Computable surface properties of therapeutic proteins like antibodies can help predict biophysical properties relevant to drug formulation. Property predictions based on single states may underestimate developability risks, motivating molecular dynamics (MD) simulation. To probe whether the initial model seeding the simulation affects risk prediction, we used five ML-based structure prediction methods and compared to simulation results using experimental structures to determine an ideal workflow. 

The generation of multispecific protein therapeutics necessitates the exploration of extensive design spaces, a task that cannot be entirely covered through wet lab experiments alone. By harnessing our systematically-collected and curated data assets, we have developed computational and machine learning-based optimization workflows for predicting molecular properties such as expression, activity, stability, and clearance. We will showcase examples of how our computational pipeline assists in navigating through the vast design space of multispecific biologics.

In this case study, we employ transformer-based language model embeddings to decode the genetic and somatic diversity of human antibody sequences. Through self-attention mechanisms, our model garners insights into the Structure-Activity Relationship (SAR) and potential disease correlations. Remarkably, despite limited sequence homology, antibodies exhibiting similar structural variations converge in a multi-dimensional latent space, illuminating subtleties in the SAR. By creating a distinct digital signature for any antibody within the adaptive human immunome, the development of language models for predicting binding and molecular developability becomes feasible.

We have developed a protein design pipeline, called EvoPro, that uses iterative rounds of deep learning-based structure prediction and sequence optimization to evolve protein sequences for prespecified design goals. Initial experimental testing has focused on the creation of small, stable proteins that bind to target binding surfaces on proteins of interest. Without any experimental optimization, low nanomolar binders were designed against a PD-L1 antagonist. 

De novo protein design seeks to learn the underlying principles of protein folding from natural proteins and to subsequently apply them to generate novel proteins with programmable functions. In this talk, I will describe the development of RFdiffusion, a generative neural network for de novo protein design which demonstrates state-of-the-art performance, both in silico and experimentally, across a broad range of therapeutically relevant design challenges.

The ability to correctly predict the structure of antibody-antigen complexes in silico would provide a lot of value for medical applications. Recent deep learning technologies have enabled the production of antibody models with significantly better quality than traditional tools. We evaluated if such quality is sufficient for successful antibody-antigen structure prediction, using traditional molecular docking tools that normally fall short in real applications where the antibody structure is unknown.

Heavy and light chain pairing in antibodies involves complex mechanisms of diversification and maturation, which remain to be elucidated. Borrowing techniques from natural language processing, we show that antibody chain pairing can be modeled as a machine translation problem. The generated synthetic library resembles natural antibody pairing preferences down to gene family level and correlates with functions previously unseen during training.

High-affinity antibodies are often identified through directed evolution, which may require many iterations of mutagenesis and selection to find an optimal candidate. Deep learning techniques hold the potential to accelerate this process, 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. Importantly, this model can quantify their likelihood to be tight binders for sequences not present in the directed evolution library, and thus greatly expand the search space to uncover the best sequences for experimental evaluation.

DeepAb, a deep learning model for predicting antibody Fv structure directly from sequence, was used to design 200 potentially stabilized variants of an anti-hen egg lysozyme (HEL) proof of concept antibody [FY1]. 85% of the clones exhibited increased thermal and colloidal stability. Of these, 11% showed a highly increased affinity for HEL (1.5- to 10-fold increase) while retaining the developability profile of the parental antibody. These data open up the possibility of in silico antibody stabilization and affinity maturation, without the need to predict the antibody-antigen interface.

Leveraging AlphaSeq, a library-on-library method for studying protein-protein interactions, we screened a large library of synthetic antibodies to identify cross-reactive binders against human TIGIT and its mouse ortholog. Cross-reactive antibodies were further optimized for binding and bio-developability following multiple cycles of data generation, computational model training, sequence proposals, and binding validation assays. This process improved binding affinity and generated hundreds of antibody candidates with favorably predicted developability properties.

Prediction of biophysical properties for protein therapeutics from calculated in silico features has potential to reduce the time and cost of delivering clinical-grade material to patients. We have developed an automated machine learning workflow designed to 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.

A single administration of CIS43 can protect against malaria infection in the field for up to 9 months. In this in silico work, we improved the potency of CIS43 antibody variants by optimizing the binding energy to its target antigen PfCSP. In silico variants yielded increased affinity and superior protective efficacy to a malaria in vivo mouse challenge model. The best variant, P3-43, showed ~10-fold improvement in protection relative to CIS43. Our in silico pipeline provides a generally applicable approach to improve antibody functionality, and leads to novel variants to be used for passive prevention against malaria.

Bispecific antibodies have emerged as attractive modalities for the development of new therapeutics but come with certain drawbacks such as fixed stoichiometry for target binding and more challenging manufacturing. Biolojic Design has developed multibodies that overcome these challenges, by using our proprietary AI technology. Multibodies have the ability to flexibly bind two different targets, while maintaining a standard IgG format. This technology enables new applications for the development of therapeutics.

Polyreactive antibodies lead to incorrect experimental results and are intractable for clinical development. We designed a set of experiments using a synthetic camelid antibody fragment ("nanobody") library to train machine learning models to assess polyreactivity from protein sequence. Our models provide quantitative scoring metrics that predict polyreactivity. We experimentally test our models’ performance on three nanobody scaffolds, finding that over 90% of predicted substitutions reduced polyreactivity.

Typically, protein engineering is labor-intensive and relies on sequential trial-and-error search. But the protein space is greater than the number of atoms in the universe. Therefore, to successfully create new designer proteins with ever-increasing complexity, it's necessary to generate to specifications, rather than search. At 310.ai, we aim to bypass trial-and-error, and instead, offer a highly parallel and coherent design solution that will impact speed, cost, and quality.

The application of machine learning for protein drug discovery combined with automated experimental systems allows the creation of advanced platforms to accelerate drug discovery. By leveraging both digital and physical automation, large-scale, low-aleatoric noise datasets can be created quickly for training and fine-tuning machine learning models. These models, in turn, can enhance the accuracy of experimental cycles, creating symbiotic feedback to deliver drugs faster than previously possible.