When can you trust an AI-predicted protein structure?

A confidence score and a structure that reproduces are not the same thing. The Lavinda Lab stress-tests predictions where chemistry decides the answer—at metals, glycans, cofactors, and interfaces—and pairs computational hypotheses with appropriate validation.

What does the Lavinda Lab study?

The lab studies chemistry-aware validation of AI structure prediction, including the TYR-TYRP1-TYRP2 melanogenic complex and human diamine oxidase, as well as validation-first AI literacy in science education.

Who runs the Lavinda Lab?

Olga Lavinda, PhD, Assistant Professor of Chemistry and principal investigator at Stern College for Women, Yeshiva University.

The Lavinda Lab · Chemistry-Aware AI Validation

When can you trust an AI-predicted structure?

Q: Where are independent AI literacy workshops offered?

This page documents academic research, teaching, and student training. Independent organizational workshops and curriculum programs are offered separately through Health AI Programs.

AI can now generate protein-complex hypotheses in seconds. The hard part is no longer making the prediction. It is knowing which predictions are chemically real. My lab builds chemistry-aware methods to tell them apart, in the cases where metals, glycans, cofactors, and interface energetics decide the answer.

Olga Lavinda, PhD · Principal Investigator
Assistant Professor of Chemistry · Stern College for Women, Yeshiva University
Katz School of Science and Health · NIH NRSA-trained (NYU Langone)

← lavinda.ai ORCID Scholar YU Faculty

Stylized protein interfaces and a dicopper active site

Structural biology · AI validation

Can you trust an AI-predicted protein complex?

AlphaFold 3 proposes a protein assembly in seconds, but not every prediction is real. We stress-test it where it tends to fail, on metal ions, sugar chains, and the randomness between runs, using the melanogenic enzyme complex TYR-TYRP1-TYRP2, whose assembly underlies albinism and is a drug target in melanoma. The finding holds across hundreds of runs: a high confidence score and a structure that actually reproduces are not the same thing. Telling them apart is the work.

The predicted TYR-TYRP1-TYRP2 melanogenic trimer (AlphaFold 3). Open the dedicated video page →

A confident score is not a reproducible structure, across hundreds of AF3 runs — Each point is one model run. A high confidence score (right) does not guarantee a structure that reproduces (top). That gap is what we measure, and one interface reproduces reliably while another never holds: a biological signal, not a modeling artifact.

Enzyme mechanism · Histamine biology

What in your diet blocks histamine breakdown?

Histamine intolerance affects a large, under-served population, yet the enzyme that clears dietary histamine, human diamine oxidase (hDAO / AOC1), is barely characterized as a diet or drug target. We use AI structure prediction and docking to rank which everyday compounds block it, including terpenes from common herbs and oils no one has tested against hDAO. It is the focus of our current NIH R15 application, with compound screening and enzyme assays running now.

AF3 folds human DAO with high, reproducible confidence across all five models — A consistent, high-confidence AF3 fold of hDAO across all five models. A reliable structure to dock into.

Protein design · ML for structure

A grammar of protein stabilization.

Some protein surfaces are strained and unstable; others lock cleanly into complexes. We are looking for the reusable rules that separate the two: a grammar of stabilization drawn from PDB structures, binding energetics, and protein language models. If those rules transfer, they tell us which proteins will pair stably, and why.

Energetic Dialects: frustrated surfaces to stabilization grammar to stabilized complex

Biocatalysis · Green chemistry

Cleaner ways to oxidize stubborn hydrocarbons.

How do enzymes selectively oxidize some of the most inert hydrocarbons, and can they do it in greener solvents? We study fungal peroxygenases and their active-site chemistry for selective C-H oxyfunctionalization of branched and cyclic alkanes in deep eutectic solvents, with an eye toward cleaner chemical manufacturing.

Fungal peroxygenase selectively oxidizes an inert C-H bond of a cyclic alkane using hydrogen peroxide in a deep eutectic solvent — One inert C-H bond, selectively oxidized under mild, greener conditions: hydrogen peroxide as the oxidant, a deep eutectic solvent as the medium.

Team

The people doing the work.

Undergraduate researchers at Stern College for Women carry the core of this work, co-mentored by MS Biotechnology graduate researchers in a near-peer structure. Every student's contribution is documented, reviewed, and credited through to authorship.

The lab and the city — Based in New York City, the lab works across the classroom, the server, and the bench.

Olga Lavinda, PhD · Principal Investigator

Scientific direction, validation-framework design, and student supervision. ACS faculty mentor, Yeshiva University student chapter.

Graduate researchers · MS Biotechnology, Katz School

Nicolas Hove · Tinashe Rabson Muyambo · Ntobeko Dube · Meli Nkau · Londiwe Moyo

Undergraduate researchers · Stern College for Women

Tiferet Aharon · Netanya Cohen · Dalit Gulkarov

Apply to the lab → New undergraduate and graduate researchers join the lab each term.

Methods

We don't trust a prediction we can't reproduce.

Rather than accepting every AlphaFold output, we treat each prediction as a hypothesis and test it where chemistry tends to break it: at metals, glycans, cofactors, and interface energetics. We then confirm it at the bench. The discipline is straightforward: predict, then verify.

500+

AlphaFold 3 runs scored for multi-seed reproducibility

enzyme systems under active study

100%

of predictions paired with a wet-lab measurement

Predict · computational

AlphaFold 3 multi-seed prediction ML-augmented molecular dynamics QM/MM active-site mechanism Protein-ligand docking (Vina) Protein language models

Verify · wet-lab & biophysical

Enzyme kinetics DSC / thermal stability ITC binding Pre-registered, reproducible pipelines

The lab

Most labs train by osmosis. This one runs a curriculum.

Students arrive with no computational background and leave having carried a project from first paper to conference poster, presented under their own name. The program is run as a documented system rather than an apprenticeship, reflecting the same operational discipline that defines the science.

Academic boundary: this page documents university research, teaching, and student training. Independent organizational workshops and curriculum programs are offered separately through Health AI Programs.

A documented system

Seven skill-gated phases, a training manual, an individual notebook for every student, and mentor sign-off at each step.

Rigorous AI literacy

Students are taught to use AI as the field now requires: with prompt discipline, blinded scoring, and results stated no more strongly than the evidence supports. The approach has been presented at ACS.

Documented authorship

Each student's work is credited: a conference poster presented under their name and a place on the resulting publication.

The seven phases · each builds a specific, transferable skill

Background & framing

Literature · biology

Structure & sequence acquisition

PDB · UniProt

Visualization & validation

PyMOL · ChimeraX

Quality assessment

MolProbity

AI structure prediction

AlphaFold 3

Interface & reproducibility analysis

Python · pandas

Figures & communication

Matplotlib · poster design

Advanced tracks

AutoDock Vina · enzyme kinetics

Undergraduate and graduate researchers welcome. No prior computational-biology experience required.

Apply to the lab →

Recognition

The science is being recognized.

In a single national meeting, four talks were accepted across four ACS divisions: structural biology, medicinal chemistry, computational science, and chemical education. Two drew particular recognition: an invited Committee-on-Science symposium and a Sci-Mix selection.

InvitedCommittee on Science

Deep Learning Co-Folding versus Physics-Based Models of Protein-Ligand Interactions

COMSCI Innovative Program · ACS Fall 2026

Selected · Sci-MixChemical Education

Teaching with AI Without Losing the Student

A validation-first framework for AI literacy in undergraduate chemistry · CHED

Biological Chemistry

AlphaFold 3 Reveals the Architecture of the Melanogenic Enzyme Complex

Metal coordination, glycosylation, and the structural basis of albinism · BIOL

Medicinal Chemistry

Docking and Allosteric Exploration of Supplement Ingredients at Human Diamine Oxidase

An AlphaFold 3 and AutoDock Vina study · MEDI

Selected publications

Origin of high diastereoselectivity in reactions of seven-membered-ring enolates

Lavinda, Witt & Woerpel · Angewandte Chemie · 2022

Biophysical compatibility of the tyrosinase-TYRP1-TYRP2 metalloenzyme complex

Frontiers in Pharmacology · 2021

Two papers in preparation on chemistry-aware AF3 benchmarking and the TYR-TYRP1-TYRP2 complex

In preparation · 2026

Full publication list →