Biology subjects

Lala, J.

Publications and source records attributed to Lala, J..

6 recordsLinked to original sources

bioRxiv2026

BAGEL-CAR: Reflections on the Bits to Binders Competition

The antigen-binding segment of chimeric antigen receptors (CARs) in CAR-T therapy has emerged as a compelling application of de novo AI protein design. In the Bits to Binders competition, our group submitted 414 BAGEL-CAR designs for CD20-directed CAR binding segments, 38.4% of which were statistically enriched in a pooled CAR-T proliferation screen, the highest among all teams and more than double the next-best. Strikingly, after detailing the three design strategies employed, we show that two of them carry the entire success of our campaign, with both of them outcompeting all the other teams' aggregate results. Crucially, during the competition, we did not refold any of the binders after backbone painting. Had we applied this typical validation step with ESMFold, all experimentally enriched designs would have been filtered out. Our findings show that each filtering step should be stress-tested per biomolecular system and intermediate outputs of design pipelines should be assayed regardless of 'common practice'. Additionally, we appear to be the only team that designed against ESMFold's predicted CD20 dimer geometry, which diverges substantially from the experimentally resolved 6Y97 structure, raising, but not resolving, the possibility that this alternative target geometry contributed to our high hit rate. We hope these reflections help inform the future of effective CAR-T and binder design campaigns.

bioinformatics↗

bioRxiv2026

An Energy Landscape Approach to Miniaturizing Enzymes using Protein Language Model Embeddings

AO_SCPLOWBSTRACTC_SCPLOWWe present a general approach to find amino acid sequences corresponding to the most compact enzyme likely to retain the structure of a given catalytic site. Our approach is based on using Monte Carlo (MC) simulations to sample an energy landscape where minima correspond, by construction, to sequences with the aforementioned properties. Building on previous work (Wu et al., 2025) and with the BAGEL package (Lala et al., 2025), we implement a route to achieve this goal using only the information extracted from a protein language model (PLM), without structural information. After generating a set of candidate sequences with this PLM-guided BAGEL optimization, we further filter potential candidates for downstream experimental validation using a two-stage protocol. First, deep-learning-based structure prediction models (ESMFold, Chai-1, Boltz-2) are used to identify a structural consensus among designs with highly conserved active-site geometries, yielding many candidates with active-site RMSD below a few angstroms relative to the wild-type and pLDDT scores above 80. Second, molecular dynamics simulations are performed on a filtered subset of sequences (based on active-site RMSD and SolubleMPNN log-likelihoods) to evaluate active-site stability when including thermal fluctuations. For the most promising enzymes, these yield RMSF values in the active site below 1.0 [A] and an active-site RMSD drift between 0.5 and 1.5 [A], making these mini-variants comparable to the wild type, though outcomes vary across enzymes. Given the protocols generality, we believe these results represent a step forward in AI-guided enzyme design. To facilitate rapid experimental validation by the broader community, we open-source all sequences generated by our computational pipeline. These include designs for four representative enzymes of this study: PETase, subtilisin Carlsberg (serine protease), Taq DNA polymerase, and VioA.

bioinformatics↗

bioRxiv2026

Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition

Artificial intelligence (AI) methods for proteins have advanced rapidly, improving structure prediction and design, particularly for de novo binders. However, most evaluations emphasize binding affinity rather than higher-order biological function. We present Bits to Binders, a global competition benchmarking de novo binder design in the context of chimeric antigen receptor (CAR) T cells. Teams from 42 countries submitted 12,000 designs of 80-amino acid binders targeting human CD20 as CAR binding domains. Designs were screened by pooled CAR-T proliferation, identifying 707 designs exhibiting significant CD20-specific enrichment, with team hit rates from 0.6% to 38.4%. Top-performing candidates were validated as individual constructs, measuring CD20-specific proliferation, expansion, cytokine production, and targeted cell lysis. We identified common design methodologies and factors correlated with DNA synthesis, expression, and target-specific T cell activation which nearly double the success rates when applied as a retrospective filter. We release this dataset as an open resource, with practical recommendations to support more effective AI-driven binder design.

bioinformatics↗

bioRxiv2025

Harnessing protein-folding algorithms to drug intrinsically disordered epitopes

Due to their lack of a specific structure and dynamical nature, targeting of epitopes that are part of an intrinsically disordered region of a protein is a notoriously difficult task. Here, we describe a computational approach to overcome this problem, based on the use of a protein-folding algorithm and its confidence metrics within a Monte Carlo optimization pipeline to generate peptide-based binders. For different protein targets, we show by accurate free energy calculations that our approach is able to design peptides with binding free energies on the order of tens of kBT, i.e., with strengths comparable to covalent interactions. Direct observation of the bound complex through molecular simulations shows that the targeted epitope folds into structured domains with lowered thermal fluctuations upon binding, while remaining unstructured and dynamic in the unbound state, suggesting that the protein-folding algorithm must have learned the principles of induced (co-)folding. Given the ubiquitous presence of unstructured regions in proteins, our results suggest a potential pathway to design drugs targeting a large variety of previously untargetable epitopes, and open new possibilities for therapeutic intervention in diseases where disordered proteins play a key role. I SIGNIFICANCESmall-molecule drugs that bind to a protein via a lock-and-key mechanism require the targeted epitope to form a well-structured, stable binding pocket, thereby preventing binding to unstructured regions. To overcome this limit, we present a general approach, based on a protein-folding algorithm, to find peptide sequences that induce the formation of such a binding interface when no pocket is normally present. In other words, we show how a protein-folding algorithm can be used to program epitope recognition by induced folding instead of rigid lock-and-key matching. In this way, we show that we can extend druggable epitopes to intrinsically disordered, dynamical regions of proteins.

bioinformatics↗

bioRxiv2025

BAGEL: Protein Engineering via Exploration of an Energy Landscape

Despite recent breakthroughs in deep learning methods for protein design, existing computational pipelines remain rigid, highly specific, and ill-suited for tasks requiring non-differentiable or multi-objective design goals. In this report, we introduce BAGEL, a modular, open-source framework for programmable protein engineering, enabling flexible exploration of sequence space through model-agnostic and gradient-free exploration of an energy landscape. BAGEL formalizes protein design as the sampling of an energy function, either to optimize (find a global optimum) or to explore a basin of interest (generate diverse candidates). This energy function is composed of user-defined terms capturing geometric constraints, sequence embedding similarities, or structural confidence metrics. BAGEL also natively supports multi-state optimization and advanced Monte Carlo techniques, providing researchers with a flexible alternative to fixed-backbone and inverse-folding paradigms common in current design workflows. Moreover, the package seamlessly integrates a wide range of publicly available deep learning protein models, allowing users to rapidly take full advantage of any future improvements in model accuracy and speed. We illustrate the versatility of BAGEL on four archetypal applications: designing de novo peptide binders, targeting intrinsically disordered epitopes, selectively binding to species-specific variants, and generating enzyme variants with conserved catalytic sites. By offering a modular, easy-to-use platform to define custom protein design objectives and optimization strategies, BAGEL aims to speed up the design of new proteins. Our goal with its release is to democratize protein design, abstracting the process as much as possible from technical implementation details and thereby making it more accessible to the broader scientific community, unlocking untapped potential for innovation in biotechnology and therapeutics.

bioinformatics↗

bioRxiv2025

Mind the Gap: An Embedding Guide to Safely Travel in Sequence Space

We present a hybrid approach combining a protein language model (pLM) with Monte Carlo (MC) sampling for generating enzyme mutants free of mutations deleterious for structural preservation. Given the amino acid sequence of the original enzyme and a set of residues for which the local environment should be conserved, i.e., the catalytic site, our approach generates mutants that differ vastly in the overall sequence while retaining the geometry of the conserved region, thereby representing promising candidates for further experimental screening. Unlike end-to-end deep learning approaches, whose results are harder to interpret and control, the use of a well-established, classic technique such as MC sampling allows us to easily interpret the generative process as the sampling of an energy landscape determined by the pLM. In turn, such an interpretation enables us to steer this generative process and control its outcome by making use of robust statistical mechanics concepts, e.g., temperature, thereby explicitly guaranteeing certain properties of the generated mutants. Given the increasing relevance of generative algorithms in the design and search for novel, optimised enzymes, we believe that our results constitute an important step for the future development of this class of techniques. To facilitate experimental verification, we finally provide hundreds of sequences for 13 different enzymes involved in catalytic processes ranging from carbon dioxide conversion to DNA replication.

synthetic biology↗