NTNeil Thomas

Neil Thomas

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Neil Thomas has a total of 106 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Tuned Fitness Landscapes for Benchmarking Model-Guided Protein Design

    by Neil Thomas, Atish Agarwala, David Belanger, Yun S Song, Lucy J Colwell

    Machine learning-guided directed evolution (MLDE) is an emerging paradigm of protein design where ML models trained on experimental data are used to guide exploration of a protein fitness landscape. Building an MLDE pipeline involves many design choices across the design-build-test-learn loop ranging from data collection strategies to modeling. The cost of collecting experimental data makes benchmarking every component of these pipelines prohibitively difficult, necessitating the development of synthetic landscapes. In this work, we develop a framework called SLIP (“Synthetic Landscape Inference for Proteins”) for constructing synthetic landscapes with tunable difficulty based on Potts models. This framework can be extended to any protein family for which there is a sequence alignment. We show that our tuning framework provides landscapes sufficiently challenging to benchmark MLDE pipelines. SLIP is open-source and is available at https://github.com/google-research/slip.

  • Top Cited Papers
    Open Access

    Language Modeling Materializes a World Model of Protein Biology

    by Salvatore Candido, Thomas Hayes, Alexander Derry, Roshan Rao, Zeming Lin, Robert Verkuil, Bryan Z. Wu, Jin Sub Lee, Elise S. Bruguera, Jehan A. Keval, et al.

    Proteins are fundamental to life. The full extent of their biology is beyond our ability to characterize with experimental approaches in the physical laboratory. Accurate digital representations could accelerate the discovery of protein biology through virtual experiments. We propose language modeling to learn unified and general representations that can be scaled to all of protein biology. Building on these representations, we develop a structure prediction model that exceeds the performance of established methods for biomolecular complex prediction across benchmarks, including for the interactions of antibodies with their targets. A simple search procedure yields high experimental success rates for the discovery of proteins with nanomolar binding affinities for both miniproteins and single-chain antibodies, a modality critical for therapeutic design. Study of the concepts in the language model's representation space reveals a systematic organization aligned with the reductionist understanding of proteins developed through empirical science. Leveraging this organization, we generate a comprehensive map of protein biology encompassing over 6.8 billion sequences and 1.1 billion predicted structures, identifying connections across known and unknown biology. As a whole, this shows language modeling as a powerful substrate for representing the biology of proteins, operating across scales from the prediction and design of protein interactions at the atomic level, to identifying properties of proteins at different levels of granularity and abstraction, to the scale of mapping connections between proteins across billions of years of evolution.


  • Top Cited Papers
    Open Access

    Engineering highly active nuclease enzymes with machine learning and high-throughput screening

    by Neil Thomas, David Belanger, Chenling Xu, Hanson Lee, Kathleen Hirano, Kosuke Iwai, Vanja Polic, Kendra D. Nyberg, Kevin G. Hoff, Lucas Frenz, et al.


  • Top Cited Papers

    Simulating 500 million years of evolution with a language model

    by Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J. Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q. Tran, Jonathan Deaton, Marius Wiggert, et al.

    More than 3 billion years of evolution have produced an image of biology encoded into the space of natural proteins. Here, we show that language models trained at scale on evolutionary data can generate functional proteins that are far away from known proteins. We present ESM3, a frontier multimodal generative language model that reasons over the sequence, structure, and function of proteins. ESM3 can follow complex prompts combining its modalities and is highly responsive to alignment to improve its fidelity. We have prompted ESM3 to generate fluorescent proteins. Among the generations that we synthesized, we found a bright fluorescent protein at a far distance (58% sequence identity) from known fluorescent proteins, which we estimate is equivalent to simulating 500 million years of evolution.


  • Open Access

    Engineering of highly active and diverse nuclease enzymes by combining machine learning and ultra-high-throughput screening

    by Neil Thomas, David Belanger, Chenling Xu, Hanson Lee, Kathleen Hirano, Kosuke Iwai, Vanja Polic, Kendra D Nyberg, Kevin G Hoff, Lucas Frenz, et al.

    Optimizing enzymes to function in novel chemical environments is a central goal of synthetic biology, but optimization is often hindered by a rugged, expansive protein search space and costly experiments. In this work, we present TeleProt, an ML framework that blends evolutionary and experimental data to design diverse protein variant libraries, and employ it to improve the catalytic activity of a nuclease enzyme that degrades biofilms that accumulate on chronic wounds. After multiple rounds of high-throughput experiments using both TeleProt and standard directed evolution (DE) approaches in parallel, we find that our approach found a significantly better top-performing enzyme variant than DE, had a better hit rate at finding diverse, high-activity variants, and was even able to design a high-performance initial library using no prior experimental data. We have released a dataset of 55K nuclease variants, one of the most extensive genotype-phenotype enzyme activity landscapes to date, to drive further progress in ML-guided design.


  • Open Access

    Functional genomics of OCTN2 variants informs protein-specific variant effect predictor for Carnitine Transporter Deficiency

    by Megan L. Koleske, Gregory McInnes, Julia E. H. Brown, Neil Thomas, Keino Hutchinson, Marcus Y. Chin, Antoine Koehl, Michelle R. Arkin, Avner Schlessinger, Renata C. Gallagher, et al.

    Genetic variants in SLC22A5 , encoding the membrane carnitine transporter OCTN2, cause the rare metabolic disorder Carnitine Transporter Deficiency (CTD). CTD is potentially lethal but actionable if detected early, with confirmatory diagnosis involving sequencing of SLC22A5 . Interpretation of missense variants of uncertain significance (VUSs) is a major challenge. In this study, we sought to characterize the largest set to date ( n = 150) of OCTN2 variants identified in diverse ancestral populations, with the goals of furthering our understanding of the mechanisms leading to OCTN2 loss-of-function (LOF) and creating a protein-specific variant effect prediction model for OCTN2 function. Uptake assays with 14 C-carnitine revealed that 105 variants (70%) significantly reduced transport of carnitine compared to wild-type OCTN2, and 37 variants (25%) severely reduced function to less than 20%. All ancestral populations harbored LOF variants; 62% of green fluorescent protein (GFP)–tagged variants impaired OCTN2 localization to the plasma membrane of human embryonic kidney (HEK293T) cells, and subcellular localization significantly associated with function, revealing a major LOF mechanism of interest for CTD. With these data, we trained a model to classify variants as functional (>20% function) or LOF (<20% function). Our model outperformed existing state-of-the-art methods as evaluated by multiple performance metrics, with mean area under the receiver operating characteristic curve (AUROC) of 0.895 ± 0.025. In summary, in this study we generated a rich dataset of OCTN2 variant function and localization, revealed important disease-causing mechanisms, and improved upon machine learning–based prediction of OCTN2 variant function to aid in variant interpretation in the diagnosis and treatment of CTD.


  • Open Access

    Tuned Fitness Landscapes for Benchmarking Model-Guided Protein Design

    by Neil Thomas, Atish Agarwala, David Belanger, Yun S. Song, Lucy J. Colwell

    Advancements in DNA synthesis and sequencing technologies have enabled a novel paradigm of protein design where machine learning (ML) models trained on experimental data are used to guide exploration of a protein fitness landscape. ML-guided directed evolution (MLDE) builds on the success of traditional directed evolution and unlocks strategies which make more efficient use of experimental data. Building an MLDE pipeline involves many design choices across the design-build-test-learn loop ranging from data collection strategies to modeling, each of which has a large impact on the success of designed sequences. The cost of collecting experimental data makes benchmarking every component of these pipelines on real data prohibitively difficult, necessitating the development of synthetic landscapes where MLDE strategies can be tested. In this work, we develop a framework called SLIP (“Synthetic Landscape Inference for Proteins”) for constructing biologically-motivated synthetic landscapes with tunable difficulty based on Potts models. This framework can be extended to any protein family for which there is a sequence alignment. We show that without tuning, Potts models are easy to optimize. In contrast, our tuning framework provides landscapes sufficiently challenging to benchmark MLDE pipelines. SLIP is open-source and is available at https://github.com/google-research/slip .


  • Open Access

    End-to-end learning of multiple sequence alignments with differentiable Smith-Waterman

    by Samantha Petti, Nicholas Bhattacharya, Roshan Rao, Justas Dauparas, Neil Thomas, Juannan Zhou, Alexander M. Rush, Peter K. Koo, Sergey Ovchinnikov

    Multiple Sequence Alignments (MSAs) of homologous sequences contain information on structural and functional constraints and their evolutionary histories. Despite their importance for many downstream tasks, such as structure prediction, MSA generation is often treated as a separate pre-processing step, without any guidance from the application it will be used for. Here, we implement a smooth and differentiable version of the Smith-Waterman pairwise alignment algorithm that enables jointly learning an MSA and a downstream machine learning system in an end-to-end fashion. To demonstrate its utility, we introduce SMURF (Smooth Markov Unaligned Random Field), a new method that jointly learns an alignment and the parameters of a Markov Random Field for unsupervised contact prediction. We find that SMURF learns MSAs that mildly improve contact prediction on a diverse set of protein and RNA families. As a proof of concept, we demonstrate that by connecting our differentiable alignment module to AlphaFold and maximizing predicted confidence, we can learn MSAs that improve structure predictions over the initial MSAs. Interestingly, the alignments that improve AlphaFold predictions are self-inconsistent and can be viewed as adversarial. This work highlights the potential of differentiable dynamic programming to improve neural network pipelines that rely on an alignment and the potential dangers of relying on black-box methods for optimizing predictions of protein sequences.


  • Open Access

    Single Layers of Attention Suffice to Predict Protein Contacts

    by Nicholas Bhattacharya, Neil Thomas, Roshan Rao, Justas Dauparas, Peter K. Koo, David Baker, Yun S. Song, Sergey Ovchinnikov

    A bstract: The established approach to unsupervised protein contact prediction estimates co-evolving positions using undirected graphical models. This approach trains a Potts model on a Multiple Sequence Alignment, then predicts that the edges with highest weight correspond to contacts in the 3D structure. On the other hand, increasingly large Transformers are being pretrained on protein sequence databases but have demonstrated mixed results for downstream tasks, including contact prediction. This has sparked discussion about the role of scale and attention-based models in unsupervised protein representation learning. We argue that attention is a principled model of protein interactions, grounded in real properties of protein family data. We introduce a simplified attention layer, factored attention , and show that it achieves comparable performance to Potts models, while sharing parameters both within and across families. Further, we extract contacts from the attention maps of a pretrained Transformer and show they perform competitively with the other two approaches. This provides evidence that large-scale pretraining can learn meaningful protein features when presented with unlabeled and unaligned data. We contrast factored attention with the Transformer to indicate that the Transformer leverages hierarchical signal in protein family databases not captured by our single-layer models. This raises the exciting possibility for the development of powerful structured models of protein family databases. 1