. They do not name additional assays in the summarized record. They compare standard models with and without FLIGHTED without enumerating architectures in the summarized record.

The statistical claim is qualitative in the accessible text: accounting for noise statistically significantly improves model performance. The data-scale claim is likewise qualitative: data size, not model scale, limits protein fitness model performance under the new benchmarking. The transfer claim is framed as an indication rather than an exhaustive validation: results indicate application to any high-throughput assay and any machine-learning model.

Authorship, affiliation and contact facts are fully specified on the Cell Systems fulltext page through Vikram Sundar, Boqiang Tu, Lindsey Guan and Kevin Esvelt and through Massachusetts Institute of Technology units in Cambridge, MA. Publication timing is fixed by New articles: Cell Systems at October 02, 2026. Together those publication facts bound what can be reported without reaching beyond the pre-cutoff public record.

Why the result matters for United States protein engineering

The work originates from Massachusetts Institute of Technology programs spanning computational and systems biology, biological engineering and the Media Lab. That institutional placement situates the advance within United States academic research in protein design and systems biology. The corresponding and lead-contact roles assigned to Vikram Sundar and Kevin Esvelt provide clear responsibility for the Bayesian method.

For United States science and technology readers tracking machine learning for protein design, the takeaway is that larger high-throughput fitness datasets help only if their noise is modeled. The Cell Systems fulltext page makes large datasets central to training and benchmarking while warning that unmodeled noise harms performance. FLIGHTED offers a Bayesian bridge between raw high-throughput readouts and usable probabilistic landscapes.

The October 02, 2026 Cell Systems debut establishes the method as peer-reviewed literature rather than an isolated preprint claim. Its appearance alongside other Cell Systems new articles underscores active interest in quantitative biology tools. As designers adopt noise-aware landscapes for selection assays, base-editing-linked assays and potentially any high-throughput assay, standard models may gain accuracy without architectural overhaul.

Sources referenced