The AI drug discovery industry has entered an exciting new era. High-profile releases like BoltzGen (from MIT/Boltz) and IsoDDE (from Isomorphic Labs) have set new benchmarks for all-atom generative design and protein-ligand (small molecule) affinity prediction.
However, even with these advancements, a critical gap, the "Last Mile" problem, persists. While these frontier models can be great at finding starting points, turning those hallucinated initial hits into viable drug candidates requires solving additional challenges that structure alone still struggles with. It requires biologics affinity maturation, manufacturing stability, and the real clinical requirements of PK, toxicity, and immunogenicity.
The reason is simple: Structural confidence is not affinity. Even with strong structure prediction, critical protein-protein interaction properties such as dynamics and solvent effects are not well encoded in most current training data. As a result, confidence scores like ipTM, iPAE, and ipSAE are best used as negative filters to remove unstable designs, rather than as reliable ways to rank true binding strength or optimize affinity.
DeepSeq.AI is Built for Finishing the Last Mile
To survey massive sequence-functional spaces and generate the proprietary data needed for optimizing therapeutic lead candidates, DeepSeq.AI has built three complementary, hyperscaled platforms:
- MI-LvL (Multiplex Interaction) Platform: Our proprietary "Library vs. Library" platform maps complex interaction landscapes at an unprecedented scale. It is specifically designed to solve the toughest clinical hurdles: PK, Toxicity, and Immunogenicity, by identifying and neutralizing off-target liabilities early in the design phase.
- Maturation Platform: A high-speed maturation engine demonstrated across our Genentech collaborations, designed to rapidly optimize initial binders into high-affinity, high-specificity leads.
- Manufacturing Platform: An explainable-AI guided approach that solves production hurdles using "non-touchable" residue insights to ensure every lead is stable and production-ready for animal and human studies.
Proven Results
In our recent work with Genentech, published in Pharmaceuticals (Feb 2026) and a follow-on article to PLOS Computational Biology (Nov 2024), we highlight the power of pairing Genentech's innovative experimental workflows with DeepSeq's engine to move beyond “zero-shot” prediction and to solve the real-world “Last Mile” drug candidate optimization problem.
- Ensuring Foldability and Yield: In our PLOS Computational Biology paper, we introduced an explainable approach probing the weights of a language model trained on large-scale yeast surface display data to identify "non-touchable" residues indispensable for spontaneous folding. This methodology ensures designed libraries are enriched with hyperstable, high-yield scaffolds, maximizing the proportion of functional members from the start.
- Rapid Affinity Maturation: Our latest work in Pharmaceuticals details the development of the VCX library, a venom-derived platform that achieved a 100% success rate against four diverse and challenging targets, including ion channels. By integrating cost-effective library design with supervised language model optimization, we achieved a 500-fold increase in affinity in a single evolutionary cycle.
We are deeply grateful to the Genentech teams and collaborators who made this innovative research possible through deep partnership and rigorous experimental validation.
The Path Forward
The biologics industry must move beyond asking, "Can AI design a molecule that structurally binds?" to "Can AI deliver a lead drug candidate that is optimized for clinical development success?" While frontier models provide essential starting points, DeepSeq focuses on scaling data generation and on building the fully integrated engine to transform computational potential into validated clinical reality.
