AI & Tech Review ⚡

📌 Navigate

📋 Exec Summary

Q4 2020 delivered three category-defining events simultaneously: AlphaFold 2 solved protein folding at CASP14 with a GDT of 92.4, proving transformers generalize beyond language to biological sequences. The Pfizer-BioNTech and Moderna mRNA EUAs validated programmable biology as a regulatory category. AbCellera's IPO (+194.5% day one) and Exscientia's Phase 1 entry established AI drug discovery as a public asset class with real-world product evidence. The competitive center of gravity shifted from "best physics simulation" to "best foundation model and training data."

📊 What Moved

📈 Trend Arcs

Arc 1: Transformers Eat Biology

Velocity: Accelerating

AlphaFold 2 used a transformer-based architecture — the same underlying mechanism as GPT-3, BERT, and every subsequent large language model — but applied it to evolutionary sequence alignments rather than text. The "evoformer" attention block, which learned to represent pairwise residue relationships from multiple sequence alignments, produced a decisive jump over prior physics-based methods. This cross-domain transfer was the key signal: the transformer was not a language model that happened to work on proteins. It was a general architecture for learning over sequences, and proteins are sequences. The progression across the quarter: October saw incremental pre-CASP coverage of competing approaches; November 30 was the inflection point with CASP14 results; December was dominated by community analysis, implications discussion, and DeepMind's commitment to open-source release. By quarter close, Oxford's Protein Informatics Group had published the most technically rigorous independent analysis, confirming the GDT claims and raising the next-order question: if transformers could crack protein structure, what other biological sequence spaces were next (RNA, DNA, protein-protein interactions, metabolomics)?

Where it stands at quarter close: AlphaFold 2 has won CASP14, transformers have demonstrated cross-domain generalization to biological sequences, and the field is beginning to ask whether a "biological foundation model" — analogous to GPT-3 for biology — is now achievable.

Arc 2: Foundation Model API Distribution

Velocity: Accelerating

GPT-3's API model established a commercial pattern in Q4 2020 that had not existed before: a frontier model accessible via API, enabling hundreds of downstream applications without access to the underlying model weights or training infrastructure. By October 2020, the GPT-3 Playground and API had spawned tools across code generation, clinical documentation, scientific literature review, and legal analysis. The Microsoft exclusive licensing deal (September 2020) set the commercial infrastructure, raising questions about what "open" meant when a frontier model was commercially locked but API-accessible. DALL-E's development in Q4 — announced January 5, 2021 — extended the pattern to multimodal generation. The progression arc: June beta → Q3 ecosystem formation → Q4 commercial differentiation → DALL-E threshold. At quarter close, zero-shot and few-shot prompting are production capabilities, and the API distribution model has been validated as a commercially viable path for deploying frontier AI.

Where it stands at quarter close: GPT-3 API is the dominant production LLM; the ecosystem of downstream applications is established; DALL-E is days from announcement; the pattern of "foundation model as API" is set.

Arc 3: AI Drug Discovery — From Science Project to Asset Class

Velocity: Accelerating

The arc moved decisively from narrative to evidence in Q4 2020. AbCellera's IPO (+194.5% day one) was the market saying AI antibody discovery is not experimental — it is a commercial category with institutional demand. Exscientia's Phase 1 entry was a clinical data point. Insilico Medicine's IND-enabling work was a pipeline signal. Bamlanivimab's real-world deployment in COVID patients was real-world product evidence. The combination — public markets, human trials, commercial product — retired the "science project" framing. Preceding context: the arc had been building through 2020, with Schrödinger's IPO in February and accelerating deal flow in AI-enabled drug discovery partnerships. Q4 was the quarter it cleared the credibility bar. The unresolved question at quarter close was whether AI discovery provided genuine efficiency gains (faster, cheaper candidates) or just narrative premium in a hot market.

Where it stands at quarter close: AI drug discovery is a public asset class with two market anchors (Schrödinger, AbCellera), a Phase 1 clinical data point (Exscientia), and a real-world commercial product (bamlanivimab).

🗺️ Landscape Shift

How the competitive map changed in Q4 2020:

The structural shift: the competitive center of gravity moved from "who has the best physics simulation" to "who has the best foundation model and the best training data." DeepMind's win was not marginal — it retired a category. The question for every computational biology company at quarter close was whether their approach was still relevant in a world where structure prediction was no longer a bottleneck.

💰 Funding & Deal Pattern

• The overall pattern: investor confidence had bifurcated — institutional capital pursued platform validation plays (AbCellera, Moderna), while early-stage capital chased API application formation. The gap between "platform with evidence" and "application on top of a platform" was becoming the primary valuation determinant.

🔍 The Counter-Narrative

• The consensus: AlphaFold 2 solved protein folding, therefore drug discovery is transformed. The reality: AlphaFold solved single-chain structure prediction — a benchmark, not a drug. At quarter close, the model was not yet open-sourced, predictions were for static single-conformation structures, and the protein-ligand co-folding problem (the commercially important question) remained unsolved. "Protein folding solved" was being conflated with "drug discovery transformed now."

• The consensus: FDA moved fast on mRNA vaccines because it was an emergency — a carve-out that would not transfer. The reality: The speed was possible because of a decade of platform work (LNP delivery, mRNA stabilization, spike protein structure) done before the pandemic. The emergency provided urgency; the platform provided capability. The lesson is not "regulatory emergencies accelerate timelines" but "platform investment compresses timelines when urgency is applied."

📐 Builder's Benchmark

API pricing (Q4 2020 reference):

• GPT-3 Davinci: $0.06/1K tokens — expensive by 2026 standards but the only option for frontier language model capability
• No other production-grade LLM API existed at quarter close

Performance benchmarks that shifted:

• Protein structure prediction: AlphaFold 2 at median GDT 92.4, decisively ahead of prior state-of-art.
• Language model capability: GPT-3 175B demonstrated few-shot performance on NLP benchmarks (LAMBADA, HellaSwag, SuperGLUE) that were competitive with fine-tuned BERT-class models, without task-specific training.
• mRNA vaccine efficacy: BNT162b2 at 95%, mRNA-1273 at 94.1% — both above the FDA's 50% minimum efficacy threshold for COVID vaccines and above most prior seasonal flu vaccine benchmarks.

Adoption curves:

• GPT-3 API: hundreds of startups by Q4 2020; developer waitlist was the constraint, not demand.
• AI drug discovery: AbCellera pre-IPO had partnerships with 10+ pharma companies; Exscientia had multiple ongoing collaborations. The adoption metric was partnership agreements and clinical entries, not consumer signups.

Open vs. closed gaps:

• Language models: GPT-3 was closed (API-only; Microsoft exclusive on underlying weights). No open-source alternative at comparable scale existed. The gap between open and closed was measured in orders of magnitude.
• Protein structure: AlphaFold 2 was committed to open-source release but not yet released. Rosetta (academic, open-source) remained the best available open alternative at quarter close, scoring roughly a third of AlphaFold's accuracy.

Time-to-ship:

• mRNA vaccine: sequence design to EUA in under 12 months — the fastest vaccine development in history by a large margin.
• AI-designed drug candidate to Phase 1: Exscientia achieved IND-enabling studies within roughly 12 months of initiating AI-driven discovery for EXS21546.

Implications for Sentium

AlphaFold 2 is proof-of-concept, not yet operational infrastructure. The CASP14 result is a scientific fact. The open-source model and database are committed but not yet released — those are 2021 events. Any Sentium workstream that depends on AlphaFold structural predictions should track the actual release date before adjusting experimental structure dependencies. The planning assumption that high-confidence structures for most human proteins may be freely available by mid-2021 is plausible but is a forward projection. What is true now: the bottleneck for structural biology will shift, in 2021, from "can we determine the structure?" to "what do we do with 200 million predicted structures?" Positioning at the structural database level — not individual structure level — will be the differentiating capability.

The mRNA EUA precedent directly affects nucleic acid development strategy. FDA's acceptance of the platform safety argument for mRNA-LNP delivery is a regulatory comparator that any Sentium program in nucleic acid therapeutics or diagnostics can cite. The precedent is specific: prior research on delivery mechanism (LNP chemistry, mRNA stability, formulation) can substitute for a conventional compound-by-compound safety database when the platform is sufficiently characterized. This does not mean FDA will authorize anything quickly — it means the argument is available. The regulatory question for any Sentium mRNA-adjacent program is not "will FDA accept this modality?" but "do we have the platform characterization data to make the argument?"

The GPT-3 API ecosystem sets the builder baseline for language model capability. Going into 2021, zero-shot and few-shot prompting on GPT-3 class models are production capabilities. Any Sentium application that relies on language model capability — clinical document parsing, regulatory submission drafting, scientific literature analysis — has a clear baseline to build against. The open-vs-closed gap is real: no open-source LLM at comparable scale exists at quarter close. If Sentium requires open-source for regulatory or data-security reasons, the capability gap is significant and is not closing in Q1 2021.

AI drug discovery credibility is now the baseline, not a differentiator. AbCellera's IPO and Exscientia's Phase 1 entry mean the market no longer rewards the AI drug discovery narrative on its own. The differentiation question for any Sentium position in this space is not "do you use AI?" but "what does your AI do that a well-resourced team with AbCellera-tier antibody discovery or Exscientia-tier small molecule discovery cannot?" The bar has moved. Competing on the AI-drug-discovery category positioning without differentiated biology or a novel modality is a losing position entering 2021.

FDA AI/ML SaMD Action Plan architecture design decision. The PCCP framework — expected January 12, 2021 — introduces a design pattern that is easiest to implement from day one and very expensive to retrofit. If Sentium has any SaMD product that involves post-market model updating (which most clinically deployed AI systems will), the architecture decision to design for PCCP compatibility is a Q1 2021 decision, not a "when we get closer to FDA submission" decision. Modular algorithm architecture with defined change boundaries and pre-specified validation methodology are the design requirements. Build them in now.

👀 What to Watch

📎 Sources

Key references for this quarter. Links provided where available; historical entries may reference publications by title and date.