Alejandro “Alex” Velez-Arce, Jesus Caraballo Anaya

Apliko Team - apliko.io

Trinity Prototype

Trinity is Apliko’s agentic-AI-based scientific super intelligence tailored to therapeutic discovery. For Calculus House’s first demo day, the team deployed the Trinity agent over an open-source single-cell protein-protein interaction network to conduct single-cell analyses tailored to therapeutic target discovery. Furthermore, the team developed a more specific definition of “AI Scientist” to emphasize novelty in this work.

Scientific Superintelligence and AI Scientists in Biomedical Sciences & Engineering

The Apliko team envisions “AI scientists” as systems that learn and reason skeptically, working alongside humans to advance biomedical research. These systems blend AI’s ability to crunch massive datasets, explore hypotheses, and handle repetitive tasks with human creativity and expertise. AI agents can plan experiments, assess knowledge gaps, and refine approaches. Powered by large language models and machine learning, they store and build on scientific knowledge, biological principles, and theories. They could transform fields like simulating cells, controlling biological traits, designing cellular circuits, or creating new treatments.

Screenshot 2025-07-01 at 17-27-27 102832.pdf.png

Vision for scientific AI in single-cell therapeutics presented by Velez-Arce et al., at NeurIPS AIDrugX 2024

Differentiation from existing works and Trinity’s contributions to the development of AI scientists

Many works exist implementing “scientific superintelligence”. Here I focus on the systems deployed by FutureHouse to emphasize differentiation.

FutureHouse deployed 4 agents as part of their scientific superintelligence platform. Of these, 3 focus on literature review and 1 is an implementation of a chemistry agent augmented with chemistry tools to complete research workflows such as chemical synthesis by invoking external APIs. The former 3 don’t design experiments and, as such, don’t meet the definition of scientific superintelligence formulated by the Apliko team. The latter does plan and execute experiments, meeting the definition in limited scope. However, Trinity differentiates from Chemcrow and FutureHouse’s agent in more thorough integration of computational and experimental chemistry. This is achieved in the following ways:

  1. Integration of expert-designed ML models. Trinity integrates predictive, generative, and foundation ML models to conduct “digital experiments” of complex tasks. During demo day, Trinity analyzed the embeddings produced by a single-cell protein-protein interaction network GNN to complete cell type labeling based on gene expression and causal relationships between genes and diseases, relevant to target nomination.
  2. Direct use of experimental data. Trinity retrieves experimental data to complete computational biology analyses. In Calculus House’s demo day 1/3, Trinity analyzed clinical trial and target nomination data to identify causal relationships between genes and diseases. Trinity also leveraged raw RNA-seq readouts to complete cell type labeling tasks.
  3. Computational biology code generation. Trinity generated code to complete requested computational biology tasks. Trinity didn’t solely rely on invoking APIs to complete these workflows, as done previously by Chemcrow.