InstaNovo, a cutting-edge transformer neural network, addresses the challenges of de novo peptide sequencing in mass spectrometry-based proteomics. Trained on 28M spectra, InstaNovo outperforms current state-of-the-art methods and showcases its utility in several applications. We further introduce InstaNovo+, a multinomial diffusion model that improves performance via iterative refinement. Together, these models unlock a plethora of opportunities across different scientific domains, including direct protein sequencing, immunopeptidomics, and dark proteome exploration.

First, we generated a large panel of binder and non-binder antibody sequences to the cancer immunotherapy targets PD-1 and CTLA-4. Next, we encoded the antibody light and heavy chain complementarity-determining regions (CDR3s) into antibody images, then built and trained convolutional neural network models to classify binders and nonbinders. We then built generative deep learning models, using generative adversarial network models to produce synthetic antibodies against PD-1 and CTLA-4.

Notwithstanding the claims, it is challenging to understand how effective ML/AI and other computational methods really are in antibody discovery and optimization. This talk will describe the launch of a blinded benchmarking competition using several Specifica NGS datasets which will be made available to the community. We hope this will become a regular CASP-like competition, with challenges of increasing difficulty.

Antibody language models are an emerging class of tools with the potential to revolutionize our ability to understand the linkage between antibody sequence, structure, and function. The paucity of natively paired antibody sequence datasets means most antibody language models have been trained exclusively using unpaired antibody sequences, however, we have shown that training with natively paired sequences allows models to learn critically important cross-chain features. This talk will explore the utility of training using natively paired antibody sequences and explore strategies for training such models in a data-limited environment.

The fusion of AI with biologics design offers unprecedented potential in drug discovery. However, building platforms that enable AI-guided drug discovery is not a straightforward process. I’ll walk through how we’ve built and scaled new biotech platforms from the ground up for de novo biologics design.

Machine learning-based models of protein fitness typically learn from either unlabeled, evolutionarily related sequences (such as in the case of protein language models) or variant sequences, with experimentally measured labels. For regimes where only limited experimental data are available, we combine both sources of information in a simple combination approach that is competitive with, and on average outperforms more sophisticated methods.

Deep learning techniques hold the potential to accelerate identification of effective antibodies but the existing methods cannot provide the confidence interval or uncertainty needed to assess the reliability of the predictions. Here we present a pipeline called RESP for efficient identification of high affinity antibodies using uncertainty-aware machine learning methods.

Antibody developability properties are dependent on the relative disposition of constituent amino acids. We develop a graph neural network with meaningful biological priors such as Delaunay-based adjacency and Kidera factors to build an efficient and explainable model to predict antibody developability.

We present a platform for synthetic protein-protein coevolution, isolating diversely remodeled interacting pairs. This dataset enables comprehensive analysis of protein pairs, uncovering insights into structural diversity, affinities, cross-reactivities, and orthogonalities. Leveraging pretrained protein language models, we expand the amino acid diversity of our coevolution screen in silico, predicting remodeled interfaces. This integrated approach simulates protein coevolution, creating protein complexes with diverse recognition properties, benefiting biotechnology and synthetic biology.

• Design parameters and objectives for cytokine engineering
• Tailoring engineering strategies to various cytokine systems
• Considerations for selecting the appropriate evolutionary and/or rational design strategies
• Key factors in choosing a computational workflow for cytokine engineering

• Which tools are out there and how good are they? 
• Which applications benefit most from de novo peptide sequencing?
• What are current limitations and opportunities for further development?

Identification of favorable biophysical properties for protein therapeutics as part of developability assessment is a crucial part of the preclinical development process. Successful prediction of such properties and bioassay results from calculated in silico features has potential to reduce the time and cost of delivering clinical-grade material to patients. We have developed and implemented an automated machine learning workflow designed to compare and identify the most powerful features from computationally derived physiochemical feature sets. We demonstrate the use of this workflow with medium-sized datasets of IgG molecules to generate predictive regression models for key developability endpoints.

Cytokines are soluble factors that signal through stimulation of their cognate transmembrane receptors on target cells to perform critical biological functions, particularly those related to immune homeostasis. We synthesized a best-in-class computational protein design software with directed evolution technologies to generate a de novo engineered cytokine mimetic with superior stability and unique biochemical activities compared to naturally-derived cytokines. Collectively, our work pioneers a novel cytokine engineering platform that integrates computational and experimental approaches to create new molecules that expand the repertoire of protein function and inform the development of next-generation immunotherapeutics.

Gene editing has the potential to solve fundamental challenges in agriculture, biotechnology, and human health. Using large language models (LLMs) trained on biological diversity at scale, we demonstrate the first successful precision editing of the human genome with programmable gene editors designed with AI. We release OpenCRISPR-1, which showed comparable or improved activity and specificity relative to SpCas9 while being 400 mutations away in sequence, for free and ethical usage.

Current Ab-specific PLMs are limited in knowledge due to the use of only unpaired sequences and the traditional masked-language-modeling approach of self-supervised training, and therefore, are not ideal for utilization for tasks such as affinity, thermal stability, contact-map / epitope / paratope prediction, etc. We aim to alleviate this knowledge limitation in our Next-Gen PLM by infusing information from multiple internal and external data modalities and develop new training strategies, to improve performance on these different tasks and beyond.