We have shown that synthetic miniproteins computationally designed for foldedness can function as developable, evolvable scaffolds for ligand discovery. We leverage the data of differential performance across designs to gain fundamental insights and develop stronger scaffolds. Moreover, we advance lead molecules as molecular therapeutics.

We're facing a challenge in the world of chemistry: our lack of data is slowing down how we can use clever computer programs, known as machine learning, to create powerful new medicines. In this piece, I'll walk you through what we're doing to solve this problem by creating data highways from millions of small molecules, peptides and small proteins. We are now able to use machine learning to discover and create new functional molecules quickly. Sometimes, these computer-designed molecules are even better than what we can make ourselves! Our next step is to create an infinite loop where we automatically design, build, and test potential new medicines.

We will describe strategies to select for cyclotide precursors on the yeast cell surface, which are then subsequently converted to cyclotides following soluble expression. These ruggedly stable binders have potential applications for cytoplasmic delivery, radioligand therapy, and oral delivery. The rules for library design, selection & counterselection strategy, and exploitation of NGS data are different than for antibody libraries, and are beginning to be explicated.

BigHat Biosciences has developed novel machine learning (ML) approaches that leverage our high speed, automated wet lab in order to rapidly and iteratively design over a thousand next generation therapeutic antibodies each week. BigHat’s algorithmic approach pairs with our unique wet lab to guide the search for better molecules by learning from each cycle of characterization across multi-objective affinity, function, and developability measures of each antibody. We’ll discuss several methodological developments in multi-parameter optimization, active learning (Bayesian Optimization), and generative humanization as well as highlight the functional and in vivo validation of our designs.

In this presentation, I will delve into recent advancements in antibody library technologies, particularly highlighting selection of pH-sensitive antibodies from our Generation 3 library. We have selected antibodies with remarkable differential binding at pH 7.4 versus pH 6 (and vice versa), with affinity differences of 50-100-fold in both orientations. Additionally, I will describe how our platform is ideal for enhancing existing AI capabilities.

CAR T and anti-CD47 therapy are two distinct immunotherapies that we found to be non-compatible in combination due to depletion of adoptively transferred T cells by macrophages. Using yeast surface display, we engineered CD47 for selective binding to be insensitive to CD47 therapy, but still function as a "don't eat me" signal. We demonstrated that the combination of CAR T cells expressing engineered CD47 and CD47 blockade results in synergistic control of multiple solid tumors by harnessing T cell and macrophage antitumor activity.

Traditional antibody generation allows limited control over the epitopes of resulting sub-optimal antibodies. We have leveraged our approach to conditional gene knockout during immunization to direct optimal antibody responses to epitopes of interest. Our technology can unlock rational antibody discovery for a new generation of antibody-based therapeutics.

Existing methods for biologics discovery offer complementary advantages but each come with their drawbacks. We present a microfluidics-based discovery engine to generate, enrich, and screen natively paired libraries, thereby combining the strengths of B cell and display-based platforms. We have extensively used this platform to rapidly isolate potent biologics across therapy areas and modalities.

SNS-101 is a yeast-based, fully human conditionally active anti-VISTA antibody designed to relieve T cell suppression driven by the VISTA-PSGL-1 immune checkpoint within the acidic tumor-microenvironment, currently in a Phase I study (NCT05864144). Biochemical and structural data show SNS-101’s exquisite selectivity for active, protonated VISTA. Preclinical data demonstrate SNS-101’s pH-selectivity abrogates target mediated drug disposition (TMDD) and significantly reduces the risk of cytokine release syndrome (CRS), suggesting SNS-101 may overcome prior hurdles of anti-VISTA biologics. SNS-101 displayed strong anti-tumor activity in tumor models as both monotherapy and in combination with PD-1 blockade.

The presentation will highlight preclinical data from the development of our clinical candidates APVO436 (CD123xCD3) and ALG.APV-527 (4-1BBx5T4); and our preclinical candidates APVO603 (4-1BBxOX40), APVO711 (PD-L1xCD40), and APVO422 (PSMAxCD3). All these molecules show very clean target-activated profile. ADAPTIR and ADAPTIR-FLEX formats allow us to create different geometry and valency. In our development strategy we implemented early functional screening of a diverse group of binding domains against a given target, and combined it with screening in multiple possible final formats with different geometry and valency and as a result with a different functional activity.

There are different kinds of challenging antibody selection’s targets. Sometimes the target differs very minimally from close relatives and needs to be detected with high accuracy and sensitivity. In this case the aim of the selection is finding multiple needles snuggly fitting different pinholes in one piece of cloth but not a second piece of cloth. Other times the target is a library to targets. In this case the aim of the selection is finding a set of needles in one haystack, matching a set of needles in a second haystack. Some other times the target is not a "thing" but it could be e.g. machine learning of antibody features correlated with a certain behavior (e.g. developability). 

In this case the aim of the selection is dividing a big haystack in smaller stacks with opposite features and everything in between. We will discuss examples of selection strategies customized to these three examples of difficult targets, including: 1) selection of antibodies against SARV CoV-2 variants not recognizing SARS-CoV-1; 2) a method for cross interrogation of yeast and phage display libraries; 3) generation of training data set for machine learning developability. 

Integral membrane proteins pose a significant challenge for antibody discovery and optimization. Adimab’s platform incorporates immunized llama or wildtype and humanized mice diversities with a highly engineered strain of yeast. Using this integrated platform approach, we can efficiently discover large numbers of clonally diverse, high-affinity antibodies, and further engineer leads for improved target product profiles, even without the aid of soluble recombinant antigen.