EVQLV’s Evolutionary Algorithm Explained

Jun 23, 2020 | Blog | 2 comments

Bhoomika Kumar

Our technology works by taking an antibody sequence of interest and performing evolutionary modeled mutagenesis, mimicking the natural selection inherent to antibody VDJ recombination and affinity maturation. Using large amounts of public and proprietary sequence data (e.g., PDB, IMGT, human antibody repertoires), our algorithm infers the patterns of evolutionarily likely fragment variable mutations. Biosequence analysis using hidden Markov Models allows for the consideration of sequences that are homologous to the reference antibody chain of interest, and the frequency of different amino acids at various positions, relative to the presence of other amino acids (epistatic modeling) is used to generate a heatmap of evolutionarily possible mutations at each position of the chain. 

In order to achieve higher-order mutational states, the algorithm traverses from first-order mutation, to second, to third and so on by exploiting the likely evolutionary possibilities. The machine searches for the best tactic, in a given mutational state, with the goal of maximizing positive likelihood in evolution. Our algorithms raise overall positive evolutionary likelihood at levels as high as 13 mutations. The algorithm is able to achieve higher order levels of mutational complexity, which could augment the work performed in the lab, all without significant time and cost.

The next step in our pipeline is the screening of the 1-2 million most evolutionarily-likely antibody variants along a host several factors such as temperature, stability, solubility, aggregation propensity, in addition to amino acid characterizations.  The performances of the generated sequences are compared to Phase III and FDA approved antibodies that are run through the same in silico assays. The algorithm then selects AI-generated antibodies that fit within the distribution of Phase 3 and FDA approved antibodies. 

Applying these in silico characterization methods enable early identification of risks such as  immunogenicity, instability, aggregation, high viscosity, or poor expression which can all preclude an antibody from becoming a therapeutic. While success is not guaranteed, this pipeline prevents developability issues early on in order to reduce downstream development risks. 

The most promising sequences from these in-silico assays are clustered by mutational and characteristic features. Both of these clustering approaches are employed so that we may produce highly diverse panels for our collaborators while working within their budget constraints. As an important note, this clustering process is not two or 3-dimensional, which is the extent of human visual perception. This is an over 80 dimensional clustering model that accounts for as much diversity as possible. 

We perform computational structure modeling to assess candidates from each cluster for binding affinity using Rosetta. Through a highly distributed file system, the relative docking scores for all the antibodies are rapidly computed and scored. 

From input of amino acid sequence, the algorithm computationally generates evolutionarily possible sequences with high affinity and lower likelihood of failure in less than a week. By failing as much as possible in the computer, wet lab researchers save immeasurable cost, time, and resources, allowing for significant acceleration of pre-clinical drug development.

Learn more about EVQLV here.


  1. Alex F

    Wow. Fascinating. Great work, Bhoomika.

  2. Fred Berg

    “The next step in our pipeline is the screening of the 1-2 million most evolutionarily-likely antibody variants along a host several factors such as temperature, stability, solubility, aggregation propensity, in addition to amino acid characterizations.”

    I would be interested in learning more about this


Leave a Reply to Alex F Cancel reply

Your email address will not be published. Required fields are marked *

EVQLV and Manhattan BioSolutions Enter into Strategic Collaboration to Discover Novel Antibody-Based Therapies

The collaboration fuses EVQLV’s expertise in AI-enabled antibody design & Manhattan BioSolutions’ expertise in antibody development in oncology

Demystifying AI In Pharma

Examining AI as automated, assistive, analytical, accelerated, or augmented intelligence is a framework for thinking about how to apply these tools to data in ways that are aligned with an organization’s goals. This framework is designed to allow those who do not routinely build or employ AI to avoid the overhyped rhetoric of AI and consider the practical outcomes of employing it.

Andrew Satz on the Amplifying Scientific Innovation Podcast with Dr. Sophia Ononye-Onyia

EVQLV CEO, Andrew Satz, recently joined the Amplifying Scientific Innovation podcast with Dr. Sophia Ononye-Onyia. They discussed his definition of scientific innovation, personal anecdotes of his CEO journey and EVQLV’s unique approach to scientific innovation through the acceleration of biologic therapies with the power of artificial intelligence (A.I.).

How Our Algorithm Models the Drug Discovery Process

Our CEO, Andrew Satz, recently spoked at “Alumni in Conversation,” hosted by Columbia University School of General Studies. During the discussion, Andrew explained how EVQLV’s algorithm models the drug discovery process.

Large Pharma’s Embrace of Artificial Intelligence

The benefits of computational drug development are well known and accepted by many in the pharmaceutical space. However, since the onset of the global pandemic, consensus went from it being an interesting concept that was almost certainly the future, to something we need implemented as soon as possible. This is one of the few silver linings of the coronavirus outbreak.

Pin It on Pinterest